Enhanced Generalization through Atomic Embeddings and Cross-Element Parameterization

Abstract

The limited transferability of machine-learned interatomic potentials (MLIPs) to unseen chemical compositions constrains their utility in computational materials discovery. We present a systematic investigation into two interdependent design principleslearned atomic representations and cross-element parameterizationthat govern MLIP generalization. Using the diverse, multielement OC20 catalyst dataset, we evaluate chemically informed atomic embeddings against conventional 1-hot encoding within two distinct model architectures: the tensor-based LATTE descriptor and the MACE message-passing neural network. We demonstrate that embeddings such as Magpie and Matscholar reduce force mean absolute errors (MAEs) by 20–30% compared to 1-hot encoding, and improve data efficiency in low-data regimes by over 40%. Furthermore, by strategically sharing parameters across elements, we reduce the model sizes by up to 98% (from 23M to 400k parameters) without remarkably compromising accuracy, and even enable meaningful few-shot prediction for underrepresented elements. This work establishes a practical framework for building compact, transferable, and data-efficient MLIPs suitable for exploration across heterogeneous chemical spaces.

Transfer learning on universal interatomic potential embeddings improves generalization in structure-property defect models
Matthew D. Witman, Sebastian Pujet, Andrew J. E. Rowberg +4
ChemRxiv29 Oct 2026PreprintLicence not statedGeneral chemistryChemRxiv
Hybridizing expert descriptors and learned embeddings yields transferable Mat2Vec-* element representations
Xiaotong Liu, Qiong Feng, Haotian Chen +2
Communications Chemistry30 Oct 2026Peer-reviewedCC BY-NC-ND 4.0Full textGeneral chemistryNatureCrossrefEurope PMCPMC
Machine learning-driven inverse design of high-hardness multi-principal element alloys through symbolic regression
Peiyi Li
Scientific Reports23 Oct 2026Peer-reviewedCC BY-NC-ND 4.0Full textMaterials & polymersCrossref
Conformational inhibitors of the HIV-1 Rev response element identified through a FRET-based screening approach
Mateusz P. Szewczyk, Saurabh Loharch, Jon Izquierdo-Pujol +10
Communications Chemistry6 Nov 2026Peer-reviewedCC BY 4.0General chemistryNatureCrossref
Thermodynamically Constrained Deep Learning for Hydrogen Sorption Prediction: Overcoming Classical Isotherm Generalization Failure through Physics-Informed Cross-Lithology Transferability
Mohammad Nooraiepour, Mohammad Masoudi, Zezhang Song +1
ACS Omega3 Nov 2026Peer-reviewedCC BY 4.0Full textMaterials & polymersCrossref

Cite

BibTeX RIS

BibTeX

@article{taleblou2026enhanced,
  title = {{Enhanced Generalization through Atomic Embeddings and Cross-Element Parameterization}},
  author = {Mina Taleblou and Franco Pellegrini and Stefano de Gironcoli},
  journal = {Journal of Chemical Information and Modeling},
  year = {2026},
  month = oct,
  doi = {10.1021/acs.jcim.6c01725},
  url = {https://doi.org/10.1021/acs.jcim.6c01725},
  note = {CC BY 4.0},
}

RIS

TY  - JOUR
TI  - Enhanced Generalization through Atomic Embeddings and Cross-Element Parameterization
AU  - Mina Taleblou
AU  - Franco Pellegrini
AU  - Stefano de Gironcoli
PY  - 2026
DA  - 2026/10/02
JO  - Journal of Chemical Information and Modeling
DO  - 10.1021/acs.jcim.6c01725
UR  - https://doi.org/10.1021/acs.jcim.6c01725
AB  - The limited transferability of machine-learned interatomic potentials (MLIPs) to unseen chemical compositions constrains their utility in computational materials discovery. We present a systematic investigation into two interdependent design principleslearned atomic representations and cross-element parameterizationthat govern MLIP generalization. Using the diverse, multielement OC20 catalyst dataset, we evaluate chemically informed atomic embeddings against conventional 1-hot encoding within two distinct model architectures: the tensor-based LATTE descriptor and the MACE message-passing neural network. We demonstrate that embeddings such as Magpie and Matscholar reduce force mean absolute errors (MAEs) by 20–30% compared to 1-hot encoding, and improve data efficiency in low-data regimes by over 40%. Furthermore, by strategically sharing parameters across elements, we reduce the model sizes by up to 98% (from 23M to 400k parameters) without remarkably compromising accuracy, and even enable meaningful few-shot prediction for underrepresented elements. This work establishes a practical framework for building compact, transferable, and data-efficient MLIPs suitable for exploration across heterogeneous chemical spaces.
SN  - 1549-9596
SN  - 1549-960X
ER  -

Abstract

Related

Cite