ai in astronomy research paper

Machine learning in astronomy

• Published: 16 October 2022
• Volume 43 , article number  76 , ( 2022 )

Cite this article

• Ajit Kembhavi   ORCID: orcid.org/0000-0001-8164-311X 1 &
• Rohan Pattnaik 2  

826 Accesses

2 Citations

1 Altmetric

Explore all metrics

Artificial intelligence techniques like machine learning and deep learning are being increasingly used in astronomy to address the vast quantities of data, which are now widely available. We briefly introduce some of these techniques and then describe their use through the examples of star-galaxy classification and the classification of low-mass X-ray binaries into binaries, which host a neutron star and those which host a black hole. This paper is based on a talk given by one of the authors and reviews previously published work and some new results.

Similar content being viewed by others

Deep Learning: A Comprehensive Overview on Techniques, Taxonomy, Applications and Research Directions

Iqbal H. Sarker

Scientific Machine Learning Through Physics–Informed Neural Networks: Where we are and What’s Next

Salvatore Cuomo, Vincenzo Schiano Di Cola, … Francesco Piccialli

Review of deep learning: concepts, CNN architectures, challenges, applications, future directions

Laith Alzubaidi, Jinglan Zhang, … Laith Farhan

Avoid common mistakes on your manuscript.

1 Introduction

Abundant astronomical data is now freely available because of surveys like the the Sloan digital sky survey (SDSS, see Abdurro’uf et al. ( 2022 ) for the latest data release and references) and the more recent Subaru hyper suprime-cam survey (see Aihara et al. 2022 for the latest data release and references). Conventional data analysis techniques will seriously constrain the scientific projects which can be undertaken with such databases due to the sheer volume of the data. To tap the full potential of the data, it is necessary to use artificial intelligence techniques like machine learning (ML) and deep learning (DL), which have evolved rapidly over the past few decades (e.g., Baron 2019 ), making them very useful for a variety of applications. These developments and the availability of software platforms like TensorFlow 2 and Keras (TensorFlow 2 is a free and open-source software library for ML, DL, etc., developed by Google researchers, see https://www.tensorflow.org/guide/effective_tf2 . Keras is a deep learning API written in Python, running on top of the machine learning platform TensorFlow 2, developed by Chollet, François et al. see, https://keras.io ) have enabled astronomers to use ML and DL for addressing the very large volumes of imaging, spectral and catalogue data that are now easily accessible to them.

Some examples of application of ML and DL to astronomy include photometric redshift estimation (D’Isanto & Polsterer 2018 ; Pasquet et al. 2019 ), gravitational lensing identification (Cheng et al. 2020 ), light curve classification (Lochner et al. 2016 ; Mahabal et al. 2019 ; Möller & de Boissière 2020 ), stellar spectrum classification and interpolation (Kuntzer et al. 2016 ; Sharma et al. 2020a , b ), galaxy morphology classification (Dieleman et al. 2015 ; Abraham et al. 2018 ; Domínguez Sánchez et al. 2018 ; Barchi et al. 2020 ; Walmsley et al. 2020 ), and star-galaxy classification (Philip et al. 2002 ; Ball et al. 2006 ; Vasconcellos et al. 2010 ; Abraham et al. 2012 ; Soumagnac et al. 2015 ; Kim & Brunner 2017 ; Clarke et al. 2020 ).

In the following, we will briefly describe a few important ML and DL techniques and two illustrative applications to astronomy: star-galaxy separation and establishing the identity of the compact object in low-mass X-ray binaries on the basis of their X-ray energy spectra.

2 Machine learning and deep learning algorithms

A number of algorithms are conventionally grouped together under ML. These are used to build computer programs which automatically improve with experience. While the algorithms can be used for classification and regression, we will be considering only the former in our examples. For a classification problem, the input to the program consists of a large number of training examples with each example having a number of measurable attributes, and belonging to one of several defined classes. The program learns from the training set to distinguish between the classes based on the attributes. After the learning is complete, the program is able to predict the class of previously unclassified objects on the basis of their attributes, i.e., the program is able to generalize the classification, learned using a finite training sample and to the examples beyond the training set. This process is known as supervised learning, because the input training set includes the known class of the objects in the sample. In unsupervised learning, the program can be asked to classify a training sample into a given number of classes on the basis of attributes with no prior classification being specified. Unsupervised learning is useful when possible novel classification schemes are to be investigated.

Machine learning includes algorithms like random forest (RF) and artificial neural networks (ANN) (see e.g., Mitchell 1997 ), while DL, which is also a part of ML, but is generally mentioned separately includes algorithms like convolutional neural networks (CNN), recurrent neural networks (RNN), generative adversarial networks (GNN), etc., (see e.g., Lecun et al. 2015 ; Goodfellow et al. 2016 ; Guo et al. 2016 ). We will briefly describe below about ANN, random forest and CNN, and two applications which will provide some insight into how the algorithms work.

2.1 Artificial neural networks

The basic unit of an ANN is a artificial neuron, which is historically loosely based on a biological neuron. An ANN with an input layer, a hidden layer and an output layer is shown in Figure 1 . A neuron in the hidden layer receives inputs from all neurons in the preceding input layer and contributes to every output neuron. The output of the \(j\mathrm{th}\) neuron is given by

where the \(w_{ji}\) is known as the weights and \(\sigma \) is a non-linear activation function. The usual forms of \(\sigma \) used are the sigmoid function

or the rectified linear unit ReLU function \(\sigma (y)=\max (0,y)\) , which is preferred because it leads to faster training for many layered ANN. To avoid dead neurons, which have output zero and which sometimes arise when ReLU is used and a leaky ReLU is used, where \(\sigma (y)=y\) for \(y\ge 0\) and \(\sigma (y) = 0.01y\) for \(y<0\) . In general, there can be any number of input nodes, and depending on the complexity of the classification boundaries in the multi-dimensional space of the inputs, there can be several hidden layers. Each neuron in the ANN is connected to all the neurons in the preceding and succeeding layers, so the ANN is said to be fully connected.

An ANN with four inputs, one hidden layer and two output classes.

For supervised learning, training the network consists of suitably adjusting the weights so that the desired known output is obtained for a given input. In the star-galaxy problem described below, a certain number of parameters are measured for objects which are known to be stars or galaxies. The parameters for a large number of such known objects are in turn input to a network, with desired output −1 for a star and 1 for a galaxy (say). A loss function is now defined which compares the output for a starting random choice of weights with the desired inputs for the training set. The loss function is then minimized over the multiple passes of the training data, using a technique known as back propagation (Mitchell 1997 ). Suitable definitions are used to measure accuracy (or precision), which is the ratio of true positives to the sum of true and false positives over all classes, and the completeness (or recall), which is the ratio of true positives to the sum of true positives and false negatives over all classes. There are many variations possible on the basic structure of an ANN described here, and there are details of many matters and technicalities which we have not mentioned.

2.2 Random forest

Random forest (RF) is an ensemble technique which is used to boost the prediction made by an individual decision tree (Breiman 2001 ). A decision tree is one of the most intuitive, yet powerful ML algorithms (Breiman et al. 1984 ). A decision tree is made up of branches of nodes, where sets of if-this-then-that rules are applied to the features of the input data, and based on the result, leads down to one of the branches of the tree. The final layer of nodes, also known as leaf nodes, contain a predicted class label which is compared to the expected class for a particular input vector. Although the decision tree algorithm has proven to be very efficient (see e.g., Vasconcellos et al. 2011 ), a decision tree, if improperly trained, can at times over-fit the data (Mitchell 1997 , Chapter 3). The idea behind RF is to combine the decisions of several such trees to improve upon the decision of a single over-trained tree. Taking a majority-vote over the decision of all the trees, helps in reducing the variance of the predictions (Breiman 2001 ). The probability of a source belonging to one class or the other is also calculated in a similar way, i.e., by dividing the number of trees that predicted the same class by the total number of trees. We illustrate the decision making process of a RF algorithm in Figure  2 .

2.3 Deep learning: convolutional neural networks

Conventional ML uses as inputs features extracted from the raw data, like images or spectra. As the data get complex, for example, when images of galaxies are to be classified, extracting a manageable number of features can be very difficult or even impossible. If all the pixels of an image were to be used directly as input to an ANN, then for CCD images with millions of pixels, the number of input nodes, hidden nodes and weights for all the connections would become so large as to make the network unmanageable. DL techniques are designed to get over these difficulties by using raw data to avoid the extraction of features to be used as inputs to the network. In fact, the network itself extracts features from the raw data by convolving it with a set of filters to provide a representation of the image at a more abstract level for the classification (Lecun et al. 2015 ). The classifier in the network uses downsized extracted features, so that number of weights remain manageable.

Illustration of the decision making procedure in a random forest algorithm.

In convolutional neural networks (CNN), is the only DL technique that we will consider in this paper. A CNN has three types of layers, convolutional layers, pooling layers and fully connected layers. In a network, the convolutional layers alternate with the pooling layers and following a number of such pairs, there are the fully connected layers, which lead to the final classified output. In a convolutional layer, the input image is convolved with a set of kernels to generate feature maps. The kernels for extracting specific features like edges at various orientations and other motifs are not provided by the user, they are learned by the network for the input data being used. A convolutional layer is followed by a pooling layer, in which the dimensionality of a feature map is reduced by taking the maximum (or average) pixels over a \(2\times 2\) array, say, which strides over a feature map. After the final pooling layer, a set of fully connected layers is used to produce the output. These perform like a conventional ANN and the output provides the categories for classification with a probability associated with each class. The weights for all the layers are trained using back propagation, as in the case of an ANN. The need for a CNN and its operation can be better understood through the problem of star-galaxy classification discussed in Section 3 .

3 Star-galaxy classification

Very compact galaxies or large galaxies at great distances, can resemble stars in their appearance in astronomical images at optical or near-infrared wavelengths. So separating such galaxies from stars for galaxy surveys can be difficult on the basis of just their images. Stars have the appearance of the point spread function (PSF) for the image, which is determined by the earth’s atmosphere to a large extent with contributions made by the telescope optics and structure, etc. In the ideal case, the PSF is a 2-dimensional circular Gaussian with a full-width at half-maximum (FWHM) of \({\sim }1\) arc-sec in good-seeing conditions, but is somewhat distorted in practice. Compact or distant galaxies which are significantly larger than the PSF can be easily distinguished as such. But the images of galaxies approaching the PSF in size are more difficult to distinguish: even though their shape is different in detail from the shape of the PSF, the differences are small and hard to discern. An expert astronomer would be needed to separate such galaxies from stars, but the task would take simply too long for large data sets, and therefore there is a very good case for the use of ML.

3.1 Star-galaxy classification with ANN

We will discuss here one of the early works, which uses ML for star-galaxy classification (Philip et al. 2002 ; a few other early results are cited in the reference). Philip et al. ( 2002 ) used R band images from the publicly available NOAO deep wide field survey (Jannuzi et al. 2000 ). The training set was constructed from a sub-image of the R band image NDWFSJ1426p3456, which had the best-seeing conditions for the data released in 2001 A total of 402 objects in the training set, in the magnitude range of 20–26, were visually classified as stars or galaxies independently by two of the authors, and the \(\sim \) 2% cases where the classification turned out to be different, were resolved by a joint inspection. The final training set had 83 stars and 319 galaxies. The number of stars was smaller than the number of galaxies because of the high galactic latitude of the field and the faintness of the objects.

Four faint compact galaxy r, g, b band composite images from the SDDS.

For each object, three parameters were measured: (1) an elongation measure, which is the ratio of the second order moments along the major- and minor-axes of the faintest isophote, (2) a standardized FWHM measure, which is the logarithm of the ratio of the FWHM of the object to the FWHM of the PSF for the image and (3) a gradient parameter which is the logarithm of the ratio of central peak count to the FWHM of the object, normalized to the standardized FWHM measure. A difference boosting neural network (DBNN, Philip et al. 2002 ) was used in the training with the three parameters as the inputs. The trained network was used to classify a test set consisting of a total of 154 stars and 558 galaxies from the two sub-images of the field, which had been previously visually classified as stars and galaxies as in the case of the training set. An overall accuracy of 98.1% classification was obtained, which was better than the 96.1% accuracy obtained for the same test set using SExtractor (Bertin & Arnouts 1996 ).

In the above project, the size of the training and test set was small because of the difficulty in visually identifying compact galaxies. The small sample can result in the overfitting of the network, which while providing good accuracy in the training, does not perform well in generalizing to a variety of images outside the training set. Moreover, it would be difficult to visually identify large samples of galaxies and to measure parameters for the training of galaxies which are faint and/or irregular, like the galaxies shown in Figure 3 . It is therefore, necessary to use a technique based on DL, so that raw images of stars and galaxies can be used in the training. It is also necessary to have large samples of known galaxies and stars for the training of a DL based network.

3.2 Star-galaxy classification with CNN

We will now consider work by Chaini et al. ( 2022 ) on star-galaxy classification which satisfies the above requirements. The authors consider a more general star-quasar-galaxy separation with the aim being to identify these objects on the basis of their photometric properties alone, but we will only consider the star-galaxy sector. The training sample consists of spectroscopically identified stars and galaxies from SDSS data release 16 (DR16, Ahumada et al. 2020 ). The spectroscopic identification is secure and no visual identification is necessary. In the training only photometric data of the five SDSS passbands u, g, r, i, z are used. The sample is limited to compact objects, which are defined as those which have the ratio

Here, half light radius, or de Vaucouleurs radius, is the radius containing half of the total light of the object, and the average of the ratio in the five passbands is used. When only a faint subset is to be considered, the criterion for faintness is that the average of the five band magnitudes \(\langle m \rangle > 20\) . The training dataset consists of 80,000 objects each for the two classes, chosen randomly from larger sets, which satisfy the criteria.

The CNN used in this case is based on the inception network (Szegedy et al. 2014 ) and has five dense layers of size 1024, 512, 256, 128 and 64, respectively (a dense layer has its neurons connected to every neuron in the preceding layer) with varying kernel sizes. Each layer is itself made up of four parallel convolutional layers, each activated by a leaky ReLU and averaging pooling layers are used. The loss function used is known as categorical-cross entropy and a total of 25,544,807 neurons are trained. The final layer contains a softmax function, which provides the output as probabilities for the two classes, star and galaxy.

When the training, validation and test samples are all drawn from the dataset described above, and compact objects are considered, the accuracy for star-galaxy separation reached with the CNN is 97.4%. If compact and faint objects with average of the five band magnitudes \(\langle m \rangle > 20\) are considered, the accuracy drops marginally near to 95.2%. The reason for the drop of course is that the fainter objects have poorer signal-to-noise ratio, so the separation into stars and galaxies is more difficult.

Chaini et al. ( 2022 ) have also used photometric parameters for each object provided by the SDSS data processing pipeline to carry out the separation using an ANN. The parameters are the magnitudes in the five bands corrected for extinction, the half light radius, FWHM of the PSF and the extinction in each of the five bands, and the colours u–g, g–r, r–i and i–z, making a total of 24 parameters. The accuracy reached for the separation with the ANN is 97.9% and 96.0%, respectively, for compact objects, and compact and faint objects. Chaini et al. (2022) further consider an ensemble of the CNN and ANN that they call MargNet, which has a combined accuracy of 98.1% and 96.9% for the two cases. Using the ensemble of CNN and ANN is therefore the best option. The important point to note here is that the CNN works directly on star and galaxy images, and does not need any measured parameters and provides high accuracy. It is therefore useful even when a dependable pipeline for measuring photometric parameters is not available. Other examples of star-galaxy-quasar classification using CNN include Kim & Brunner ( 2017 ) and Clarke et al. ( 2020 ).

4 Low-mass X-ray binaries

Low-mass X-ray binaries (LMXBs) are binary systems where one of the components is a black hole (BH) or a neutron star (NS); the other component is a less massive star, usually on main sequence or an evolved star of mass \(M<1~M_{\odot }\) . Some LMXBs show long quiescent periods, which can last from a few months to decades, when the source is very faint. There are also short periods, lasting from days to months, when the source is in outburst, with the flux increasing by several orders of magnitude (see e.g., McClintock & Remillard 2006 ).

The energy spectra of LMXB systems are described by two main components: (1) a thermal component which is usually described by a multi-color disc blackbody, thought to be produced by an accretion disc and (2) a hard component thought to be produced by a corona, which is a region of hot plasma around the compact object. This component is usually described by a thermal Comptonization model. The contribution of these components to the X-ray emission varies during an outburst, leading to modification of its spectral and timing properties. References to the details about LMXB energy spectra are provided in Pattnaik et al. ( 2021 ).

One of the important questions about LMXBs is whether the compact object in the binary is a NS or a BH. The nature of the compact object has a significant impact on the physical interpretation of the observations. With the large scale sky surveys and transient search programs, e.g., INTEGRAL/JEM-X (Lund et al. 2003 ), Swift/BAT transient monitor (Krimm et al. 2013 ), MAXI (Matsuoka et al. 2009 ), eROSITA (Merloni et al. 2012 ), the sample of LMXBs keeps increasing. Such newly detected transient sources are usually characterized by their fast variation (days) of luminosity by orders of magnitude. Early identification of the nature of the compact object is very important for the community to be able to plan observing campaigns (Middleton et al. 2017 ).

There are only a few methods for identifying the nature of the compact object. For example, coherent pulsations and the presence of thermonuclear bursts (for reviews see, Lewin et al. 1993 ; Cumming 2004 ; Galloway et al. 2008 ; Strohmayer et al. 2018 ), indicate that the compact object is a NS. Based on the mass function of the X-ray binary system, if the mass of the compact object is estimated to be greater than about 3  \(M_\odot \) , then the compact object can be taken to be a black hole. Apart from that, one can surmise the nature of the compact object by comparing its X-ray timing and spectral properties and X-ray-radio correlation with those of sources where the nature of the compact object is known.

One technique that is yet to be fully explored to classify LMXBs is the use of ML algorithms. ML has been used by Huppenkothen et al. ( 2017 ) to classify light curves of the unusual BH X-ray binary GRS 1915 \(+\) 105. It has also been used by Gopalan et al. ( 2015 ) to distinguish between different types of X-ray binaries. We describe below how ML can be applied to the X-ray energy spectra of LMXB to identify the nature of the compact object.

We used the Rossi X-ray timing explorer (RXTE) mission (Bradt et al. 1993 ) data archive, Footnote 1 which provides more than 8500 observations of 33 NS systems and more than 6000 observations of 28 BH systems. We used data from the proportional counter array (PCA, Glasser et al. 1994 ) instrument aboard RXTE, which has an energy range of 2–60 keV to create the energy spectra. We selected a total of 61 sources, which are classified as BH or NS binaries, with classification well established (see e.g., Corral-Santana et al. 2016 ; Tetarenko et al. 2016 , for BH). In the dataset, we have a fairly balanced representation of the two classes, with 8669 observations from 33 sources identified as neutron-star LMXBs (58%) and 6216 observations from 28 sources identified as black-hole LMXBs (42%). The number of observations per source varies greatly from source to source. A few sources have >1000 observations while some have <20 observations. Some details of the procedure followed to obtain the energy spectra for our analysis are described in Pattnaik et al. ( 2021 ).

For each observation, we used 43 channels in the energy range of 5–25 keV. The number of channels is kept fixed at 43 for all the observations since ML algorithms require each observation used in the training and testing to have the same size. The 43 count rate values are used directly as an input vector for the algorithm.

4.2 Random forest for classifying LMXB

We wish to determine the nature of the compact object on the basis of the energy spectrum. The object can be any one of two types, a black hole or a neutron star. The training set consists of 14,885 spectra of 61 X-ray binaries for which we know which of the two kinds the compact object is. This is a supervised binary classification problem of classifying an X-ray spectrum into two labeled classes. There are several ML algorithms that can be used for handling this type of binary classification problem. From those, we have to choose the one which provides the best accuracy, i.e., the highest percentage of correct classifications. We experimented with a number of algorithms including classification and regression trees (CART), more commonly known as decision trees (Breiman et al. 1984 ), random forest (RF) (Breiman 2001 ) which we briefly described in Section 2.2 , XGBoost (XGB) (Chen & Guestrin 2016 ), logistic regression (LR) (Cox 1958 ), K-nearest neighbors (KNN) (Cover & Hart 2006 ) and support vector machines (SVM) (Cortes & Vapnik 1995 ). These are all traditional ML algorithms that are usually known to show satisfactory performance even with a limited amount of data. They also have significantly lower execution times compared to DL methods (see e.g., Kotsiantis et al. 2007 ).

To establish the best algorithm, we compared their performance using accuracy as a metric. Here accuracy is defined as the ratio of the number of observations correctly classified to their class (neutron star or black hole), to the total number of observations. Using a k-fold cross-validation technique (Burman 1989 ), in which the set of 14,885 observations is split into training and test sets in many different ways, we find that RF provides the best accuracy of \(91\pm 2\) % and use it in the subsequent analysis. We implement the RF algorithm using the scikit-learn Footnote 2 (Pedregosa et al. 2011 ) library of python.

Plot showing individual source wise accuracy using the leave-one source out method of cross-validation. The filled circles are black hole binaries while open circles are neutron star binaries. The area of the points corresponds to the number of observations in each source. The figure is based on (Pattnaik et al. 2021 ).

4.3 Methods and results

We apply the RF algorithm with the best combination of hyper-parameters to the dataset described above. Hyper-parameters are a set of parameters defined prior to the training process that are used to tune the performance of the ML algorithm. Since the dataset contains 14,885 observations for 61 individual X-ray sources, each source has multiple observations taken at different times. The LMXB are variable in nature, so observations for the same source taken at different times typically sample a different physical spectral state, which correspond to different geometrical configurations in the source. We can therefore, assume that each observation for a given source is considered independent of the other observations. Traditionally, the data set of 14,885 spectra would be randomly split into a training set containing 90% (say) of the sources with the other 10% (say) forming the test set. However, this can lead to biases because in this method, some spectra of a given source can belong to the training set, while other spectra of the same source could belong to the test set. This can lead to overestimation of the accuracy reached and affect the predictive power of the nature of the compact source for the spectra of new LMXB. We therefore, chose not to use this method.

We find that optimal use of the data is made when we keep all observations from one source as the test data, while using all the remaining sources for training, and this experiment is repeated for each source. This provides the results for all the observations from each of the 61 sources. The size of the training and test sets vary in each run and each model uses one source less than the total number in the data. The final model is trained on the entire dataset. We show in Figure  4 , the accuracy obtained for each source using this method. There are four sources that lie below the 50% average accuracy mark. The sigma-clipped average accuracy is \(87\pm 13\) %, which gives a lower bound proxy on the performance of our final model.

While the RF algorithm provides good overall classification of low-mass X-ray binary sources into BH and NS types, there are a few sources for which the accuracy is low and most of the observations of those sources are misclassified. Four sources, XTE J1118 \(+\) 480 (BH), XTE J1748−288 (BH), IGR J00291 \(+\) 5934 (NS) and 1A 1246−588 (NS), have <50% accuracy, out of which the observations of XTE J1118 \(+\) 480 and XTE J1748−288 are consistently misclassified with overall accuracy percentage of \(\sim \) 30%. Two factors that can influence the energy spectra are the signal-to-noise ratio (SNR) and the physical state of LMXB systems at the time of observation.

SNR is defined as the net count rate divided by the error in the net count rate for each spectrum. The SNR over the sample ranges from <4 to >5800. We find that for observations with SNR <100, the distribution of predicted probabilities peaks at 0.58. For SNR in the range of 100–1000 and for SNR \(> 1000\) , the distribution peaks are at 0.87 and 0.91, respectively. The performance of the classification model therefore, improves with the increase in SNR. Among the sources which were misclassified, only 1A1246−588 had an average SNR \(<100\) . Therefore, there are reasons other than low SNR for the poor classification of sources. We find that the algorithm performs better for soft-state observations as compared to hard-state observations for individual sources (see Figures 8 and 9 of Pattnaik et al. 2021 ).

4.4 Prediction for a sample of sources

We have used the RF model trained on all 61 sources to predict the classification of 13 systems which have a total of 766 spectra, but where the nature of the compact object is still not established (Table  1 ). If \({>}50\) % of the spectral observations of a source were predicted to belong to a particular class, then that class was assigned to the source. It is seen from the table that five sources have very few observations ( \({<}10\) ) that meet our criteria for good data and it is difficult to make any comments on the predicted classes for these sources. The remaining eight sources all have >30 observations each and six of these sources are classified as BH LMXBs, while two sources are classified as NS LMXBs. Six of these eight sources have prediction percentage \({>}60\) %, while the remaining two sources have prediction in the range of 50–60% . All the 13 sources have an average SNR \({<}100\) , which is the region where the algorithm has the worst performance.

4.5 Discussion

Our classification model is trained specifically on RXTE data and cannot be used directly to classify the energy spectra from other X-ray missions. It is in principle possible to train a classification model for different missions using data from the missions, but there may not be enough data in every case to train a ML algorithm. However, the concept of transfer learning could be employed to train an algorithm for another instrument with limited data using our pre-trained classification model for RXTE data.

Adding more information as input to the algorithm can also be explored as a means of improving the current level of accuracy reached for all the sources in our dataset. One way of doing that would be to combine the energy spectra with the power spectra of all the observations for each source.

There is now a considerable amount of data obtained with LAXPC detector on AstroSat (Yadav et al. 2021 ). It should be possible to apply ML and DL techniques to the data to have further useful information from it than has been done so far using conventional methods.

https://heasarc.gsfc.nasa.gov/cgi-bin/W3Browse/w3browse.pl .

https://scikit-learn.org/stable/ .

Abdurro’uf, Accetta K., Aerts C., et al. 2022, The Astrophysical Journal Supplement, 259, 35. https://doi.org/10.3847/1538-4365/ac4414

Abraham S., Philip N. S., Kembhavi A., Wadadekar Y. G., Sinha R. 2012, Monthly Notices of the Royal Astronomical Society, 419, 80

Article   ADS   Google Scholar  

Abraham S., Aniyan A. K., Kembhavi A. K., Philip N. S., Vaghmare K. 2018, Monthly Notices of the Royal Astronomical Society, 477, 894

Ahumada R., Prieto C. A., Almeida A., et al. 2020, The Astrophysical Journal Supplement, 249, 3. https://doi.org/10.3847/1538-4365/ab929e

Aihara H., AlSayyad Y., Ando M., et al. 2022, Publications of the Astronomical Society of Japan, 74, 247. https://doi.org/10.1093/pasj/psab122

Ball N. M., Brunner R. J., Myers A. D., Tcheng D. 2006, The Astrophysical Journal, 650, 497

Barchi P. H., et al. 2020, Astronomy and Computing, 30, 100334

Article   Google Scholar  

Baron D. 2019, Machine learning in astronomy: a practical overvie, 1904.07248

Bertin E., Arnouts S. 1996, Astronomy & Astrophysics, 117, 393. https://doi.org/10.1051/aas:1996164

Bradt H., Rothschild R., Swank J. 1993

Breiman L. 2001 Machine Learning, 45, 5

Breiman L., Friedman J., Olshen R., Stone C. 1984, Group, 37, 237

Burman P. 1989, Biometrika, 76, 503

Article   MathSciNet   Google Scholar  

Chaini S., Bagul A., Gondkar A., Sharma K., Vivek M., Kembhavi A. 2022, Photometrical identification of compact galaxies, stars and quasars using multiple neural networks, in preparation

Chen T., Guestrin C. 2016, in Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, p. 785

Cheng T.-Y., Li N., Conselice C. J., Aragón-Salamanca A., Dye S., Metcalf R. B. 2020, Monthly Notices of the Royal Astronomical Society, 494, 3750

Clarke A. O., Scaife A. M. M., Greenhalgh R., Griguta V. 2020, Astronomy & Astrophysics, 639, A84

Corral-Santana J. M., Casares J., Muñoz-Darias T., et al. 2016, The Astrophysical Journal, 587, A61

Google Scholar  

Cortes C., Vapnik V. 1995, Machine learning, 20, 273

Cover T., Hart P. 2006, IEEE Trans. Inf. Theor., 13, 21

Cox D. R. 1958, Journal of the Royal Statistical Society: Series B (Methodological), 20, 215

MathSciNet   Google Scholar  

Cumming A. 2004, Nuclear Physics B Proceedings Supplements, 132, 435

Dieleman S., Willett K. W., Dambre J. 2015, Monthly Notices of the Royal Astronomical Society, 450, 1441

D’Isanto A., Polsterer K. L., 2018, Astronomy & Astrophysics, 609, A111

Domínguez Sánchez H., Huertas-Company M., Bernardi M., Tuccillo D., Fischer J. L. 2018, Monthly Notices of the Royal Astronomical Society, 476, 3661

Galloway D. K., Muno M. P., Hartman J. M., Psaltis D., Chakrabarty D., 2008, The Astrophysical Journal Supplement Series, 179, 360

Glasser C. A., Odell C. E., Seufert S. E. 1994, IEEE Transactions on Nuclear Science, 41, 1343

Goodfellow I., Bengio Y., Courville A. 2016, Deep Learning, The MIT Press

Gopalan G., Vrtilek S. D., Bornn L. 2015, The Astrophysical Journal, 809, 40

Guo, Y., Liu, Y., Oerlemans, A., et al. 2016, Neurocomputing, 187, 27

Huppenkothen D., Heil L. M., Hogg D. W., Mueller A. 2017, Monthly Notices of the Royal Astronomical Society, 466, 2364

Jannuzi B. T., Dey A., Tiede G. P., Brown M. J. I., NDWFS Team 2000, AAS

Kim E. J., Brunner R. J., 2017, Monthly Notices of the Royal Astronomical Society, 464, 4463

Kotsiantis S. B., Zaharakis I., Pintelas P. 2007, Emerging artificial intelligence applications in computer engineering, 160, 3

Krimm H. A., et al. 2013, The Astrophysical Journal Supplement Series, 209, 14

Kuntzer T., Tewes M., Courbin F. 2016, Astronomy & Astrophysics, 591, A54

Lecun Y., Bengio Y., Hinton G. 2015, Nature, 521, 436. https://doi.org/10.1038/nature14539

Lewin W. H. G., van Paradijs J., Taam R. E. 1993, Space Science Reviews, 62, 223

Lochner M., McEwen J. D., Peiris H. V., Lahav O., Winter M. K. 2016, The Astrophysical Journal https://doi.org/10.3847/0067-0049/225/2/31

Lund N., et al. 2003, The Astrophysical Journal, 411, L231

Mahabal A., et al. 2019, Publications of the Astronomical Society of the Pacific, 131, 038002

Matsuoka M., et al. 2009, Publications of the Astronomical Society of Japan, 61, 999

McClintock J. E., Remillard R. A., 2006, Black hole binaries, 157

Merloni A., et al. 2012, 1209.3114

Middleton M. J., et al. 2017, New Astronomy, 79, 26

Mitchell T. 1997b, Machine Learning (New York: McGraw-Hill)

MATH   Google Scholar  

Möller A., de Boissière T. 2020, Monthly Notices of the Royal Astronomical Society, 491, 4277

Pasquet J., Bertin E., Treyer M., Arnouts S., Fouchez D. 2019, Astronomy & Astrophysics, 621, A26

Pattnaik R., Sharma K., Alabarta K., et al. 2021, Monthly Notices of the Royal Astronomical Society, 501, 3457

ADS   Google Scholar  

Pedregosa F., et al. 2011, Journal of Machine Learning Research, 12, 2825

Philip N. S., Wadadekar Y., Kembhavi A., Joseph K. B., 2002, Astronomy & Astrophysics, 385, 1119

Sharma K., Kembhavi A., Sivarani T., Abraham S., Vaghmare K., 2020a, Monthly Notices of the Royal Astronomical Society, 491, 2280

Sharma K., et al. 2020b, Monthly Notices of the Royal Astronomical Society, 496, 5002

Soumagnac M. T., et al. 2015, Monthly Notices of the Royal Astronomical Society, 450, 666

Strohmayer T. E., et al. 2018, The Astronomer’s Telegram, 11507

Szegedy C., Liu W., Jia Y., et al. 2014, 1409.4842

Tetarenko B., Sivakoff G., Heinke C., Gladstone J. 2016 The Astrophysical Journal Supplement Series, 222, 15

Vasconcellos E. C., de Carvalho R. R., Gal R. R., et al. 2010, Astro-Phys. https://doi.org/10.1088/0004-6256/141/6/189

Vasconcellos E., De Carvalho R., Gal R. 2011, The Astronomical Journal, 141, 189

Walmsley M., et al. 2020, Monthly Notices of the Royal Astronomical Society, https://doi.org/10.1093/mnras/stz2816 ,491, 1554

Yadav J. S., Agrawal P. S., Misra R., Roy J., Pahari M. R. K. 2021, Journal of Astrophysics and Astronomy, 496, 5002

Download references

Acknowledgements

This paper is based on a talk given by one of the authors, Ajit Kembhavi at the ‘Astrophysical jets and observational facilities: National perspective’ meeting at ARIES, Nainital in April 2021, which described ML and DL techniques as well as work on star-galaxy classification by Chaini et al. ( 2022 ) and on the classification of LMXB by Pattnaik et al. ( 2021 ). The authors wish to thank an anonymous referee for suggestions which helped to significantly improve the manuscript. The data underlying this paper are publicly available in the High Energy Astrophysics Science Archive Research Center (HEASARC) at https://heasarc.gsfc.nasa.gov/db-perl/W3Browse/w3browse.pl and the SDSS archives.

Author information

Authors and affiliations.

Inter-University Centre for Astronomy and Astrophysics (IUCAA), Pune, 411007, India

Ajit Kembhavi

School of Physics and Astronomy, Rochester Institute of Technology, Rochester, NY, 14623, USA

Rohan Pattnaik

You can also search for this author in PubMed   Google Scholar

Corresponding author

Correspondence to Ajit Kembhavi .

Additional information

This article is part of the Special Issue on “Astrophysical Jets and Observational Facilities: A National Perspective”.

Rights and permissions

Reprints and permissions

About this article

Kembhavi, A., Pattnaik, R. Machine learning in astronomy. J Astrophys Astron 43 , 76 (2022). https://doi.org/10.1007/s12036-022-09871-2

Download citation

Received : 09 December 2021

Accepted : 03 June 2022

Published : 16 October 2022

DOI : https://doi.org/10.1007/s12036-022-09871-2

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Low-mass X-ray binaries
• star-galaxy classification
• machine learning
• classification

Advertisement

• Find a journal
• Publish with us
• Track your research

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

• View all journals
• Explore content
• About the journal
• Publish with us
• Sign up for alerts
• Published: 23 January 2023

Welcome to the AI future?

Nature Astronomy volume  7 ,  page 1 ( 2023 ) Cite this article

6717 Accesses

1 Citations

33 Altmetric

Metrics details

• Communication

The launch of ChatGPT late last year has school teachers, conference organizers, Google and others worried, for different reasons. Where should we draw the line when it comes to artificial intelligence?

Artificial intelligence (AI) has the potential to revolutionize many fields, including astronomy. In recent years, astronomers have begun to increasingly rely on AI to analyse large datasets, identify patterns and make predictions. While the use of AI in astronomy has brought many benefits, it also comes with some pitfalls that need to be considered.

One of the major benefits of AI in astronomy is the ability to analyse and process vast amounts of data quickly and accurately. Astronomy generates huge amounts of data from telescopes, satellites and other instruments, and it can be overwhelming for humans to analyse all of this information. AI algorithms, on the other hand, can quickly and efficiently process this data, identifying patterns and trends that might not be immediately obvious to humans. This can save astronomers a lot of time and effort, and allow them to make new discoveries more quickly.

AI can also help astronomers make predictions and forecasts. For example, AI algorithms can be used to analyse data on the movements of celestial bodies and make predictions about their future behaviour. This can help astronomers plan observations and make more accurate forecasts about things like the likelihood of a comet hitting the Earth or the likelihood of a particular exoplanet being habitable.

In addition to these practical benefits, the use of AI in astronomy can also help to advance our understanding of the Universe. By identifying patterns and relationships in data that humans might not have noticed, AI can help to uncover new insights and theories about the way the Universe works. This can lead to new discoveries and a better understanding of the cosmos.

Despite these benefits, there are also some pitfalls to consider when using AI in astronomy. One of the main concerns is the accuracy of the predictions and forecasts made by AI algorithms. While these algorithms can be very accurate, they are only as good as the data they are trained on. If the data are incomplete or biased, the predictions made by the algorithms may also be flawed. This can lead to incorrect or misleading conclusions.

Another concern is the potential for AI to replace human astronomers. While AI algorithms can be very effective at analysing data and making predictions, they cannot replace the human ability to ask questions, make creative connections and think critically about the data. There is a risk that the reliance on AI could lead to a reduction in human creativity and curiosity in the field of astronomy.

In conclusion, the use of AI in astronomy has the potential to bring many benefits, including the ability to analyse and process large amounts of data, make predictions and forecasts, and advance our understanding of the Universe. However, it is important to be aware of the pitfalls of AI, including the risk of inaccurate predictions and the potential for it to replace human thinking and creativity. By being mindful of these potential pitfalls, astronomers can make the most of the benefits of AI while also maintaining the unique strengths of human intelligence.

The above text was written by ChatGPT, a conversational language-model chatbot launched by OpenAI on 30 November 2022. We asked it to “write a 500-word editorial on the benefits and pitfalls of AI to astronomy research” and it did so in a matter of seconds. The content is somewhat generic and the style formulaic, but it is at least correct if incomplete. And ChatGPT is very upfront about its own limitations, especially concerning training: it was trained on data up until the end of 2021. As an example, we also asked it to tell us the most exciting astronomy to look forward to in 2023. Highlights include:

The launch of the James Webb Space Telescope: This long-awaited successor to the Hubble Space Telescope is set to launch in March 2023, and it is expected to revolutionize our understanding of the Universe.

The arrival of the OSIRIS-REx spacecraft at asteroid Bennu: In 2023, the OSIRIS-REx spacecraft is set to arrive at asteroid Bennu, where it will collect a sample of surface material and return it to Earth for analysis.

The first manned mission to Mars: In 2023, NASA and SpaceX are set to launch the first manned mission to Mars, with the goal of establishing a permanent human presence on the red planet.

The potential discovery of new exoplanets: With the launch of the James Webb Space Telescope and the ongoing work of other telescopes and spacecraft, 2023 could be the year that we discover a plethora of new exoplanets.

The observation of a supermassive black hole: In 2023, astronomers will have the opportunity to observe a supermassive black hole in unprecedented detail, thanks to the Event Horizon Telescope. This global network of telescopes will allow astronomers to capture the first-ever image of a black hole, giving us a glimpse into the mysteries of these mysterious objects.

Aside from the exoplanet one — and arguably, JWST is mainly for exoplanet characterization rather than discovery — the predictions are all incorrect, by years. AI is a tool, after all. It is not a complete replacement for humans. Our Editorials will continue to be written by the editors. But by all means use AI to do tedious or time-consuming tasks. You might find more time to be creative or disruptive. And that is a good way to start 2023.

Rights and permissions

Reprints and permissions

About this article

Cite this article.

Welcome to the AI future?. Nat Astron 7 , 1 (2023). https://doi.org/10.1038/s41550-023-01891-4

Download citation

Published : 23 January 2023

Issue Date : January 2023

DOI : https://doi.org/10.1038/s41550-023-01891-4

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

Quick links

• Explore articles by subject
• Guide to authors
• Editorial policies

Sign up for the Nature Briefing: AI and Robotics newsletter — what matters in AI and robotics research, free to your inbox weekly.

We have completed maintenance on Astronomy.com and action may be required on your account. Learn More

• Login/Register
• Solar System
• Exotic Objects
• Upcoming Events
• Deep-Sky Objects
• Observing Basics
• Telescopes and Equipment
• Astrophotography
• Space Exploration
• Human Spaceflight
• Robotic Spaceflight
• The Magazine

How artificial intelligence is changing astronomy

When most people picture an astronomer, they think of a lone person sitting on top of a mountain, peering into a massive telescope. Of course, that image is out of date: Digital cameras have long since done away with the need to actually look though a telescope.

But now the face of astronomy is changing again. With the advent of more powerful computers and sky surveys that generate unimaginable quantities of data, artificial intelligence is the go-to tool for the keen researcher of space. But where is all of this data coming from? And how can computers help us learn about the universe?

AI’s appetite for data

Chances are you’ve heard the terms “artificial intelligence” and “machine learning” thrown around recently, and while they are often used together, they actually refer to different things. Artificial intelligence (AI) is a term used to describe any kind of computational behavior that mimics the way humans think and perform tasks. Machine learning (ML) is a little more specific: It’s a family of technologies that learn to make predictions and decisions based on vast quantities of historical data. Crucially, ML creates models which exhibit behavior that is not pre-programmed, but learned from the data used to train it.

The facial recognition in your smartphone, the spam filter in your emails, and the ability of digital assistants like Siri or Alexa to understand speech are all examples of machine learning being used in the real world. Many of these technologies are now being used by astronomers to investigate the mysteries of space and time. Astronomy and machine learning are a match made in the heavens, because if there’s one thing astronomers have too much of — and ML models can’t get enough of — it’s data.

We’re all familiar with megabytes (MB), gigabytes (GB), and terabytes (TB), but data at that scale is old news in astronomy. These days, we’re interested in petabytes (PB). A petabyte is about one thousand TB, a million GB, or a billion MB. It would take around 10 PB of storage to hold every single feature-length movie ever made in 4K resolution — and it would take over a hundred years to watch them all.

The Vera C. Rubin Observatory , a new telescope under construction in Chile, will be tasked with mapping the entire night sky in unprecedented detail , every single night. Over a 10-year survey, Vera Rubin will produce about 60 PB of raw data — studying everything from asteroids in our solar system, to galaxies in the distant universe. No human being could ever hope to analyze all that data — and that’s from just one of the next-generation observatories being built, so the race is on among astronomers in every field to find new ways to leverage the power of AI.

Planet hunters

One area of astronomy where AI has made a significant impact is in the search for exoplanets. There are many ways to look for their signals, but the most productive methods with current technology usually involve studying the variation of a star’s brightness over time. If a star’s light curve shows a characteristic dimming, it could be a sure sign of a planet transiting in front of the host star. Conversely, a phenomenon called gravitational microlensing can cause a large spike in a star’s brightness, when the exoplanet’s gravity acts as a lens and magnifies a more distant star along the line of sight. Detecting these dips and spikes means sifting through millions of light curves, studiously collected by space telescopes like NASA’s Kepler and TESS (Transiting Exoplanet Survey Satellite).

Using the huge libraries of observed light curves, astronomers have been able to develop ML-based models that can outperform humans in the search for exoplanets. But AI can do much more than just find exoplanets: It can also lead astronomers to new insights into how those techniques work .

In a paper published May 23 in Nature Astronomy , a team of researchers reported that ML algorithms had helped them discover a more elegant understanding of exoplanet microlensing, unifying multiple interpretations of how the exoplanet’s configuration with its host star might vary. The report came just months after researchers at DeepMind reported in Nature new AI-aided fundamental insights into mathematics .

Astronomers also hope that in the near future, machine learning will help them identify which planets might be habitable. Using next-generation observatories like the Nancy Grace Roman Telescope and James Webb Space Telescope (JWST), astronomers intend to use ML to detect water, ice, and snow on rocky planets .

Galactic forgeries

While many ML models are trained to distinguish between different types of data, others are intended to produce new data. These generative models are a subset of AI techniques that create artificial data products, such as images, based on some underlying understanding of the data used to train it.

The series of DALL-E models developed by the research company OpenAI — and the free-to-use imitator it inspired, DALL-E mini — have pushed this concept into the public eye. These models generate an image from any written prompt and have set the internet alight with their uncanny ability to construct images of, for instance, Garfield inserted into episodes of Seinfeld .

You might think that astronomers would be wary of any kind of fake imagery, but in recent years, researchers have turned to generative models in order to create galactic forgeries. A paper published Jan. 28 in Monthly Notices of the Royal Astronomical Society describes using the method to produce incredibly detailed images of fake galaxies, which can be used to test predictions from enormous simulations of the universe. They can also help develop and refine the data processing pipelines for next-generation surveys.

Some of these algorithms are so good that even professional astronomers can struggle to distinguish between the real and the fake. Take this recent entry into NASA’s Astronomy Picture of the Day webpage , which features dozens of synthetically generated images of objects in the night sky — and just one real image.

Searching for serendipity

AI is also primed to make discoveries that we cannot predict. There’s a long history of discoveries in astronomy that happened because someone was in the right place, at the right time. Uranus was discovered by chance when William Herschel was scanning the night sky for faint stars, Vesto Slipher measured the speed of spiral arms in what he thought were protoplanetary disks — eventually leading to the discovery of the expanding universe — and Jocelyn Bell Burnell’s famous detection of pulsars happened while she was analyzing measurements of quasars.

Perhaps soon, an AI could join these ranks of serendipitous discoverers though a field of techniques called anomaly detection . These algorithms are specifically trained to sift through mountains of images, light curves, and spectra, looking for the samples that don’t look like anything we’ve seen before. In the next generation of astronomy, with its petabytes of raw data from observatories like the Rubin and JWST, we can’t possibly imagine what these algorithms might find.

While millions watched the eclipse, these Purdue University scientists listened

The best telescopes if you want to spend somewhere from $90 to $11,000

The search is on for extraterrestrial life on worlds like Enceladus

2024 Full Moon calendar: Dates, times, types, and names

The clever prince Ulugh Beg built — and used — an ingenious observatory in the 15th century

The largest digital camera ever made for astronomy is done

NASA seeks faster, cheaper options to return Mars samples to Earth

Why good science is needed more than ever, and so is good faith

How to help astronomers study gamma-ray bursts.

Subscribe to the PwC Newsletter

Join the community, edit social preview.

Add a new code entry for this paper

Remove a code repository from this paper, mark the official implementation from paper authors, add a new evaluation result row.

• DECISION MAKING

Remove a task

Add a method, remove a method, edit datasets, applications of ai in astronomy.

3 Dec 2022  ·  S. G. Djorgovski , A. A. Mahabal , M. J. Graham , K. Polsterer , A. Krone-Martins · Edit social preview

We provide a brief, and inevitably incomplete overview of the use of Machine Learning (ML) and other AI methods in astronomy, astrophysics, and cosmology. Astronomy entered the big data era with the first digital sky surveys in the early 1990s and the resulting Terascale data sets, which required automating of many data processing and analysis tasks, for example the star-galaxy separation, with billions of feature vectors in hundreds of dimensions. The exponential data growth continued, with the rise of synoptic sky surveys and the Time Domain Astronomy, with the resulting Petascale data streams and the need for a real-time processing, classification, and decision making. A broad variety of classification and clustering methods have been applied for these tasks, and this remains a very active area of research. Over the past decade we have seen an exponential growth of the astronomical literature involving a variety of ML/AI applications of an ever increasing complexity and sophistication. ML and AI are now a standard part of the astronomical toolkit. As the data complexity continues to increase, we anticipate further advances leading towards a collaborative human-AI discovery.

Code Edit Add Remove Mark official

Tasks edit add remove, datasets edit, results from the paper edit add remove, methods edit add remove.

ARTIFICIAL INTELLIGENCE AT NORTHWESTERN

• Research & Applications
• Use of AI in Research Exploration

Use of AI in Astronomy

Northwestern's Center for Interdisciplinary Exploration and Research in Astrophysics (CIERA) is home to several groups at the forefront of AI research in astronomy. Vicky Kalogera 's group uses AI approaches to explore massive binary systems lives, properties, and fates. Adam Miller 's group works at the intersection of wide-field time-domain astronomical surveys and science, and uses AI techniques in the era of big data astronomy. Jason Wang 's group is developing AI image reconstruction techniques for the next generation of exoplanet imaging instruments. Gravity Spy , developed at CIERA, combines the crowd-sourcing power of citizen science with machine learning to classify glitches in gravitational-wave (GW) data. Glitches are transient bursts of noise, of a wide range of origins, which impact our ability to analyze GW data. Gravity Spy uses a convolutional neural net to classify glitches based on spectrogram images; citizen scientists perform a similar task, expanding our training set, and aiding classification of uncertain glitches. The machine learning algorithm allows us to rapidly classify glitches, producing near real-time reports for assessing candidate GW events, and enabling us to compile a large glitch database for study by detector experts. The Gravity Spy glitch classes are determined by experts, but as new detectors are commissioned, new glitches can arise. Gravity Spy empowers its citizen scientists to identify new glitch classes by allowing citizens to search for morphologically similar glitches to an interesting example. Volunteers have already identified multiple new glitch types, demonstrating that the public can make meaningful contributions to cutting-edge science. Learn more about CIERA's use of AI in astronomy

Learn more about Northwestern's use of AI in research exploration

Academia.edu no longer supports Internet Explorer.

To browse Academia.edu and the wider internet faster and more securely, please take a few seconds to  upgrade your browser .

Enter the email address you signed up with and we'll email you a reset link.

• We're Hiring!
• Help Center

Artificial Intelligence in Astronomy

2020, International Journal for Research in Applied Science and Engineering Technology -IJRASET

Artificial intelligence (AI) is a domain of computer science that emphasizes the creation of intelligent and smart machines that work and reacts like humans. Artificial Intelligence is a trendy field in computer science and it has been a boon to humans in many areas. A major contribution of AI is in the field of astronomy. With a vast range of data sets in astronomy, it becomes difficult for humans to have correct calculations and predictions. How these vast data sets can be effectively processed and analyzed is a challenging task. Here, Artificial Intelligence promises to deal with this vast data in an efficient and effective manner and provide with the accurate predictions. This paper summarizes the developments of artificial intelligence (AI) techniques, and examines their application to astronomy.

Related Papers

Theory and Practice of Natural Computing

Large archives of astronomical data (images, spectra and catalogues of derived parameters) are being assembled worldwide as part of the Virtual Observatory project. In order for such massive heterogeneous data collections to be of use to astronomers, development of Computational Intelligence techniques that would combine modern machine learning with deep domain knowledge is crucial. Both fields-Computer Science and Astronomy-can hugely benefit from such a research program. Astronomers can gain new insights into ...

Publications of the Astronomical Society of the Pacific

International Journal of Advanced Engineering Research and Applications

Pankhudi Saraswat

Machine learning is one of the most immensely growing field these days. It is gaining attention quickly and has people all around the world in awe of the tasks it can achieve. Machine and deep learning are a huge milestone while also being a great step in direction of technology as they can help us in achieving results with high precision. Astronomy is one such field that is reaching new levels with help of AI and machine learning. Through this paper, authors try to bridge the difference between the two fields and try to show how various machine learning algorithms have been implemented in the field of space research and have made work much easier and efficient. This article also addresses how deeply connected these two fields are how they can completely change the future of inventions and space information.

ESO ASTROPHYSICS SYMPOSIA

Roberto Tagliaferri , Antonino Staiano

harinder singh

Rosario De Rosa

Advances in Machine Learning and Data Mining for Astronomy documents numerous successful collaborations among computer scientists, statisticians, and astronomers who illustrate the application of state-of-the-art machine learning and data mining techniques in astronomy. Due to the massive amount and complexity of data in most scientific disciplines, the material discussed in this text transcends traditional boundaries between various areas in the sciences and computer science. The books introductory part provides context to issues in the astronomical sciences that are also important to health, social, and physical sciences, particularly probabilistic and statistical aspects of classification and cluster analysis. The next part describes a number of astrophysics case studies that leverage a range of machine learning and data mining technologies. In the last part, developers of algorithms and practitioners of machine learning and data mining show how these tools and techniques are use...

Abstracts Accepted for Publication

Natalia Shevchenko

Alexander Gorina Sánchez

The work exemplifies a way to enhance the research formation of students of Computer Science through the application of a didactic procedures system created by the authors themselves. The exemplification is distinguished by simulating the resolution process carried out by professors and students in a practical class, using a problematic situation, with the purpose of serving as a guide in the preparation and development of said classes. This process is conducted by the teacher through the three procedures of the system, which are: interpretive of the user system, translator of the intermediary and valorative system of the computer information system. It was concluded that the system of didactic procedures, exemplified, influences positively in the development of investigative skills.

RELATED PAPERS

Jasmin Peskoller

Scientific Investigation in Dentistry

Geraldo Griza

Journal of Neurology

jane Newman

Arch Mainous

El Carnaval de los Desencuentros, la construcción de los derechos sexuales y reproductivos en adolescentes de Bogotá

Liz Rincón Suárez

American Studies

The Journal of Chemical Thermodynamics

Deepak Ekka

Arindam Chakrabarti

FEMS Microbiology Ecology

Seshagiri Raghukumar

Molecular &amp; Cellular Proteomics

Leticia Rebollo Couto

RELACult - Revista Latino-Americana de Estudos em Cultura e Sociedade

Gabriele Vieira Neves

tahleho Tseole

ACTIO Journal of Technology in Design, Film Arts and Visual Communication

juan miceli

Natalia Kryzhanovskaya

Gut Pathogens

Marcel Behr

Brazilian Journal of Microbiology

Ulisses de Padua Pereira

Informatik-Spektrum

Uwe Schöning

Asian Journal of Research in Business Economics and Management

Kirtikumar Ekhande

Dinda Dwi julita

hyutrTT hytutr

arXiv (Cornell University)

Pavlos Efraimidis

Defect and Diffusion Forum

philippe azais

RELATED TOPICS

•   We're Hiring!
•   Help Center
• Find new research papers in:
• Health Sciences
• Earth Sciences
• Cognitive Science
• Mathematics
• Computer Science
• Academia ©2024

share this!

April 17, 2023

This article has been reviewed according to Science X's editorial process and policies . Editors have highlighted the following attributes while ensuring the content's credibility:

fact-checked

peer-reviewed publication

trusted source

Astronomers create AI to better communicate their stellar research

by University of Manchester

An international team of scientists, led by a researcher at The University of Manchester, have developed a novel AI (artificial intelligence) approach to distill technical astronomy terminology into simple understandable English in their recent publication.

The new research is a result of the international RGZ EMU (Radio Galaxy Zoo EMU) collaboration and is transitioning radio astronomy language from specific terms, such as FRI (Fanaroff-Riley Type 1), to plain English terms such as "hourglass" or "traces host galaxy."

The paper is published in the journal Monthly Notices of the Royal Astronomical Society .

In astronomy, technical terminology is used to describe specific ideas in efficient ways that are easily understandable among professional astronomers. However, this same terminology can also become a barrier to including non-experts in the conversation. The RGZ EMU collaboration is building a project on the Zooniverse citizen science platform, which asks the public for help in describing and categorizing galaxies imaged through a radio telescope.

Modern astronomy projects collect so much data that it is often impossible for scientists to look at it all by themselves, and a computer analysis can still miss interesting things easily spotted by the human eye.

Micah Bowles, Lead author and RGZ EMU data scientist, said, "Using AI to make scientific language more accessible is helping us share science with everyone. With the plain English terms we derived, the public can engage with modern astronomy research like never before and experience all the amazing science being done around the world."

Radio telescopes work in a very similar way to satellite dishes, but instead of picking up television signals they can be used to pick up the radio light generated by very energetic astrophysical objects—such as black holes in other galaxies. For many decades, these " radio galaxies " have been categorized into different types by astronomers to help them understand the origins and evolution of the universe.

Recently, dramatic improvements to radio telescopes around the world have revealed more and more of these radio galaxies, not only making it impossible for professional astronomers to look at each one individually and categorize it, but also introducing new variations that aren't already captured by existing radio galaxy types. Instead of trying to invent more and more new technical terminology for different types of radio galaxy—and train people to recognize them—the RGZ EMU team saw a different path forward that would enable citizen scientists to participate more fully in their research project.

The RGZ EMU team first asked experts to describe a selection of radio galaxies with their technical terms, and then asked non-experts to describe them in plain English. Using a first-of-its-kind AI-based approach they had developed, they then identified the plain English descriptions that carried the most scientific information. These descriptions("tags") can now be used by anyone to describe radio galaxies—in a way which is meaningful for any English speaker—without any specialist training at all. This work will not only be crucial for the RGZ EMU project, but with ever-increasing volumes of data across many areas of science, this new AI approach could find use in many more situations where simplified language can accelerate research, collaboration and communication.

Led from Manchester, this research was conducted by researchers from the UK, China, Germany, the U.S., the Netherlands, Australia, Mexico, and Pakistan. The data, code and results are all available online .

Journal information: Monthly Notices of the Royal Astronomical Society

Provided by University of Manchester

Explore further

Feedback to editors

Baby white sharks prefer being closer to shore, scientists find

24 minutes ago

Key protein regulates immune response to viruses in mammal cells

4 hours ago

Unraveling the mysteries of consecutive atmospheric river events

7 hours ago

Research team resolves decades-long problem in microscopy

RNA's hidden potential: New study unveils its role in early life and future bioengineering

8 hours ago

Smoother surfaces make for better accelerators

Scientists reveal hydroclimatic changes on multiple timescales in Central Asia over the past 7,800 years

Research reveals a surprising topological reversal in quantum systems

9 hours ago

NASA's Juno gives aerial views of mountain and lava lake on Io

Toxic fireproof chemicals can be absorbed through touch, 3D-printed skin model shows

Relevant physicsforums posts, our beautiful universe - photos and videos.

3 hours ago

Orientation of the Earth, Sun and Solar System in the Milky Way

13 hours ago

The linear polarization and brightness of pulsars

16 hours ago

Will we ever communicate with extraterrestial life in a reasonable time frame?

17 hours ago

Recommendations for International Research Competitions

Apr 18, 2024

Where Are the Missing Black Holes in the Milky Way?

Apr 16, 2024

More from Astronomy and Astrophysics

Related Stories

Astronomers discover fast radio bursts that skewer nearby galaxy

Apr 12, 2023

Help find the location of newly discovered black holes in the LOFAR Radio Galaxy Zoo project

Feb 26, 2020

Untangling a knot of galaxy clusters

Jan 30, 2023

Researchers study size, distribution and densities of intergalactic hydrogen clouds

Jan 13, 2023

Astronomers may have uncovered how galaxies change their shape

Dec 16, 2022

Astronomers discover ghosts of supernovas in nearby galaxy

Oct 30, 2019

Recommended for you

Researchers investigate three star-forming regions, identify hundreds young stellar objects

14 hours ago

Hubble goes hunting for small main belt asteroids

12 hours ago

Astronomers discover the most metal-poor extreme helium star

Apr 17, 2024

Researchers unveil unique tidal disruption event with pronounced early optical bump

NASA observations find what helps heat roots of 'moss' on sun

Researchers propose new formation model for massive hot subdwarfs

Let us know if there is a problem with our content.

Use this form if you have come across a typo, inaccuracy or would like to send an edit request for the content on this page. For general inquiries, please use our contact form . For general feedback, use the public comments section below (please adhere to guidelines ).

Please select the most appropriate category to facilitate processing of your request

Thank you for taking time to provide your feedback to the editors.

Your feedback is important to us. However, we do not guarantee individual replies due to the high volume of messages.

E-mail the story

Your email address is used only to let the recipient know who sent the email. Neither your address nor the recipient's address will be used for any other purpose. The information you enter will appear in your e-mail message and is not retained by Phys.org in any form.

Newsletter sign up

Get weekly and/or daily updates delivered to your inbox. You can unsubscribe at any time and we'll never share your details to third parties.

More information Privacy policy

Donate and enjoy an ad-free experience

We keep our content available to everyone. Consider supporting Science X's mission by getting a premium account.

E-mail newsletter

AI for Earth: How NASA’s Artificial Intelligence and Open Science Efforts Combat Climate Change

As extreme weather events increase around the world due to climate change, the need for further research into our warming planet has increased as well. For NASA, climate research involves not only conducting studies of these events, but also empowering outside researchers to do the same. The artificial intelligence (AI) efforts spearheaded by the agency offer a powerful tool to accomplish these goals.

In 2023, NASA teamed up with IBM Research to create an AI geospatial foundation model. Trained on vast amounts of NASA’s widely used Harmonized Landsat and Sentinel-2 (HLS) data, the model provides a base for a variety of AI-powered studies to tackle environmental challenges. In keeping with open science principles, the model is freely available for anyone to access .

Foundation models serve as a baseline from which scientists can develop a diverse set of applications, enabling powerful and efficient solutions. “Foundation models only know what things are represented in the data,” explained Manil Maskey, the data science lead at NASA’s Office of the Chief Science Data Officer (OCSDO). “It’s like a Swiss Army Knife—it can be used for multiple different things.”

Once a foundation model is created, it can be trained on a small amount of data to perform a specific task. To date, the Interagency Implementation and Advanced Concept Team (IMPACT) along with collaborators have demonstrated the geospatial foundation model’s capabilities by fine-tuning it to detect burn scars, to delineate flood water, and to classify crop and other land use categories.

Because of the computational resources required to create the initial foundation model, a partnership was necessary for success. In this case, NASA brought the data and scientific knowledge, while IBM brought the computing power and AI algorithm optimization expertise. The team’s shared commitment to making their research accessible through open science principles ensures that their model can be useful to as many researchers as possible.

“To build a foundation model at scale, we realized early on that it's not feasible for one institution to build it,” Maskey said. “Everything we have done on our foundation models has been open to the public, all the way from pre-training data, code, best practices, model weights, fine-tuning training data, and publications. There’s transparency, so researchers can trace why certain things were used in terms of data or model architecture.”

Following on from the success of their geospatial foundation model, NASA and IBM Research are continuing their partnership to create a new, similar model for weather and climate studies. They are collaborating with Oak Ridge National Laboratory (ORNL), NVIDIA, and several universities to bring this model to life.

This time, the main dataset will be the Modern-Era Retrospective analysis for Research and Applications, Version 2 (MERRA-2) , a huge collection of atmospheric reanalysis data that spans from 1980 to the present day. Like the geospatial foundation model, the weather and climate model is being developed with an open science approach, and will be available to the public in the near future.

Covering all aspects of Earth science would take several foundation models trained on different types of datasets. However, Maskey believes those future models might someday be combined into one comprehensive model, leading to a “digital twin” of the Earth that would provide unparalleled analysis and predictions for all kinds of climate and environmental events.

Whatever innovations the future holds, NASA and IBM’s geospatial and climate foundation models will enable leaps in Earth science like never before. Though powerful AI tools will enhance researchers’ work, the team’s dedication to open science supercharges the possibilities for discovery by allowing anyone to put those tools into practice and pave the way for groundbreaking research to help better care for the planet.

For more information about open science at NASA, visit: https://science.nasa.gov/open-science/

By Lauren Leese Web Content Strategist for the Office of the Chief Science Data Officer

Related Terms

Explore more.

Climate Change Research

Science in Space: April 2024 Everyone on Earth is touched by the effects of climate change, such as hotter temperatures, shifts in rain patterns, and sea level rise. Collecting climate data helps communities better plan for these changes and build more resilience to them. The International Space Station, one of dozens of NASA missions contributing […]

Tech Today: Taking Earth’s Pulse with NASA Satellites

Natural disasters like volcanic eruptions, floods, and tornados can dramatically change the surface of Earth to the point where alterations are visible in space. Changes driven by human actions and interventions, such as mining and deforestation, are also visible in satellite imagery. For over 50 years, NASA’s Landsat satellites have recorded our planet’s changing surface. […]

Kate A. McGinnis: Ready to “Go” with PACE Testing

PACE Systems Engineer Kate A. McGinnis spent years planning the vibration, acoustic, and shock tests for the PACE spacecraft before saying “Systems is ‘go’” to start each test. Name: Kate A. McGinnisTitle: Plankton, Aerosol, Cloud, and Ocean Ecosystem Mission (PACE) Systems EngineerFormal Job Classification: Systems engineerOrganization: Mission Systems Engineering Branch, Engineering and Technology Directorate (Code […]

AI Index Report

Welcome to the seventh edition of the AI Index report. The 2024 Index is our most comprehensive to date and arrives at an important moment when AI’s influence on society has never been more pronounced. This year, we have broadened our scope to more extensively cover essential trends such as technical advancements in AI, public perceptions of the technology, and the geopolitical dynamics surrounding its development. Featuring more original data than ever before, this edition introduces new estimates on AI training costs, detailed analyses of the responsible AI landscape, and an entirely new chapter dedicated to AI’s impact on science and medicine.

Read the 2024 AI Index Report

The AI Index report tracks, collates, distills, and visualizes data related to artificial intelligence (AI). Our mission is to provide unbiased, rigorously vetted, broadly sourced data in order for policymakers, researchers, executives, journalists, and the general public to develop a more thorough and nuanced understanding of the complex field of AI.

The AI Index is recognized globally as one of the most credible and authoritative sources for data and insights on artificial intelligence. Previous editions have been cited in major newspapers, including the The New York Times, Bloomberg, and The Guardian, have amassed hundreds of academic citations, and been referenced by high-level policymakers in the United States, the United Kingdom, and the European Union, among other places. This year’s edition surpasses all previous ones in size, scale, and scope, reflecting the growing significance that AI is coming to hold in all of our lives.

Steering Committee Co-Directors

Ray Perrault

Steering committee members.

Erik Brynjolfsson

John Etchemendy

Katrina Ligett

Terah Lyons

James Manyika

Juan Carlos Niebles

Vanessa Parli

Yoav Shoham

Russell Wald

Staff members.

Loredana Fattorini

Nestor Maslej

Letter from the co-directors.

A decade ago, the best AI systems in the world were unable to classify objects in images at a human level. AI struggled with language comprehension and could not solve math problems. Today, AI systems routinely exceed human performance on standard benchmarks.

Progress accelerated in 2023. New state-of-the-art systems like GPT-4, Gemini, and Claude 3 are impressively multimodal: They can generate fluent text in dozens of languages, process audio, and even explain memes. As AI has improved, it has increasingly forced its way into our lives. Companies are racing to build AI-based products, and AI is increasingly being used by the general public. But current AI technology still has significant problems. It cannot reliably deal with facts, perform complex reasoning, or explain its conclusions.

AI faces two interrelated futures. First, technology continues to improve and is increasingly used, having major consequences for productivity and employment. It can be put to both good and bad uses. In the second future, the adoption of AI is constrained by the limitations of the technology. Regardless of which future unfolds, governments are increasingly concerned. They are stepping in to encourage the upside, such as funding university R&D and incentivizing private investment. Governments are also aiming to manage the potential downsides, such as impacts on employment, privacy concerns, misinformation, and intellectual property rights.

As AI rapidly evolves, the AI Index aims to help the AI community, policymakers, business leaders, journalists, and the general public navigate this complex landscape. It provides ongoing, objective snapshots tracking several key areas: technical progress in AI capabilities, the community and investments driving AI development and deployment, public opinion on current and potential future impacts, and policy measures taken to stimulate AI innovation while managing its risks and challenges. By comprehensively monitoring the AI ecosystem, the Index serves as an important resource for understanding this transformative technological force.

On the technical front, this year’s AI Index reports that the number of new large language models released worldwide in 2023 doubled over the previous year. Two-thirds were open-source, but the highest-performing models came from industry players with closed systems. Gemini Ultra became the first LLM to reach human-level performance on the Massive Multitask Language Understanding (MMLU) benchmark; performance on the benchmark has improved by 15 percentage points since last year. Additionally, GPT-4 achieved an impressive 0.97 mean win rate score on the comprehensive Holistic Evaluation of Language Models (HELM) benchmark, which includes MMLU among other evaluations.

Although global private investment in AI decreased for the second consecutive year, investment in generative AI skyrocketed. More Fortune 500 earnings calls mentioned AI than ever before, and new studies show that AI tangibly boosts worker productivity. On the policymaking front, global mentions of AI in legislative proceedings have never been higher. U.S. regulators passed more AI-related regulations in 2023 than ever before. Still, many expressed concerns about AI’s ability to generate deepfakes and impact elections. The public became more aware of AI, and studies suggest that they responded with nervousness.

Ray Perrault Co-director, AI Index

Our Supporting Partners

Analytics & Research Partners

Stay up to date on the AI Index by subscribing to the  Stanford HAI newsletter.

To revisit this article, visit My Profile, then View saved stories .

• Backchannel
• Newsletters
• WIRED Insider
• WIRED Consulting

Amanda Hoover

Students Are Likely Writing Millions of Papers With AI

Students have submitted more than 22 million papers that may have used generative AI in the past year, new data released by plagiarism detection company Turnitin shows.

A year ago, Turnitin rolled out an AI writing detection tool that was trained on its trove of papers written by students as well as other AI-generated texts. Since then, more than 200 million papers have been reviewed by the detector, predominantly written by high school and college students. Turnitin found that 11 percent may contain AI-written language in 20 percent of its content, with 3 percent of the total papers reviewed getting flagged for having 80 percent or more AI writing. (Turnitin is owned by Advance, which also owns Condé Nast, publisher of WIRED.) Turnitin says its detector has a false positive rate of less than 1 percent when analyzing full documents.

ChatGPT’s launch was met with knee-jerk fears that the English class essay would die . The chatbot can synthesize information and distill it near-instantly—but that doesn’t mean it always gets it right. Generative AI has been known to hallucinate , creating its own facts and citing academic references that don’t actually exist. Generative AI chatbots have also been caught spitting out biased text on gender and race . Despite those flaws, students have used chatbots for research, organizing ideas, and as a ghostwriter . Traces of chatbots have even been found in peer-reviewed, published academic writing .

Teachers understandably want to hold students accountable for using generative AI without permission or disclosure. But that requires a reliable way to prove AI was used in a given assignment. Instructors have tried at times to find their own solutions to detecting AI in writing, using messy, untested methods to enforce rules , and distressing students. Further complicating the issue, some teachers are even using generative AI in their grading processes.

Detecting the use of gen AI is tricky. It’s not as easy as flagging plagiarism, because generated text is still original text. Plus, there’s nuance to how students use gen AI; some may ask chatbots to write their papers for them in large chunks or in full, while others may use the tools as an aid or a brainstorm partner.

Students also aren't tempted by only ChatGPT and similar large language models. So-called word spinners are another type of AI software that rewrites text, and may make it less obvious to a teacher that work was plagiarized or generated by AI. Turnitin’s AI detector has also been updated to detect word spinners, says Annie Chechitelli, the company’s chief product officer. It can also flag work that was rewritten by services like spell checker Grammarly, which now has its own generative AI tool . As familiar software increasingly adds generative AI components, what students can and can’t use becomes more muddled.

Detection tools themselves have a risk of bias. English language learners may be more likely to set them off; a 2023 study found a 61.3 percent false positive rate when evaluating Test of English as a Foreign Language (TOEFL) exams with seven different AI detectors. The study did not examine Turnitin’s version. The company says it has trained its detector on writing from English language learners as well as native English speakers. A study published in October found that Turnitin was among the most accurate of 16 AI language detectors in a test that had the tool examine undergraduate papers and AI-generated papers.

Will Knight

Stephen Clark, Ars Technica

Emily Mullin

Andrew Couts

Schools that use Turnitin had access to the AI detection software for a free pilot period, which ended at the start of this year. Chechitelli says a majority of the service’s clients have opted to purchase the AI detection. But the risks of false positives and bias against English learners have led some universities to ditch the tools for now. Montclair State University in New Jersey announced in November that it would pause use of Turnitin’s AI detector. Vanderbilt University and Northwestern University did the same last summer.

“This is hard. I understand why people want a tool,” says Emily Isaacs, executive director of the Office of Faculty Excellence at Montclair State. But Isaacs says the university is concerned about potentially biased results from AI detectors, as well as the fact that the tools can’t provide confirmation the way they can with plagiarism. Plus, Montclair State doesn’t want to put a blanket ban on AI, which will have some place in academia. With time and more trust in the tools, the policies could change. “It’s not a forever decision, it’s a now decision,” Isaacs says.

Chechitelli says the Turnitin tool shouldn’t be the only consideration in passing or failing a student. Instead, it’s a chance for teachers to start conversations with students that touch on all of the nuance in using generative AI. “People don’t really know where that line should be,” she says.

You Might Also Like …

In your inbox: The best and weirdest stories from WIRED’s archive

Jeffrey Epstein’s island visitors exposed by data broker

8 Google employees invented modern AI. Here’s the inside story

The crypto fraud kingpin who almost got away

It's shadow time! How to view the solar eclipse, online and in person

Benj Edwards, Ars Technica

Makena Kelly

Matt Burgess

Lauren Goode

Reece Rogers

Steven Levy

More From Forbes

5 tips to enhance your research paper’s visibility and altmetric score.

• Share to Facebook
• Share to Twitter
• Share to Linkedin

US evangelist Billy Graham addresses a crowd of football supporters at Stamford Bridge, London, ... [+] during half-time at the match between Chelsea and Newcastle United. (Photo by Edward Miller/Getty Images)

I previously wrote about the importance of attracting public attention to scientific research . In today’s world, where billions of people are attached to their digital devices watching the very addictive but often useless TikTok content or receiving instant gratification by engaging in meaningless debates about celebrities, scientists need to find creative ways to have their research noticed. Popularizing scientific research helps inspire the younger generations to go into science and provide the general public with a sense of optimism enabling the government to channel more resources into science. People do need inspiration. But very often, even very important scientific breakthroughs requiring many years, hard work, skill, funding, and genuine serendipity go largely unnoticed by the general public.

One of the best ways to measure expert and public attention is the cumulative Altmetric Attention Score , originally developed by Digital Science and adopted by many prestigious publishers, including Nature Publishing Group. Every Nature paper and the papers published by pretty much every credible publisher are tracked by Digital Science by the Document Object Identification (DOI) or the Unique Resource Locator (URL) . While Altmetric has many limitations, for example, it does not track LinkedIn posts and may not adequately cover the impact of top-tier media coverage, at the moment it is the blueprint for tracking attention.

Altmetric Score in The Age of Generative AI

Media attention is likely to be very important in the age of generative AI. Many modern generative systems, such as ChatGPT, Claude, Mistral, and Gemini, as well as hundreds of Large Language Models (LLMs) in China, use the data from the same sources referenced in Altmetric to learn. The more times generative systems see the same concept presented in different contexts, the better they learn. So if you want to contribute to the training of AI systems that may thank you for it in the future - Altmetric is the way to go.

So what can a research group do to ensure they are communicating their findings effectively and increasing the visibility of their research to ensure it gets reflected in the Altmetric Attention Score?

JPMorgan Joins Goldman Sachs In Serious Bitcoin Halving Price Warning

Google makes a major new sale offer to pixel 8 buyers, the crazy true story behind netflix s baby reindeer what happens to martha.

Altmetric openly discloses the weights of the various sources and the scoring algorithm is relatively straightforward. It is easy to learn, and there are multiple online resources providing advice on how to share your research in ways that will be captured by Altmetric. Cambridge University Press published a guideline to Altmetric for the authors on how to popularize their research with Altmetric in mind. Wolters Kluwer put out a guide and the editor of Toxicology and Pathology wrote a comprehensive overview of Altmetric and how to use it. Surprisingly, this overview got an Altmetric Attention Score of only 4 at the time of the writing, but was cited 137 times according to Google Scholar .

Altmetric monitors social networks, including X (formerly Twitter), Facebook, and Reddit; all major top-tier mainstream media, mainstream science blogs, policy documents, patents, Wikipedia articles, peer review websites, F1000, Syllabi, X (formerly Twitter), tracked Facebook pages, Reddit, one of the Stack Exchange sites, and Youtube. Unfortunately, several powerful platforms, including LinkedIn, are not currently tracked.

The popularity of the paper depends on many factors. Firstly, it has to be novel, trendy, and newsworthy. You are unlikely to get high Altmetric Score with a boring topic. Secondly, papers coming out of popular labs in top-tier academic institutions and in top journals are likely to attract more attention. Often, the communications officers in these academic institutions work closely with the media to amplify notable research. Celebrity companies, for example, Google DeepMind, consistently get higher coverage.

Screenshot of the Altmetric Attention Score "Flower" showing several tracked sources

Here are the five tips for increasing the visibility of your work and ensuring that reach is tracked and reflected by Altmetric:

1. Understand How Altmetric System Works

Congratulations, if you read this article and looked at what sources are tracked by Altmetric. Most likely, you got the basics and will be able to get a “balanced flower” by making a press release, tweeting the DOI of the paper on X, posting a video overview of your paper on Youtube, announcing on Reddit (I still need to learn how to do this).

To understand how Altmetric works, I emailed a few questions to Miguel Garcia, Director of Product and Data Analytics Hub at Digital Science and my first question was wether the Altmetric algorithm is open source. “The Altmetric Attention Score's calculation is not open source but we try to provide as much information as possible around how we calculate it here, and are currently considering what steps we might take to make our algorithms more transparent.” He also provided a link to how the Altmetric Attention Score is calculated.

Many professionals use LinkedIn as the primary social media resource and I was wondering why Altmetric stopped tracking it. Bad news - technical reasons prevent tracking DOIs on LinkedIn. Good news - they are actively seeking ways to appropriately track mentions on LinkedIn and we may see some news toward the end of the year.

My other big question was how does Altmetric count tweets and retweets on X. What if there are many posts from the same account? Miguel’s response was: “Re-tweets count less than original tweets. In addition to that, modifiers are applied to the type of account that is tweeting in order to reduce the weight of the tweet in situations where we find signals of bias or promiscuity (for example a journal publisher only tweeting their own articles). Besides that, we have conditions around the maximum number of retweets in order to limit the maximum impact they would have.”

So tweeting the article many times will not help you. But if other scientists tweet you paper with a DOI - these tweets will get counted. So tweet others as you would like to be tweeted.

2. Make a Press Release and Distribute to Science-focused Media

If your paper is significant, for example, you elucidated novel disease biology, discovered a new drug, developed a new fancy algorithm, designed a new material, or developed a new application for a quantum computer, it is worthwhile investing some time and resources in writing a press release. If you are working for an academic institution, most likely they have a communications office that will help you. If you do not have this luxury, you will need to learn how to write a press release. Plenty of free online guides cover the basics of press release writing. And press releases are one area where ChatGPT and other generative tools do surprisingly well. Upload your paper and ask it to write a press release, check for errors or exaggerations, edit, and you are ready to go. Just make sure to include the DOI and the URL of your paper. A proper business press release on BusinessWire or PRNewswire may cost several thousand dollars. In my opinion, these resources are dramatically overcharging while providing little service. I don't remember a case where a journalist picked up our news based on a commercial press release. But these releases are often reposted by other online press release distributors and the boost to Altmetric may be considerable. The default news release distribution service for research news is EurekAlert. This resource may sometimes result in journalistic coverage as many reporters are using it for science news. There are many free resources you can use if you do not have any budget.

Once the press release is issued, share it with the media. Share the resulting news coverage via your social networks and contacts. Many journalists track the popularity of their news articles and giving them several thousand extra views from professional audience and increasing their social following increases the chances that they will cover the next important research paper.

3. Make a Blog Post

Writing a blog post can be longer and more comprehensive than the press release. Make sure to add fancy diagrams and graphical explainers. You can share the blog post with the journalists at the same time as the press release. Your blog may serve as a source of inspiration for third party news coverage. Make sure to reference the DOI and URL of your paper.

If your paper is in one of the Nature journals, consider writing a “Behind the Paper" blog post on Nature Bioengineering Community. Surprisingly, these blogs are rarely picked up by Altmetric but may serve as a source of inspiration for the journalists and social media influencers. Plus, it is a resource by the Nature Publishing Group.

4. Tweet and Ask Your Team Members to Tweet

Each post on X gives you a quarter of an Altmetric point. If your paper goes viral on X, your Altmetric score can be considerable. Plus, once journalists notice that it went viral, they will be more likely to cover the story, further increasing the score.

Try to choose the time of the post, the hashtags, and the images wisely. Since Elon Musk took over X and opened the algorithm it became very transparent and easy to optimize for. Here are the top 10 tips for boosting attention for a post on X. Make sure to include the DOI or the URL of the paper for Altmetric to find the post.

5. Experiment, Learn, Repeat

My highest Altmetric Attention Score core to date was around 1,500 for a paper in Nature Biotechnology published in 2019, where we used a novel method for designing small molecules called Generative Tensorial Reinforcement Learning (GENTRL) to generate new molecules with druglike properties that got synthesized and tested all the way into mice. In 2024, we went further and showed that an AI-generated molecule for an AI-discovered target was tested all the way up to Phase II human trials, but the paper published in Nature Biotechnology, let’s call it the TNIK paper , has achieved a score ofjust over 600 to date. So what has changed and what can we learn from these two papers?

The popularity of the paper depends on many factors. Ones which capture the public imagination or have widespread appeal are of course, much more likely to gain traction online. When we published the GENTRL paper in 2019, Generative AI was in its infancy, and there are pretty much no other companies that I heard of at the intersection of generative AI and drug discovery. We also published multiple articles in this field in the years leading to that paper and many key opinion leaders (KOLs) followed us. That following included a small army of generative AI skeptics who not only contributed to multiple rejections in peer-reviewed journals but also openly criticized this approach in social networks. This criticism also helped boost the Altmetric Score and bring more attention to the study. So first learning from this exercise - negative publicity helps overall publicity. As long as you are certain that your research results are honest - leave room for criticism and even help expose your paper’s weaknesses. Critics are your greatest Altmetric boosters. Since readers and, by extension journalists, react to negative news and drama stronger than to positive news, critical reviews will boost your Altmetric as long as the DOI or URL of the paper is properly referenced.

Secondly, papers coming out of popular labs in top-tier academic institutions and in top journals are likely to attract more attention. Often, the communications officers in these academic institutions work closely with the media to amplify notable research. Celebrity companies, for example, Google DeepMind, always get a higher level of coverage. For example, the AlphaFold paper published in July 2021 in Nature got an Altmetric Attention Score of over 3,500 . Even though I have not seen any drugs out of AlphaFold reaching preclinical candidate status, I predict the popularity of this tool will result in the first Nobel Prize in this area. Therefore, in order to become famous and popularize your research more effectively, it is a good idea to build up the public profile of yourself and your work. For example, Kardashians are famous for being famous .

Be careful with Wikipedia. I made a mistake explaining the importance of Wikipedia to students when lecturing on the future of generative AI, and one or two of them got banned for expanding the articles with paper references. Wikipedia requires that these are added by independent editors rather than the authors of papers themselves, but if some editors do not like it, they will not go deep or investigate - they will assume wrongdoing. So it is better to avoid even talking about Wikipedia. References there should happen naturally and often some of the more popular papers get picked up and referenced by veteran editors.

Experimenting with Altmetric will also help you explore new strategies for popularizing scientific research and develop new strategies for inspiring people to learn or even get into the new exciting field. UNESCO estimates that there was just over 8 million full-time equivalent (FTE) researchers in 2018 globally. Only a fraction of these are in biotechnology - less than 0.01% of the global population. If you motivate a million students to go into biotechnology by popularizing your research, you double this number.

• Editorial Standards
• Reprints & Permissions

This paper is in the following e-collection/theme issue:

Published on 17.4.2024 in Vol 26 (2024)

Correction: Application of AI in Multilevel Pain Assessment Using Facial Images: Systematic Review and Meta-Analysis

Authors of this article:

There are no citations yet available for this article according to Crossref .

Help | Advanced Search

Astrophysics > Instrumentation and Methods for Astrophysics

Title: astrollama: towards specialized foundation models in astronomy.

Abstract: Large language models excel in many human-language tasks but often falter in highly specialized domains like scholarly astronomy. To bridge this gap, we introduce AstroLLaMA, a 7-billion-parameter model fine-tuned from LLaMA-2 using over 300,000 astronomy abstracts from arXiv. Optimized for traditional causal language modeling, AstroLLaMA achieves a 30% lower perplexity than Llama-2, showing marked domain adaptation. Our model generates more insightful and scientifically relevant text completions and embedding extraction than state-of-the-arts foundation models despite having significantly fewer parameters. AstroLLaMA serves as a robust, domain-specific model with broad fine-tuning potential. Its public release aims to spur astronomy-focused research, including automatic paper summarization and conversational agent development.

Submission history

Access paper:.

• Other Formats

References & Citations

• INSPIRE HEP
• Google Scholar
• Semantic Scholar

BibTeX formatted citation

Bibliographic and Citation Tools

Code, data and media associated with this article, recommenders and search tools.

• Institution

arXivLabs: experimental projects with community collaborators

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs .