Abstract

Machine learning techniques often follow the end-to-end approach in that they estimate outputs of quantum mechanical calculations such as structural energies or dipole moments based on geometric descriptions of the underlying structure. Following a recent paradigm shift of blurring the distinction between explicit quantum mechanical and modeling steps, there has been an interest, instead, in machine learning the ingredients of electronic structure, such as the effective single-particle Hamiltonian from which properties of interest may be derived. In this talk, I will describe how we can leverage existing techniques in the framework of atom-centered density representations (ACDCs) to model electronic Hamiltonians¹. I will motivate the merits of this symbiotic integration of fundamental physical relations with data-driven methods, not only in terms of their accuracy and transferability, but also on their role in the prediction of more complex properties such as electronic excitations [2].

[1] J. Nigam, M. Willatt, M. Ceriotti, JCP 156 , 014115, 2022 [2] E. Cignoni, D. Suman, J. Nigam et al. arXiv:2311.00844 (Accepted)