Research map / Probabilistic and scientific modeling / Molecules and proteins

Stein Kernelized Molecular Dynamics for Active Learning of Interatomic Potentials

NeurIPS 2026 · Probabilistic and scientific modeling › Molecules and proteins · approach: Machine-learned interatomic potentials · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Stein kernelized molecular dynamics (SKMD), an asynchronous stochastic SVGD-based sampler with a global-descriptor kernel and an adaptive stopping criterion for online active learning of MLIPs.

Problem

Standard MD and existing enhanced-sampling active learning methods get trapped in energy basins, produce correlated training data, and do not preserve the Boltzmann distribution of the MLIP, limiting MLIP improvement per iteration.

Approach

Adapts stochastic Stein variational gradient descent to MD with asynchronous single-particle updates and a kernel over global atomic descriptors; the drift preserves the Boltzmann distribution asymptotically (mean-field argument), and the norm of the Stein biasing force defines an adaptive stopping/acquisition criterion for selecting non-redundant training configurations online.

Key results

• On Muller-Brown active learning over 50 trials, a-SKMD converges to lower energy/force RMSE in fewer iterations than UDD and Langevin-based sampling (qualitative; no numbers stated in extracted text).
• On alanine dipeptide, SKMD-generated data yields the fastest and largest reductions in energy and force RMSE over 10 AL iterations (100 samples each) versus unbiased 300 K and 700 K underdamped MD.
• In sampling tests on Muller-Brown, SKMD gives lower mean and variance of Wasserstein-2 error than overdamped Langevin dynamics for stopping times up to 10^3.

Surprising findings

• Increasing the SKMD stopping time increases discrepancy from the invariant distribution, though it still beats Langevin at stopping time 10^3.
• A variable (median) kernel bandwidth makes the acquisition function norm nearly constant, so online acquisition fails and offline acquisition performs better.

Limitations

• Adaptive stopping criterion requires a fixed kernel bandwidth; variable bandwidth weakens the acquisition signal.
• Asymptotic Boltzmann guarantee relies on limits (small step, small stopping time, infinite particles) that are only approximated in practice.
• Evaluated only on a 2D toy potential and alanine dipeptide with a single fine-tuning setup.

Experimental setup

Datasets: SPICE MACE-OFF-23 training data (0.1% subset) Muller-Brown potential samples

Benchmarks: Muller-Brown potential active learning alanine dipeptide MACE fine-tuning

Metrics: energy RMSE force RMSE Wasserstein-2 distance

Baselines: overdamped Langevin dynamics uncertainty-driven dynamics underdamped MD 300 K underdamped MD 700 K

Builds on: Stein variational gradient descent uncertainty-driven dynamics stochastic SVGD Stein points MACE

Closest papers by meaning

• Physics-Informed Weakly Supervised Learning For Interatomic Potentials (ICML 2025 · similarity 0.88)
• FlashTP: Fused, Sparsity-Aware Tensor Product for Machine Learning Interatomic Potentials (ICML 2025 · similarity 0.88)
• FlashMD: long-stride, universal prediction of molecular dynamics (NeurIPS 2025 · similarity 0.88)
• Higher-Rank Irreducible Cartesian Tensors for Equivariant Message Passing (NeurIPS 2024 · similarity 0.87)
• The Importance of Being Scalable: Improving the Speed and Accuracy of Neural Network Interatomic Potentials Across Chemical Domains (NeurIPS 2024 · similarity 0.87)
• Conditional Normalizing Flows for Active Learning of Coarse-Grained Molecular Representations (ICML 2024 · similarity 0.87)