Amir Natan

 

 

 

 

 

 

Talk: Machine Learning Models in Atomistic Simulations: When Can We Trust Their Predictions? 

 Accurate interatomic forces and total energies are essential for atomistic simulations such as molecular dynamics (MD) and Monte Carlo (MC) for the correct prediction of materials properties and processes. First-principles methods such as density-functional theory (DFT) and other quantum methods provide reliable forces and energies but are often prohibitively expensive. In the last two decades, machine-learning (ML) interatomic models have demonstrated close to first-principles accuracy while remaining several orders of magnitude faster than DFT (but still slower than classical force fields). This raises two important questions: How trustworthy are simulations driven by ML models, and how can we quantify their predictive uncertainty? I examine those questions in three different projects: (1) deep-learning simulations of interlayer sliding in defective graphene and the role of chemical bond formation and rapture in controlling friction; (2) evaluation of the performance of uncertainty-quantification (UQ) signals for direct DL force/energy models; and (3) analysis of how ML errors can affect the inference of materials properties such as the melting temperature at various pressures.  

Workshop Home Page

Return to the TAU-UIUC Workshop Home Page

Home→