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.
Abstract
Machine-learned interatomic potentials (MLIPs) have emerged as a transformative tool for computational materials science and chemistry, with universal potentials trained on large and diverse datasets now routinely deployed as'foundation models'for downstream fine-tuning in targeted chemical spaces. Many scientific applications of the resulting models, such as molecular dynamics (MD), require high inference and training speeds as well as accuracy. In this work, we examine the limits of equivariant MLIPs, which directly encode physical symmetries in model architectures, to achieve these competing targets -- particularly in the regime of extremely large datasets where data efficiency is less critical. We show how this trade-off can be addressed, and present a family of foundation potentials in the NequIP and Allegro equivariant MLIP architectures which achieve leading inference speeds and strong scalability as well as excellent accuracies across a range of community benchmarks -- spanning materials discovery, thermal conductivity prediction, and near-equilibrium mechanical and thermodynamic properties. Accelerations implemented within the NequIP infrastructure now permit training of high-accuracy foundation potentials on ultra-large datasets with dramatically reduced computational cost. Alongside, we show that efforts to improve model accuracy for materials discovery should focus on dataset diversity and improved, consistent descriptions of transition metal compound energy surfaces.
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.
DPA4C is introduced, an equivariant potential whose architecture and compressed CUDA operators are co-designed under deployment constraints to pursue accuracy and efficiency together and brings quantum-trained universal accuracy into a regime of speed and system size previously associated with empirical potentials.
Tian Li, Jianming Xue, Linfeng Zhang et al.· 0 citations
AI2Pot is presented, a scalable and unified MLIP framework that seamlessly integrates model training, evaluation, and large-scale MD simulations with PyTorch-compatible ecosystem, and offers an user-friendly end-to-end framework for the developing, training, and deploying MLIPs for large scale MD.
Hanyu Liu, Linggang Zhu, Xuanguang Zhang et al.· 0 citations
The Active Learning Framework (ALF), an open-source Python package designed to streamline the design and deployment of MLIP training datasets on High Performance Computing resources, is introduced, illustrating ALF’s effectiveness in compiling datasets that capture essential chemical and structural regimes.
V. Grizzi, P. Lohr, Nikita Fedik et al.· Journal of Chemical Theory a...· 0 citations
Dyna-Mat-v1.0 shows that end-to-end finite-temperature validation is essential for quantifying the predictive behaviour of foundation MLIPs, and provides a simple, scalable route for assessing them beyond static and harmonic benchmarks relevant to materials design.
Mikołaj J Gawkowski, Nongnuch Artrith, Silvia Bonfanti et al.· 1 citation
Universal machine learning interatomic potentials (uMLIPs) bridge quantum-mechanical accuracy and large-scale molecular dynamics, but the cost of high-accuracy calculations such as r$^2$SCAN limits training to datasets that remain small relative to the open materials space. Strong average benchmark performance also does not guarantee reliable energy--force predictions for every structure. We propose Adaptive Multi-Teacher Routing (ATR), which reformulates high-fidelity data construction as a structure-wise decision problem under uncertainty. Using a small set of real r$^2$SCAN labels, ATR calibrates multiple pretrained uMLIP teachers and combines structural descriptors, teacher identity, and inter-teacher disagreement to estimate the reliability of each structure--teacher pair. It selects high-confidence predictions for pseudo-label generation and rejects structures for which no teacher is sufficiently reliable. With real r$^2$SCAN labels for only 0.2\% of candidate structures, ATR distils 2.89 million traceable r$^2$SCAN-level pseudo-labels for pretraining. On held-out r$^2$SCAN structures and the MP-r$^2$SCAN benchmark, a lightweight CHGNet trained on the ATR-generated dataset consistently outperforms the baseline and non-routed controls. Finite-temperature molecular dynamics further shows that ATR improves dynamical robustness across multiple material systems, maintaining stable trajectories where baseline simulations undergo catastrophic structural collapse. These results establish active rejection as an effective mechanism for converting multiple pretrained uMLIPs into a scalable and reliable data-construction system for high-fidelity uMLIPs.
Mingxiang Luo, Xinnan Mao, Lu Wang et al.· 0 citations
