Speaker
Description
Molecular dynamics (MD) simulations are a fundamental tool for investigating the atomistic behavior of complex systems, offering deep insights into reaction mechanisms, phase transitions, and emergent properties in both condensed and soft matter. Recent advances in machine learning (ML) have determined a paradigm shift in atomistic simulations, allowing the development of force-fields that closely mimic quantum mechanical interactions with exceptional accuracy and efficiency—achieving this at a fraction of the computational cost of ab initio methods. However, while standard non-linear ML potentials such as Gaussian Approximation Potentials and Neural Network Potentials deliver excellent descriptions of the atomic environment, their numerous fitting parameters often restrict the size of the systems and the duration of simulations due to increased computational demands from high-dimensional parameter spaces. Furthermore, realizing these potentials is labor-intensive, as they generally require extensive training datasets and are vulnerable to overfitting. In this presentation, I will introduce the Chebyshev Interaction Model for Efficient Simulation (ChIMES), which leverages a linear expansion in Chebyshev polynomials to accurately reproduce atomic forces, energies and stress tensors in molecular and condensed-phase systems while requiring comparatively less data. I will present overall findings and discusse future perspectives: these ChIMES force-fields for atomistic simulations are pivotal for the multiscale computational characterization of crystalline oxides like HfO₂. These materials are promising candidates for the coatings of multilayer systems in the Einstein Telescope, where a detailed understanding of their structure and atomic environment is essential for predicting performance under stress and the relationship between mechanical losses and thermoelastic properties.
| AI keywords | chebychev polynomials; linear models; active learning |
|---|
