Machine learning potentials for cationic zirconiumcomplexes grafted on amorphous silica

Abstract

Accurately modeling heterogeneous catalytic systems while maintaining computational efficiency is a persistent challenge, as conventional methods like Density Functional Theory (DFT) offer high accuracy but are computationally expensive, whereas classical force fields provide efficiency without precision. In recent year, Machine Learning Potentials (MLPs) have emerged as a powerful tool to bridge the gap between the efficiency of classical force fields and the precision of first-principles methods. In this study, I assess the accuracy, efficiency, and limitations of MACEMLP-models when applied to a challenging catalytic system: cationic zirconocene hydride grafted onto an amorphous silica slab model. My results demonstrate that MACE models, even with minimal training data, achieve impressive accuracy with energy RMSE below 0.05 eV/atom and force errors under 0.2 eV/Å, highlighting the efficiency of foundational models. Nonetheless, challenges such as a sub-unity slope in energy predictions and dynamically unstable MD simulations due to catastrophic forgetting remain, even with the application of active learning techniques. A novel multihead replay technique shows promise in enhancing stability, though additional validation is necessary. Furthermore, thermodynamic reweighting proves effective in refining bond length distributions, especially with hybrid functionals like PBE0+D3, but its robustness remains sensitive to model accuracy and bias. Overall, these results demonstrate the potential of MLP-based approaches in accelerating calculations by orders of magnitude while emphasizing the importance of thorough validation for accurate predictions.