Machine-learned interatomic potentials (MLIPs) have become the state-of-the-art for performing accurate, scalable molecular dynamics (MD) simulations. It is, therefore, crucial to understand and quantify the reliability of MLIPs for downstream property predictions. Uncertainty in predicted properties can arise from limitations in first-principles training data, intrinsic MLIP model errors in representing the data, and the statistical noise introduced during subsequent MD simulations. Using ion transport in Li7P3S11 as a case study, we systematically assess the impact of training set size and selection, neural network stochasticity, and MD sampling statistics on predicted diffusivity and activation energy. We find that when using equivariant MLIP architectures with standard MD protocols, uncertainty arising from MD sampling dominates over model-induced errors. In contrast, MLIP errors relative to the underlying first-principles data are consistently minor. Given this, there are two main routes to improving the accuracy of predictions based on MLIP potentials: adopting higher accuracy reference data generation methods and improving the MD sampling statistics.
Machine Learning interatomic potentials (MLIPs) have emerged as powerful tools for molecular dynamics (MD) simulations with their competitive accuracy and computational efficiency. However, MLIPs often exhibit unphysical behavior when encountering configurations that deviate significantly from their training data distribution, leading to simulation instabilities and unreliable dynamics. This limits their reliability for materials simulations. We therefore present a physics-informed pretraining strategy that leverages simple empirical potentials to improve the robustness and stability of MLIPs for MD simulations. We demonstrate this approach through a pretraining-finetuning pipeline where MLIPs are initially pretrained on data labeled with embedded atom model (EAM) potentials and subsequently finetuned on the quantum mechanical ground truth data. Evaluation across three material systems (phosphorus, silica, and a subset of Materials Project) and three representative MLIP architectures (CGCNN, M3GNet, and TorchMD-NET) demonstrates that this physics-informed pretraining consistently improves both prediction accuracy as well as stability in MD compared to the baseline models.
Unknown authors· Journal of Chemical Informat...· 0 citations
Machine-learning interatomic potentials (MLIPs) have become a powerful tool for rare event sampling in molecular dynamics, offering near ab initio accuracy at a fraction of the computational cost. However, the uncertainty associated with these models remains a major challenge. Existing uncertainty quantification approaches have largely focused on point-wise quantities, such as energies and forces, or on equilibrium thermodynamic observables. In this work, we introduce a framework for propagating MLIP uncertainty to the averaged committor probability, a kinetic observable that enables reaction-rate calculations. Our approach combines rare event sampling methods such as Adaptive Multilevel Splitting with Girsanov reweighting to estimate the sensitivity of committor probabilities to variations in MLIP parameters, without requiring the costly resampling of reactive trajectories for each parameter realization. We derive exact and approximate Girsanov-based estimators for uncertainty propagation and validate them on several benchmark systems, including a rugged Muller-Brown potential, a dimer in a solvent, and the conformational transition of butane. The proposed framework enables the construction of uncertainty-aware probability distributions for rare event observables and successfully recovers reference rare event probabilities from uncertain surrogate models. Under mild assumptions on the accuracy of the MLIP within metastable basins, the framework can also provide uncertainty bounds on reaction rates through Hill's relation. These results demonstrate that path-space reweighting provides an efficient route for propagating MLIP uncertainty to rare event kinetics.
Léonard Moracchini, Thomas Pigeon, Morgane Menz et al.· 0 citations
Machine learning interatomic potentials bridge the gap between quantum chemical precision and classical computational speed, enabling molecular dynamics simulations with first-principles accuracy. Their reliability is often improved through active learning, which iteratively expands the training set by identifying uncertain, out-of-distribution configurations. Existing uncertainty-quantification methods often involve a trade-off between computational cost and reliability, and generally cannot account for redundancy as an acquisition batch is assembled. Here, we introduce AdaptNTK, a single-model framework that measures uncertainty as a regularized Mahalanobis distance in empirical neural tangent kernel (NTK) feature space. With the NTK features fixed during acquisition, the uncertainty depends on the acquired configurations but not their reference labels. This allows the uncertainty to be updated recursively after each selection without retraining, reducing redundancy within an acquisition batch. On held-out rMD17 data, AdaptNTK achieves the highest mean correlations with force errors (Spearman 0.68, Pearson 0.71) and matches a three-member ensemble in error retention. In active learning experiments, AdaptNTK achieves the lowest force errors across rMD17 and Transition-1X, with particularly strong performance on transition-state configurations in Transition-1X. AdaptNTK provides a 2.6-fold speedup per Transition-1X cycle relative to the ensemble, providing efficient single-model uncertainty estimation with sequential updates for data-efficient active learning.
The limits of equivariant MLIPs are examined, and a family of foundation potentials in the NequIP and Allegro equivariant MLIP architectures are presented which achieve leading inference speeds and strong scalability as well as excellent accuracies across a range of community benchmarks.
Seán R. Kavanagh, Chuin Wei Tan, Menghang Wang et al.· 0 citations
Machine-learned interatomic potentials (MLIPs) have become an increasingly important tool for molecular dynamics (MD) simulations, enabling near quantum-mechanical accuracy at significantly reduced computational cost. Recent studies indicate that the Graph Atomic Cluster Expansion (GRACE) neural network architecture delivers strong performance in materials chemistry. In this work, we assess the GRACE architecture for the prediction of potential energy surfaces for organic molecules and introduce GRACE-OFF (GRACE Organic Force Field). GRACE models of varying depth (one-layer and two-layer) and size (small, medium, large) are trained on the SPICE v2.0 data set. We validate the resulting models using a variety of benchmarks. These include single-point energy and force predictions, torsional energy profiles, condensed phase properties of organic liquids and water (thermodynamic properties, self-diffusion coefficients, radial distribution functions, and temperature-dependent water density), as well as the stability of biomolecular MD simulations for gas-phase Ala15 and solvated crambin. For the single-molecule benchmarks (single point energies and forces, torsional energy profiles), the one-layer models showed only mediocre performance, whereas the two-layer models outperformed the MACE-OFF models to which we compare. For the condensed phase properties, the two-layer models gave consistently better results than the MACE-OFF family of MLIPs. For water and hexane, the GRACE-OFF models also beat the much more expensive small UMA/OMol25 (S) model. The two-layer GRACE-OFF models accurately reproduce experimental water radial distribution functions and predict water densities in close agreement with experimental data over a temperature range from 270 to 330 K. Benchmarks demonstrate that GRACE-OFF achieves higher MD performance than comparable MACE-OFF models in both single and double precision. This establishes GRACE-OFF as an accurate and computationally efficient foundation potential for routine simulations of organic liquids and biomolecular systems.
Anna Katharina Picha, Johannes Karwounopoulos, Linus C. Erhard et al.· Journal of Chemical Theory a...· 1 citation
Pretrained machine-learning interatomic potentials, so-called universal or foundation models offer an appealing starting point for atomistic simulations, but their accuracy for material-specific observables often remains limited without additional reference data (fine-tuning). Here, we systematically quantify how much first-principles data are required to convert universal models into ab initio-accurate material-specific potentials, and ask whether fine-tuning is necessarily preferable to training from scratch. We compare five universal MLIP frameworks, MACE-MP-0, SevenNet-0, GRACE-1L-OAM, MatterSim-v1-5M and ORB-v2, across seven chemically diverse systems incorporating rare and reactive events. Fine-tuning on only 10 AIMD-derived configurations is insufficient for the investigated systems; 200 configurations succeed in favorable cases, but the outcome remains strongly system-dependent. By contrast, 2000 AIMD configurations constitute a robust default, yielding low force and energy errors and reproducing the target material-specific observables. Moderately dense sub-sampling of the AIMD trajectory reduces the required trajectory length tenfold with little loss in model quality. Training from scratch on the same datasets is competitive with, and often slightly more accurate than, naive fine-tuning for MACE and SevenNet, whereas GRACE requires more data. The energy profile for a sulfur-vacancy jump in MoS$_2$ reveals that low trajectory-level errors do not guarantee a correct reaction profile, highlighting the need for observable-level validation. Finally, we show that averaging independently trained models improves predictions in scarce-data regimes at no additional first-principles cost. Together, these results provide practical guidelines for converting limited AIMD reference data into reliable material-specific MLIPs for nanosecond-timescale simulations at near-DFT accuracy.
Jonas Hänseroth, Christian Dreßler· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
