A semiparametric interatomic potential is introduced based on a generalization of the Abell--Tersoff bond-order potential, incorporating a chemically informed functional form and explicit high-order many-body correlations to address this challenge of reliability for out-of-distribution configurations far beyond the training domain.
Abstract
Machine-learning interatomic potentials enable accurate and efficient atomistic simulations, yet their reliability for out-of-distribution configurations far beyond the training domain remains a significant challenge. Here, we introduce a semiparametric interatomic potential based on a generalization of the Abell--Tersoff bond-order potential, incorporating a chemically informed functional form and explicit high-order many-body correlations to address this challenge. The model is trained and evaluated on various datasets of silicon, carbon, water, and small molecules, achieving interpolation accuracy comparable to existing MLIP models while exhibiting improved extrapolation to unseen configurations, including those at high pressures and temperatures. These results provide insights into the design of specific inductive biases for reliable extrapolation in interatomic potentials. inductive biases that can be used to control extrapolation in machine-learning interatomic potentials.
The convergence of molecular dynamics simulations and machine-learned interatomic potentials (MLIPs) promises density functional theory (DFT) level accuracy at near-classical force-field computational costs. However, while the average fidelity to reference energies and forces approaches perfection, several failure modes limit MLIP reliability in production simulations. These include spurious bond formation, inconsistent reproduction of long-range interactions, and inconsistent spin-state references. Here, the origins of these behaviors are studied by benchmarking the UMA, ORB, MACE, and AIMNet2 models against reference DFT bond dissociation curves for an illustrative range of species. These benchmarks reveal that models without explicit atomic charge resolution predict spurious stable bonds between like-charged halide anions, effectively transmuting two Cl- ions into neutral Cl2. Models with atomic partial charge equilibration correctly predict repulsion in these systems. Conversely, several secondary limitations are exposed in these benchmarks, including inconsistent agreement with unrestricted DFT (uDFT) versus restricted DFT (rDFT) energies and inconsistent core-region treatment. This comparative analysis suggests that, while artifacts related to core repulsion and asymptotic electrostatics are readily repairable through improved physical priors and better data curation, the issue of spurious bond formation is intrinsic to the inability of global charge specification to disambiguate similar local geometries at different charge and spin states.
Ericka Roy Miller, V. Sathyaseelan, Dylan M Gilley et al.· Journal of Chemical Theory a...· 0 citations
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.
High-entropy alloys (HEAs) exhibit exceptional structural and functional properties arising from their complex local chemical environments, and their vast compositional space offers considerable flexibility to further tune and optimize these properties. Atomistic simulations based on density functional theory (DFT) have played a central role in elucidating the thermodynamic, mechanical, magnetic, and defect-related properties of HEAs. However, DFT simulations are severely limited by the intrinsic chemical and configurational complexity of these alloys, particularly because reliable predictions require extensive statistical sampling over chemically diverse configurations and access to extended spatial and temporal scales. In this review, we summarize recent advances in atomistic simulations of HEAs, with particular emphasis on machine learning interatomic potentials (MLIPs), which extend beyond conventional DFT approaches. We discuss how MLIPs enable statistically robust simulations with near-DFT accuracy while dramatically reducing computational cost, thereby allowing explicit treatment of chemical short-range order, vibrational contributions to Gibbs energies, point defects, diffusion, dislocation behavior, grain boundaries, and hydrogen absorption in chemically complex alloys. Particular attention is devoted to the role of local chemical environments, many-body interactions, and configurational sampling in determining HEA properties. We further review recent developments in universal/foundation MLIPs trained on chemically diverse datasets and discuss their potential for rapid and transferable atomistic simulations of HEAs across broad compositional and configurational spaces. We discuss current limitations and open challenges, including transferability to highly distorted defect configurations, treatment of magnetic and charge degrees of freedom, incorporation of finite-temperature excitations, and construction of representative training datasets for chemically and structurally complex systems. This review aims to provide a comprehensive perspective on the ongoing transition from conventional DFT-based simulations toward scalable, statistically rigorous, and predictive atomistic modeling frameworks for HEAs and related compositionally complex materials.
Yuji Ikeda, Xiang Xu, Pranav Kumar et al.· Journal of Materials Science· 0 citations
New methods and workflows to overcome the challenges inherent to automating unrestricted coupled cluster calculations are developed and a transferable MLIP for gas-phase reactions, trained on unrestricted CCSD(T) data is developed.
Alice E. A. Allen, Rui Li, Sakib Matin et al.· Journal of Chemical Theory a...· 0 citations
Machine-learning interatomic potentials (MLIPs) trained by directly learning the total interatomic interaction energies can suffer from limited transferability, unphysical behavior beyond a finite cutoff, and large errors for out-of-distribution geometries such as transition states and uncommon conformers. We evaluate Δ-learning (delta-learning) as a remedy by training an ANI-style high-dimensional neural network (HDNNP) potential as a correction to predict PBE0/aug-cc-pVTZ energies from a third-order tight-binding density functional theory (DFTB3) baseline model. On a held-out test set derived from the modified ANI-1x training set, the Δ-learning model (named ANIDFTB-Δ) achieves a mean absolute error (MAE) of 0.82 kcal/mol, while the HDNNP model (ANIPBE0-Direct) has an MAE of 2.01 kcal/mol. On selected GMTKN55 benchmarks, the Δ model systematically improves relative energies for conformers and tautomers and avoids catastrophic outliers on challenging structures. The good physical description of DFTB3 significantly reduces the error in proton-transfer transition states in the PX13 benchmark and intermolecular interactions in the DES370K dataset in comparison with reference target PBE0. The long-range electrostatics in DFTB3 also partially correct the long-range behavior of local descriptor MLIPs outside of their predetermined cutoff, reducing the MAE from 0.098 to 0.019 kcal/mol. The computational cost of this method is incrementally more expensive than DFTB3 alone, making it practical for extensive simulations of moderately sized systems, although the DFTB3 step makes the Δ model much more costly than a simple HDNNP alone.
Nguyen Thien Phuc Tu, Christopher N. Rowley· Journal of Chemical Physics· 0 citations
The application of density functional theory to heterogeneous catalysis is hindered by the shortcomings of conventional density functional approximations. We combine machine learning with explicitly non-local physically informed descriptors and introduce an exchange-correlation functional (CIDER26SS) framework regularized for wide transferability. CIDER26SS is size-extensive, highly efficient, provides a balanced and accurate description of both molecular and solid-state systems, and is specifically well-optimized for transition metal surface chemistry. Surpassing existing conventional functionals, CIDER26SS resolves the CO/Pt puzzle, identifying the correct binding site for CO adsorption on the Pt(111) surface, along with an accurate adsorption energy, Pt lattice constant, and surface energy. Predictions agree well with the experimental values, even when all bulk and surface data for Pt are excluded from the training set. Remarkably, CIDER26SS exceeds the accuracy of semilocal approximations even for systems far outside the training domain.
M. S. Abdallah, Zhuotao Jin, Boris Kozinsky et al.· 0 citations
