The AMC Seminar is back! After the summer break, our Winter Term Seminar Series on Atomistic Modeling-related topics will resume, taking place every Tuesday.

We’re happy to welcome Dr. Paolo Pegolo as our third speaker. At the École Polytechnique Fédérale de Lausanne, he develops frameworks that use machine learning to either directly predict quantum mechanical properties like dipole moments or to infer intermediate quantities that enable the computation of multiple properties through physics-based analysis.

In his talk “Tensors à la carte with the metatensor ecosystem”, Pegolo will explore recent advances in machine learning for atomistic simulations, focusing on predicting tensorial properties like dipoles and polarizabilities using the metatensor and metatomic libraries. He will present two approaches: one leveraging symmetry-based tensor decompositions, and another using data-driven models to learn rotational equivariance.

Date: Tuesday, October 28, 2025, 3:30 pm

Location: MIBE Lecture Hall

Abstract:

Machine learning has become a cornerstone of modern atomistic simulations, providing access to potential energy surfaces at ab initio accuracy through models capable of predicting energies and their derivatives, forces and stresses. Extending these approaches beyond scalar quantities to general tensorial observables such as dipoles, polarizabilities, and response tensors requires a careful treatment of rotational symmetry. In this talk, I will discuss recent developments in learning tensorial properties within the metatensor and metatomic ecosystem, a set of open-source libraries designed to bridge traditional atomistic simulation codes with modern machine-learning frameworks. I will focus on two complementary paradigms: one in which tensorial targets are expressed as the product of a scalar function of the atomic geometry and a small basis of tensors with the appropriate symmetry, and another that employs unconstrained architectures which learn approximate O(3) equivariance directly from data.