Dr. Dan Mendels
26/02/2026
אודיטוריום ע"ש דויד וואנג, בניין מידן, קומה 3
13:30
A key challenge in computational molecular and materials modelling is simulating systems and processes with time scales beyond the reach of standard methods. A common solution involves introducing artificial bias potentials to flatten free energy landscapes, significantly accelerating sampling. These bias potentials are guided by collective variables (CVs), which must be carefully designed for each problem. However, designing CVs is a complex task that has traditionally relied on intuition and trial-and-error, spurring recent advancements in machine learning-based approaches to automate and improve the process.
In this talk, we present a machine learning-based approach for systematically constructing CVs with minimal prior system knowledge. We show how this method improves sampling efficiency in various cases and highlight its interpretability.
Furthermore, we explore how ML-based CVs can be used to shape free energy surfaces, allowing us to manipulate thermodynamic and kinetic behaviors for insights into structure-dynamics-function relationships and practical applications. Finally, we explore the use of a graph neural network-based simulators for enhanced sampling and efficient system-to-property learning, highlighting its potential to engineer the dynamics and functionality of complex molecular systems.