Engineering atoms with tailored properties using physics-informed Kolmogorov– Arnold Networks

Research question

Design of new materials is often limited by the need to accurately predict the atomistic and electronic structure with minimal computational costs. The large number of microscopic model parameters and heavy load of numerical computations which are needed to obtain the data on macroscopic properties of the engineered materials require new developments from both the quantum chemistry methods and Machine Learning (ML) techniques. Combining embedding techniques from quantum chemistry with physics-informed Kolmogorov-Arnold networks (PI-KANs), physics-informed neural networks (PINNs) and active learning has a strong potential to override current limitations of the existing theoretical framework.  This project will create an open-source computational tool which can be used for accurate and efficient modeling of electronic structure of new materials.

The proposed approach is broadly applicable, but given the constraints of quantum chemical methods and the requirement for locality in the quantum region, typical applications include (i) light-harvesting systems (especially those with a central transition metal), (ii) solid-state electrolytes (where single ions move within the solid framework), (iii) solid-state lasers (formed as point defects of transition or lanthanides
atoms in ionic crystals), (iv) atomic clocks (formed by f-element point defects in ionic crystals), (v) biologically significant proteins (particularly those containing transition metals), and (vi) catalytically
active non-metallic surfaces.

Sustainability aspects

The development of new computational protocols for the electronic structure calculations of crystals will reduce the computational costs of such calculations.