Accelerate materials discovery using machine learning interatomic potentials

by Carlos Bornes (Department of Physical and Macromolecular Chemistry, Faculty of Science, Charles University, Prague)

Tuesday 28 Oct 2025, 15:00 → 16:00 Portugal

Sala de Conferências (Departamento de Física FCTUC)

Computational modeling and solid-state NMR are combined to gain atomistic insights into the structure, dynamics, and reactivity of catalytic materials. However, experiments typically probe complex, realistic samples containing defects, varying hydration, and diverse compositions, measured at moderate temperatures. In sharp contrast, conventional computational studies often rely on simplistic models, such as static simulations at 0 K in a vacuum or very short molecular dynamics trajectories, using pristine structures. This disconnect between experimental reality and computational models hinders the development of a true atomic-level understanding and the rational design of new catalysts.
Reactive machine learning interatomic potentials (MLIPs) can maintain the accuracy of ab initio methods while being orders of magnitude faster, around 1000x, enabling nanosecond-scale molecular dynamics simulations. This capability allows for the robust exploration of complex potential energy surfaces, capturing the effects of temperature, dynamics, structural defects, and reactivity. Besides modeling energies and forces, MLIPs can be integrated with regression methods to predict experimental observables, such as chemical shifts and tensors, and thus allowing a stronger link between predicted values and experiment results.