Michele Ceriotti: Physically (un)inspired modeling: how much physics do we need to machine learn the quantum properties of materials?

Physically (un)inspired modeling: how much physics do we need to machine learn the quantum properties of materials?

Machine-learning techniques are often applied to perform "end-to-end" predictions, making black-box estimates of a property of interest using only a coarse description of the corresponding inputs. In contrast, atomic-scale modeling of matter is most useful when it allows one to gather mechanistic insight into the microscopic processes that underlie the behavior of molecules and materials.

In this talk, Dr. Michele Ceriotti will critically discuss how, and to what extent, physical-chemical ideas can and should be integrated into a machine-learning framework aimed at simulating matter at the atomic scale. He will discuss how physical priors, such as smoothness or symmetry of the structure-property relations, are used to inform the mathematical structure of a generic ML approximation—an approach that has become ubiquitous in the field.

Dr. Ceriotti will also discuss the emergence of unconstrained models that can directly learn physical constraints from large amounts of data and that can outperform, in speed and accuracy, models that enforce physical constraints—showing, however, that care should be applied to use these models safely so that the lack of built-in physics does not result in unphysical results. Examples will include PET-MAD, a lightweight unconstrained model that achieves competitive accuracy across the periodic table despite being trained on a relatively small dataset, and that incorporates many advanced capabilities such as uncertainty quantification and direct force estimation.

Our Speaker:

Michele Ceriotti

Michele Ceriotti received his Ph.D. in Physics from ETH Zürich working with Michele Parrinello, and spent three years in Oxford as a Junior Research Fellow at Merton College and as a member of the group of David Manolopoulos. Since 2013, he has led the Laboratory for Computational Science and Modeling at EPFL, focusing on method development for atomistic modeling of matter, bridging quantum mechanics, statistical physics, and machine learning. He has contributed to several open-source software packages, including MetaTensor, i-PI, and Chemiscope, and served the atomistic modeling community as an associate editor of the Journal of Chemical Physics, as a moderator of the physics.chem-ph section of the arXiv, and as an editorial board member of Physical Review Materials. More details on his publications can be found here.