This paper introduces a distance measure that assesses the output of material generative models by capturing both quality and novelty in a single distribution-based evaluation framework, and introduces the Coarse-Fine Transport Distance (CFTD), which is used as guidance for a material generative model.
Abstract
Generative machine learning is increasingly used for inorganic crystal structure generation. Most models and the corresponding evaluation approaches rely on simple forms of crystal structure representation. In this paper, we showcase the power of atom-averaged features from pretrained Machine-Learning Interatomic Potentials (MLIPs), such as MACE, for such tasks. We first introduce a distance measure that assesses the output of material generative models by capturing both quality and novelty in a single distribution-based evaluation framework. In particular, we introduce the Coarse-Fine Transport Distance (CFTD) using two different featurizers, where the quality component is based on coarse MACE features. We showcase CFTD's versatility in capturing crystal-structure quality while also detecting memorization, and compare it with the recently introduced continuous SUN metrics. We further show that coarse MACE features can be used as guidance for a material generative model.
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.
Pre-trained or'foundational'machine-learned interatomic potentials (MLIPs) are now widely used in materials modelling. However, early pre-trained models and benchmarks have largely focused on ordered, crystalline structures, and their transferability to non-crystalline solids remains unclear. Here, we show that the amorphous state is indeed a central challenge for future universal MLIPs, based on a systematic evaluation of current mainstream models in this domain. We introduce a benchmarking framework built on a curated reference dataset of canonical amorphous systems, as well as validation for structures and properties. Our study identifies limitations in the transferability of many current pre-trained models and investigates fine-tuning strategies tailored to disordered phases. Together, our results can facilitate future applications of MLIPs in the fast-growing field of amorphous functional materials, and they provide guidance for designing next-generation training datasets and transferable atomistic models.
Natascia L. Fragapane, Volker L. Deringer· 2 citations
Electronic structure calculations remain a major bottleneck in atomistic simulations and, not surprisingly, have attracted significant attention in machine learning (ML). Most existing approaches learn a direct map from molecular geometries, typically represented as graphs or encoded local environments, to molecular properties or use ML as a surrogate for electronic structure theory by targeting quantities, such as Fock or density matrices expressed in an atomic orbital (AO) basis. Inspired by the Hohenberg-Kohn theorem, in this work, we propose an operator-centric framework in which the external (nuclear) potential, expressed in an AO basis, serves as the model input. From this operator, we construct hierarchical, body-ordered representations of atomic configurations that closely mirror the principles underlying several popular atom-centered descriptors. At the same time, the matrix-valued nature of the external potential provides a natural connection to equivariant message-passing neural networks. In particular, we show that successive products of the external potential provide a scalable route to equivariant message passing and enable an efficient description of nonlocal effects. We demonstrate that this approach can be used to model molecular properties, such as energies and dipole moments, from the external potential or to learn effective operator-to-operator maps, including mappings to the Fock matrix from which multiple molecular observables can be simultaneously derived.
Jigyasa Nigam, T. Smidt, G. Dusson· Journal of Chemical Physics· 2 citations
Foundation machine-learning interatomic potentials (MLIPs) enable atomistic simulations at substantially lower computational cost than first-principles methods, but their reliability across structural geometries remains insufficiently understood. Here, we construct a density-functional-theory dataset of ZrO2 configurations spanning bulk, slab, particle, neck, and atomically thin wire environments motivated by an experimentally observed ZrO2 desintering process involving neck thinning and atomic wire formation. We first benchmark 26 pretrained MLIPs and observe pronounced geometry-dependent degradation in zero-shot predictions. Without any training, after only reference-energy alignment, the best zero-shot model (ORB-V3) reaches energy and force root-mean-square errors of 6 meV/atom and 197.3 meV/{\AA}, respectively, with the largest force errors in neck and wire configurations. We then compare zero-shot inference, fine-tuning, and training from scratch strategies. Fine-tuning yields lower energy and force errors than training from scratch, while both require comparable wall-clock time. Geometry-specific fine-tuning improves in-domain accuracy but frequently produces negative transfer to other structural classes, whereas mixed-geometry fine-tuning reduces cross-geometry errors. Evaluations of elastic and vibrational properties, surface energies, and neck dynamics further show that rankings based on average energy and force errors do not universally predict property-level behavior. These results demonstrate that geometry-diverse target data and independent physical validations are necessary when adapting foundation MLIPs to low-coordination (ionic) nanostructures.
P. Zanineli, B. Focassio, G. R. Schleder· 0 citations
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.
Seán R. Kavanagh, Chuin Wei Tan, Menghang Wang et al.· 0 citations
Foundation machine-learned interatomic potentials (MLIPs) are trained on large quantum-mechanical datasets and generalise across broad regions of chemical and configurational space. Beyond their usual role in accelerating sampling-based simulations, their internal representations encode chemically rich local atomic environments. Here, we introduce Rem3Di, a representation-learning framework that repurposes latent features from atomistic foundation models as transferable molecular descriptors for property prediction and virtual screening. Rem3Di combines a potential's per-atom features into a single fixed-length descriptor of the whole molecule that varies smoothly with three-dimensional structure and is invariant to the ordering of the atoms. The descriptor can be used directly or fine-tuned for specific prediction tasks. To capture molecular handedness, Rem3Di constructs pseudoscalar features, which are unchanged by rotation but reverse sign under mirror reflection. This lets the descriptor distinguish enantiomers, which can differ in activity and toxicity. The transformer is pretrained on large molecular datasets by reconstructing corrupted atom features, so no experimental labels are required. Across public drug-property benchmarks, Rem3Di matches or exceeds published baselines without relying on classical 2D fingerprints. Additionally, the same descriptor yields chemically meaningful differentiation of transition-metal complexes without predefined bonding rules or handcrafted representations. Rem3Di therefore provides a route from simulation-trained atomistic representations to transferable, chirality-aware molecular representations for chemical machine learning.
Steffen Wedig, Felix Burton, Rokas Elijošius et al.· 0 citations