Berlin 2024 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
O: Fachverband Oberflächenphysik
O 82: Electronic Structure Theory I
O 82.6: Talk
Thursday, March 21, 2024, 11:45–12:00, MA 043
Electronic excited states from physically-constrained machine learning — Edoardo Cignoni1, •Divya Suman2, Jigyasa Nigam2, Lorenzo Cupellini1, Benedetta Mennucci1, and Michele Ceriotti2 — 1Dipartimento di Chimica e Chimica Industriale, Università di Pisa, Pisa, Italy — 2Laboratory of Computational Science and Modeling (COSMO), IMX, École Polytechnique Fédérale de Lausanne, Switzerland
The integration of machine learning (ML) techniques with quantum mechanical (QM) calculations has opened new avenues in predicting a variety of electronic properties of molecules. Our work investigates a fundamental question, which is of great relevance to drive these developments: should machine learning (ML) be directly employed to predict desired properties or be synergistically combined with physically-grounded operations? To this end we introduce an integrated modelling approach in which we build a symmetry-adapted ML model that targets the properties of interest while learning the minimal-basis, single-particle electronic Hamiltonian as an intermediate. This approach also enables predictions of properties other than the ones used during the training process, like the molecular excited states. The resulting architecture, therefore, inherits the accuracy of QM calculations, as well as the transferability to larger, more complex molecules as well as a variety of ground and excited states properties, while being orders of magnitude faster compared to the traditional electronic structure methods.
Keywords: Machine learning; Hamiltonian; Excited states