Hierarchical machine learning framework pushes boundaries of multi-property material discovery

Recent developments in artificial intelligence have broadened applications in the field of materials science. Still, relatively limited data, multiple desired properties, and data dispersion over multiple domains remain hurdles for using machine learning to accelerate the materials discovery.

Kong et al. present the hierarchical correlation learning for multiproperty prediction (H-CLMP) framework for predicting properties in new composition spaces.

By performing multitarget regression from the materials’ composition, the model learns element and property relationships at multiple scales. Learning property relationships in such a way allows the framework’s generative transfer learning approach to augment prediction in the target domain.

“A key contribution of our work is a novel way of transferring knowledge from theoretical work into the experimental setting, by developing a new type of generative model that quickly generates ‘pseudo’ data,” said author Carla Gomes. “The data encapsulates and transfers the prior scientific knowledge of theoretical physics-based computational models.”

The group demonstrated H-CLMP by predicting sunlight absorbing properties of new metal oxides. H-CLMP used training data from high-throughput experiments to predict the absorption spectra for combinations of three cations from 31 elements.

The group hopes the broader materials community will help push the framework in new ways as subfields use it.

“While we design models to be as general as possible, we often find new settings that challenge the state of the art,” said author John Gregoire. “Such challenges provide the impetus for generating new methods, which we will continue to approach by making connections among computing challenges in a broad range of scientific disciplines.”

Source: “Materials representation and transfer learning for multi-property prediction,” by Shufeng Kong, Dan Guevarra, Carla P. Gomes, and John M. Gregoire, Applied Physics Reviews (2021). The article can be accessed at https://doi.org/10.1063/5.0047066 .

This paper is part of AIP’s Autonomous Materials Science special. More papers can be found at https://aip.scitation.org/toc/are/collection/10.1063/are.2020.MLMAT2020.issue-1