A meta-machine-learning method for identifying effective descriptors of materials properties

Luca Ghiringhelli
Fritz-Haber-Institut der Max-Planck-Gesellschaft

Identifying the functional dependency of a target material property on the material composition and structure, the descriptor, is of fundamentally important for the mechanism understanding and novel material prediction. Yet, methods for descriptor
identification are still not well established. In this work, we present a systematic and efficient approach for identifying descriptors of either quantitative or qualitative categorical) material properties, within the framework of compressed sensing. We demonstrate the efficiency of our (meta-)method on a well-studied example of finding descriptors for predicting the crystal structure of octet binary materials via the quantitative property of energy difference. Finally, we apply this approach to find a descriptor for the validation classification of binary materials as metals, with all training data from experiment.

Back to Workshop I: Optimization and Optimal Control for Complex Energy and Property Landscapes