Transformer Atomic Cluster Expansion (TRACE) is introduced, an energy-conserving architecture that combines atomic cluster expansion density correlations with local multihead cross-attention that captures multi-species crystallization, liquid structures, phase diagrams, and chemical reactivity.
Abstract
Designing machine-learning interatomic potentials involves achieving the precise representation of complex many-body interactions alongside the efficiency required for scalable molecular dynamics. We introduce Transformer Atomic Cluster Expansion (TRACE), an energy-conserving architecture that combines atomic cluster expansion density correlations with local multihead cross-attention. The correlations form an O(3)-equivariant state for each center, which queries tensorial neighbor features that remain fixed functions of species and geometry. No learned state is passed between atoms. On a laptop MacBook-M1, we train and test TRACE for polymorphic cesium lead iodide, liquid water, and intramolecular methyl migration against experiments. For cesium lead iodide, TRACE reproduces the r$^2$SCAN+rVV10 ordering of four polymorphs and gives a classical edge-sharing hexagonal non-perovskite($\delta$) to corner-sharing cubic perovskite($\alpha$) Gibbs-free-energy crossing $\simeq$580K near the experimental observations of $\simeq$600K. By employing enhanced sampling to cross high energy barriers, the same TRACE potential successfully captures the $\delta$-to-$\alpha$ perovskite transformation without any reinforcement learning. A water potential trained on a reduced set of CCSD(T) configurations places the first oxygen--oxygen maximum at 2.85~\AA, compared to the experimental value of 2.80~\AA{}. For the gas-phase methyl migration in 2,2-dimethylisoindene, umbrella sampling yields an activation free energy of $27.92\pm0.03$~kcal~mol$^{-1}$, in close agreement with the experimental measurement of $29.2\pm1.1$~kcal~mol$^{-1}$. Across these diverse benchmarks, a single unified architecture successfully captures multi-species crystallization, liquid structures, phase diagrams, and chemical reactivity.
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
The prediction of the structural stability of octet $AB$-type binary compounds is a classical materials informatics problem. The challenge is to capture the relative stability of 4-fold coordinated atoms in zincblende ($\beta$-ZnS) structure and 6-fold coordinated atoms in rocksalt (NaCl) structure, modulated by charge transfer and atomic-size differences. Previous structure maps and machine-learning approaches used atomic features such as valence-electron count, ionization potential and atomic radii, using either physical intuition or symbolic regression. Here, we demonstrate that explicitly incorporating the domain knowledge of the interatomic bonds can significantly and systematically improve the prediction of $\beta$-ZnS/NaCl stability. We encode this bonding information through a coarse-grained representation of the local electronic structure obtained by a recursive solution of a tight-binding bond model. The underlying pairwise Hamiltonians are taken from downfolded eigenstates of density-functional theory calculations for diatomic molecules and thereby include domain knowledge of the bond between specific $A-B$ pairs. The benefit of this description is demonstrated with an ensemble of independently trained Kernel Ridge or symbolic regression models combined with sequential feature selection. The obtained models are compared to a previous symbolic-regression model using the same set of \emph{ab initio} calculations for octet binaries as training data. We find a significant improvement in the prediction of the formation energy difference of $AB$ compounds as compared to previous works and demonstrate that an increasing amount of bond-informed recursion features improves the predictive accuracy.
Rohan D. Kumar, Mariano Forti, A. Naik et al.· 0 citations
In this paper, we provide a systematic investigation of SO(2) theory to machine learning interatomic potentials (MLIPs) and identify the limitations of conventional SO(2) Linear architectures relative to SO(3) Clebsch-Gordan Tensor Products (CGTP). Building on these insights, we propose direct Cartesian construction and recursive Clebsch-Gordan construction of Wigner D-matrices and introduce two novel interaction building blocks. First, we propose the Edge Complex Product Basis based on Generalized Asymmetric Contraction, a new formulation for many-body expansion that directly constructs higher-order interactions on edges through complex-valued equivariant multiplications. Second, we introduce Radial Rotary Complex Attention(RRA), which enhances extrapolation performance and surpasses existing attention vector formulations. We also introduce several improvements to the Atomic Cluster Expansion module. Building on these advances, we train our models on OMat24, sAlex, and MPTrj, and introduce TECE-OAM-RRA-1.0, which achieve state-of-the-art (SOTA) performance on the Matbench Discovery.
Graph neural networks have become the dominant machine-learning architecture for predicting materials properties from crystal structures. Yet the initialization of atomic node features has received comparatively little attention, and conventional approaches rely on static elemental descriptors that carry no information about the quantum-mechanical electronic environment of each atom in its crystalline host. Here we show that augmenting atomic node representations with site-projected orbital density of states (pDOS) fingerprints, computed directly from density functional theory calculations, yields systematic and substantial improvements in predictive performance.These representations are fused with Pettifor elemental embeddings at each atomic site before message passing. For the superconducting critical temperature $T_c$ and the optical dielectric constant $\epsilon_{\infty}$,the pDOS augmentation reduces prediction errors by 22.9% and 27.9%, respectively, relative to the elemental-descriptor baseline. These improvements are comparable to those achieved by doubling the training-set size. The gains are, however, contingent on training-set size. For the magnetic exchange energies of Heusler compounds, a substantially smaller dataset, the improvement is reduced,indicating that pDOS augmentation is most effective when the training data exceeds the length of the pDOS feature vector. We introduce an interpretable spectral attention-gating mechanism that reveals that the model autonomously learns to prioritize the orbital channels and energy windows most physically relevant to each target property. These results establish pDOS-augmented graph nodes as a broadly applicable strategy for infusing first-principles electronic-structure knowledge into graph networks, opening a practical route to high-accuracy property prediction in data-scarce regimes.
Paulo R. Pires, Pierre-Paul De Breuck, Mauro Fava et al.· 0 citations
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
This work proposes an operator-centric framework in which the external (nuclear) potential, expressed in an AO basis, serves as the model input and builds hierarchical, body-ordered representations of atomic configurations that closely mirror the principles underlying several popular atom-centered descriptors.
Jigyasa Nigam, T. Smidt, G. Dusson· Journal of Chemical Physics· 2 citations