Speakers
Description
Recent years have seen an explosion of work related to machine-learning (ML) models trained on electronic structure data. To push the boundaries of what can be done with physics-based ML, we have explored how ML can be directly incorporated into electronic structure calculations. This dispenses of the separation between electronic structure theory as a mere data generation device on one hand, and ML as a mathematical fitting framework on the other. In this case, the ML model is used to predict more fundamental physical relationships, rather than the full structure-property mapping. This has the advantage of making the model more transferable and data-efficient, since the mathematical framework of the electronic structure calculation is retained.
To this end, we recently reported a size-extensive and rotationally invariant ML representation based on the atomic decomposition of the electron density obtained from mean-field electronic structure calculations.[1] This was used to develop so-called Kernel Density Functional Approximations (KDFAs), trained on coupled cluster correlation energies. Importantly, KDFAs are pure functionals of the electron density that are nonetheless able to capture the non-local nature of the correlation energy. Currently, we are further developing this concept to obtain a self-consistent version of KDFA. This enables the prediction of electronic properties like the dipole moment and the density on the same footing as the energy and thus offers a route to CC-quality electron densities for systems of unprecedented size.
[1] Nat. Commun. 12, 344 (2021).
