Séminaire de l'équipe PCMT: Bastien Casier (Maître de conférences, Université d'Artois, Laboratoire UCCS)

Bastien Casier, Maître de Conférences, Université d'Artois, Laboratoire UCCS

Résumé:

Quantum chemical calculations play a central role in many areas of chemistry, yet accurate methods are often hindered by prohibitive computational costs. Recent advances in Machine Learning (ML) have opened new opportunities to accelerate these approaches. In this work, we propose a perturbative selective Configuration Interaction (SCI) method driven by a binary ML classifier. Our approach employs a simple feedforward neural network (FNN), deliberately kept lightweight to ensure fast and efficient training. Within this framework, we demonstrate that the performance of traditional SCI methods, such as CIPSI (Configuration Interaction using a Perturbative Selection done Iteratively), can be significantly accelerated without loss of accuracy. While the feasibility of using ML to identify the most important Slater determinants has previously been well established, our method is, to our knowledge, the first to achieve quantum chemical accuracy (error < 1.0 kcal/mol) across diverse chemical systems for both ground and excited states. The results show improved convergence compared to CIPSI, particularly for excited states involving strained geometries and conformational changes. Furthermore, the ability to classify Slater determinants across multiple geometries of potential energy curves and surfaces paves the way for the development of new regression-based approaches to molecular electronic structure.