Machine Learning in Materials Science
Machine learning is a branch of artificial intelligence and is currently revolutionizing research practices in the natural sciences. Machine learning models are trained on materials data already available from experiments or computations by creating statistically optimized relationships between the given data. Once the model is trained sufficiently, it can make predictions for new materials or infer correlations with almost the same accuracy as the data generation method, but in only a fraction of the time and with a fraction of the computational or experimental effort. We currently pursue two main machine learning research lines: BOSS and ARTIST.
BOSS: Bayesian Optimization Structure Search is an active learning technique for global exploration of energy and property phase space, and for accelerated structure determination.
ARTIST: Artificial Intelligence for Spectroscopy is a suite of machine learning methods for excited states and spectral properties. We are exploring kernel ridge regression for individual excitation energies and neural networks for excitation spectra. We are also developing descriptors for atomistic representations. These descriptors are available in the DScribe Phython library.
Chemical diversity in molecular orbital energy predictions with kernel ridge regression, A. Stuke, M. Todorović, M. Rupp, C. Kunkel, K. Ghosh, L. Himanen, and P. Rinke, J. Chem. Phys. 150, 204121 (2019)
Deep Learning Spectroscopy: Neural Networks for Molecular Excitation Spectra, K. Ghosh, A. Stuke, M. Todorović, P. B. Jørgensen, M. N. Schmidt, A. Vehtari and P. Rinke, Adv. Sci. 6, 1801367 (2019)
DScribe: Library of descriptors for machine learning in materials science, L. Himanen, M. O. J. Jäger, E. V. Morooka, F. F. Canova, Y. S. Ranawat, D. Z. G