Hierarchical incremental learning deciphers multi-component materials

Identifying meaningful patterns of atomic and molecular arrangements from molecular simulations is crucial for revealing microscopic mechanisms in materials. Unraveling these patterns is challenging for the multi-component systems frequently encountered in advanced materials, energy and environmental applications. This limits the understanding of the microscopic mechanisms that ultimately govern the performance of devices based on these systems. Here, we propose a hierarchical incremental learning research protocol named HiDiscover to systematically expedite the mechanistic exploration in multi-component materials. As illustrations, we study Li-ion transport and gas adsorption in nanoporous framework materials, as well as molecular packing in organic active layers for photovoltaics. The HiDiscover protocol enables the detailed differentiation and facile extraction of ionic and molecular arrangements, and reveals quantitative microscopic features that are difficult to discern through conventional molecular simulations, thereby informing materials design. Our approach is seen to improve the reliability of mechanistic descriptions for three different processes in three different classes of materials. More broadly, this report highlights the potential of incremental learning to provide atomistic insight into complex materials and processes.