Title: Leveraging Stein Kernelized Molecular Dynamics for the Active Learning of Interatomic Potentials

Original: arXiv:2606.04100v1 Announcement Type: New Abstract

Machine learning interatomic potentials (MLIPs) facilitate precise and computationally efficient atomistic simulations; however, their performance is heavily reliant on the richness and variety of the underlying training datasets. To address this, we present Stein kernelized molecular dynamics (SKMD), a novel enhanced sampling technique designed to generate high-value training configurations for the active learning and refinement of MLIPs. SKMD operates as a stochastic adaptation of Stein variational gradient descent, tailored for molecular dynamics through the integration of asynchronous particle updates and a kernel based on global atomic descriptors. This kernel establishes a symmetry-aware metric for assessing configurational similarity.

In contrast to other enhanced sampling methods employed in molecular dynamics, SKMD maintains the Boltzmann distribution as the asymptotic state of the system’s dynamics. This characteristic ensures an optimal equilibrium between exploring a wide range of configurations and directing the system toward high-probability areas within the energy landscape. Additionally, we introduce a method for efficient online data collection that utilizes an adaptive stopping criterion to identify and select non-redundant training samples throughout the simulation process.

We validate the effectiveness of SKMD through two case studies: the active learning of a neural network model applied to the Müller-Brown potential and the fine-tuning of a MACE interatomic potential for alanine dipeptide. Our results indicate that, relative to active learning baselines, SKMD delivers superior model accuracy with fewer training iterations while requiring an equivalent volume of acquired training data.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC