Publications (2)0 Total impact
ABSTRACT: Observational astronomy has changed drastically in the last decade: manually driven target-by-target instruments have been replaced by fully automated robotic telescopes. Data acquisition methods have advanced to the point that terabytes of data are flowing in and being stored on a daily basis. At the same time, the vast majority of analysis tools in stellar astrophysics still rely on manual expert interaction. To bridge this gap, we foresee that the next decade will witness a fundamental shift in the approaches to data analysis: case-by-case methods will be replaced by fully automated pipelines that will process the data from their reduction stage, through analysis, to storage. While major effort has been invested in data reduction automation, automated data analysis has mostly been neglected despite the urgent need. Scientific data mining will face serious challenges to identify, understand and eliminate the sources of systematic errors that will arise from this automation. As a special case, we present an artificial intelligence (AI) driven pipeline that is prototyped in the domain of stellar astrophysics (eclipsing binaries in particular), current results and the challenges still ahead.
ABSTRACT: Achieving maximum scientific results from the overwhelming volume of astronomical data to be acquired over the next few decades will demand novel, fully automatic methods of data analysis. Artificial intelligence approaches hold great promise in contributing to this goal. Here we apply neural network learning technology to the specific domain of eclipsing binary (EB) stars, of which only some hundreds have been rigorously analyzed, but whose numbers will reach millions in a decade. Well-analyzed EBs are a prime source of astrophysical information whose growth rate is at present limited by the need for human interaction with each EB data-set, principally in determining a starting solution for subsequent rigorous analysis. We describe the artificial neural network (ANN) approach which is able to surmount this human bottleneck and permit EB-based astrophysical information to keep pace with future data rates. The ANN, following training on a sample of 33,235 model light curves, outputs a set of approximate model parameters (T2/T1, (R1+R2)/a, e sin(omega), e cos(omega), and sin i) for each input light curve data-set. The whole sample is processed in just a few seconds on a single 2GHz CPU. The obtained parameters can then be readily passed to sophisticated modeling engines. We also describe a novel method polyfit for pre-processing observational light curves before inputting their data to the ANN and present the results and analysis of testing the approach on synthetic data and on real data including fifty binaries from the Catalog and Atlas of Eclipsing Binaries (CALEB) database and 2580 light curves from OGLE survey data. [abridged] Comment: 52 pages, accepted to ApJ