Challenging Atomistic Foundation Models on Disordered Carbons
Abstract
Foundation machine-learned interatomic potentials (MLIPs) promise broad transferability, yet their out-of-domain reliability is hard to establish. Here, we use pure and O/H-containing disordered carbon spanning multiple phases, competing hybridisations and porous frameworks to stress-test six popular foundation MLIPs. We first screen all six on density functional theory (DFT) energy and force errors and binding curves, then subject four to melt-quench-anneal molecular dynamics and evaluate the resulting structural observables. Binding curves and O/H-containing trajectories reveal unphysical extrapolation and dynamical failures that low DFT energy and force errors fail to reveal. Of the models tested, MACE-MH-1 is the most consistently robust, with low DFT errors and stable MD. Despite differing architectures, OMAT-based models share structural biases, highlighting the importance of training coverage. All four models generate structures of nanoporous carbon that reproduce the principal experimental and specialised carbon potential signatures. Incorporating O/H leaves short-range order largely unchanged but shifts pore-size distributions towards larger pores. These results emphasise the need for application-specific validation and extend molecular dynamics modelling of nanoporous carbon towards experimentally realistic structures.
Related
- Nucleus-Independent Chemical Shifts in disordered nanoporous carbons are governed by the long-range magnetic response of the carbon framework
- Thermal conductivity predictions with foundation atomistic models
- Fast and Accurate Equivariant Foundation Models for Atomistic Simulation
- Structural response and competitive pore filling of phenol and caffeine in hydrophobic microporous carbons
- Understanding the Relationship Between Synthetic Conditions and H ₂ Uptake Capacity in Oxidatively Activated Turbostratic Carbons through Machine Learning
BibTeX
@misc{hart2026challenging,
title = {{Challenging Atomistic Foundation Models on Disordered Carbons}},
author = {Joseph Hart and Samuel G. H. Brookes and Fabian L. Thiemann and Christoph Schran},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.26701},
eprint = {2610.26701},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.26701},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Challenging Atomistic Foundation Models on Disordered Carbons
AU - Joseph Hart
AU - Samuel G. H. Brookes
AU - Fabian L. Thiemann
AU - Christoph Schran
PY - 2026
DA - 2026/10/19
PB - arXiv
DO - 10.48550/arxiv.2610.26701
UR - https://arxiv.org/abs/2610.26701
AB - Foundation machine-learned interatomic potentials (MLIPs) promise broad transferability, yet their out-of-domain reliability is hard to establish. Here, we use pure and O/H-containing disordered carbon spanning multiple phases, competing hybridisations and porous frameworks to stress-test six popular foundation MLIPs. We first screen all six on density functional theory (DFT) energy and force errors and binding curves, then subject four to melt-quench-anneal molecular dynamics and evaluate the resulting structural observables. Binding curves and O/H-containing trajectories reveal unphysical extrapolation and dynamical failures that low DFT energy and force errors fail to reveal. Of the models tested, MACE-MH-1 is the most consistently robust, with low DFT errors and stable MD. Despite differing architectures, OMAT-based models share structural biases, highlighting the importance of training coverage. All four models generate structures of nanoporous carbon that reproduce the principal experimental and specialised carbon potential signatures. Incorporating O/H leaves short-range order largely unchanged but shifts pore-size distributions towards larger pores. These results emphasise the need for application-specific validation and extend molecular dynamics modelling of nanoporous carbon towards experimentally realistic structures.
ER -