🔬 materials science

Cross-Geometry Transferability Assessment of Universal Machine Learning Interatomic Potentials: From Bulk Materials to Atomic Nanowires

This study evaluates the cross-geometry transferability of pretrained machine learning interatomic potentials for ZrO2, revealing that while fine-tuning outperforms training from scratch, zero-shot models suffer significant degradation in low-coordination nanostructures and that average error metrics fail to reliably predict specific physical property performance, necessitating geometry-diverse data and independent physical validations for robust adaptation.

Original authors: Pedro H. M. Zanineli, Bruno Focassio, Gabriel R. Schleder

Published 2026-08-10

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Pedro H. M. Zanineli, Bruno Focassio, Gabriel R. Schleder

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to build a video game world where atoms are the characters. To make the game run fast, you can't calculate every single interaction from scratch every time a character moves; that would take too long. Instead, you teach a smart computer program a "rulebook" based on a few thousand examples of how atoms behave. This rulebook is called a Machine Learning Interatomic Potential (MLIP). Think of it like a chef who has memorized thousands of recipes for perfect cakes (bulk materials). Now, imagine you ask that chef to bake something completely different, like a delicate, thin strand of spun sugar (an atomic wire) or a crumbly cookie (a nanoparticle). The chef might try to use the same "cake rules" for the sugar, but the result could be a disaster because the physics of a thin strand are totally different from a big block of cake. Scientists care about this because if we want to simulate new materials for batteries or electronics, we need our computer "chefs" to be reliable even when the shape of the material changes drastically, not just when it stays the same old block.

This paper is like a stress test for a bunch of these computer chefs. The researchers wanted to see if the best "universal" AI models, which are trained on huge datasets of standard, blocky materials, could handle a specific, tricky scenario: Zirconia (ZrO2). They focused on a weird process where Zirconia grains pull apart, getting thinner and thinner until they form incredibly thin, wire-like strands. It's a journey from a solid block to a fragile thread, passing through slabs, particles, and necks along the way.

The team first tested 26 different pre-trained AI models without teaching them anything new (a "zero-shot" test). The results were a bit of a shock: while the models were okay at describing the big, blocky parts, they completely fell apart when it came to the thin wires and necks. The errors in predicting how hard the atoms were pushing or pulling were huge—up to 197.3 meV per Ångström for the forces in the wire sections. It was like the chef trying to bake a soufflé using a recipe for a brick.

To fix this, the researchers tried three strategies: using the model as-is, "fine-tuning" it (giving it a quick crash course on Zirconia), or "training from scratch" (teaching it everything from the beginning). They found that fine-tuning produced more accurate predictions for both energy and force than training from scratch, while requiring a comparable amount of time to run. However, there was a catch. If they only taught the model about one specific shape (like just the wires), it got really good at wires but forgot how to handle the blocks or slabs. This is called "negative transfer"—getting better at one thing made it worse at everything else.

The real winner was "mixed-geometry" fine-tuning. When they fed the model a little bit of everything—blocks, slabs, particles, necks, and wires—it learned to handle the whole process smoothly. The errors dropped significantly, with the best model reaching an energy error of just 6 meV per atom and a force error of 197.3 meV Å⁻¹ for the tricky wire parts.

But here is the most important twist: being good at predicting energy and force numbers didn't guarantee the model was good at everything. When the researchers tested the models on real physical properties, the rankings changed. For example, the models that were best at predicting the stiffness of the material (elasticity) were actually the ones that hadn't been trained much at all. But when it came to simulating the actual breaking of the wire (molecular dynamics), the fine-tuned models were the only ones that didn't immediately fall apart. The paper suggests that to trust these AI models for complex, shape-shifting materials, you can't just look at a single average score. You need to feed them a diverse diet of shapes and test them on the specific physical behaviors you care about, because a model that is "good on average" might still fail spectacularly when the atoms start to dance.

Technical Summary: Cross-Geometry Transferability Assessment of Universal Machine Learning Interatomic Potentials

Problem Statement
Machine-learning interatomic potentials (MLIPs) offer a pathway to extend atomistic simulations beyond the scales accessible to first-principles methods. While foundation models pretrained on large datasets (e.g., Materials Project, Alexandria) have shown promise for broad elemental transferability, their reliability across diverse structural geometries remains insufficiently understood. Existing pretraining datasets are dominated by periodic bulk crystalline structures, leading to a potential performance gap when these models are applied to low-coordination, non-equilibrium, or low-dimensional environments such as surfaces, nanoparticles, necks, and atomic wires. This limitation is particularly critical for processes involving continuous structural evolution, such as the desintering of zirconia ( $ZrO_2$ ), where coordination regimes shift dynamically. Furthermore, standard benchmarks relying on average energy and force errors on bulk datasets do not necessarily guarantee accurate derived physical properties or stable molecular dynamics in these out-of-distribution regimes.

Methodology
The authors constructed a density-functional-theory (DFT) dataset of $ZrO_2$ configurations motivated by an experimentally observed desintering process involving neck thinning and atomic wire formation. The dataset encompasses five distinct geometry classes: bulk, slab, particle, neck, and atomically thin wire.

Dataset Generation: Structures were generated via ab initio molecular dynamics (AIMD) and quasistatic elongation, capturing non-equilibrium configurations, bond rearrangements, and coordination reduction. A force filter ( $<3$ eV Å $^{-1}$ ) was applied after geometry-stratified splitting to ensure physical validity without biasing the test set.
Models Evaluated: The study benchmarked 26 pretrained MLIPs, categorized as Materials Project-compliant (MP-C) or non-compliant (MP-NC). Two representative models, MACE-MP-0 (MP-C) and ORB-V3 (MP-NC), were selected for deeper analysis.
Adaptation Strategies: The authors compared three strategies:
1. Zero-shot inference: Direct application of pretrained models with only reference-energy alignment.
2. Fine-tuning: Full-parameter fine-tuning on target data.
3. Training from scratch: Training models from random initialization on the same target data.
Evaluation Protocols:
- Cross-Geometry Transfer: Models were fine-tuned on specific geometry classes (e.g., bulk only) and tested on others to assess negative transfer. Mixed-geometry fine-tuning was also evaluated.
- Property-Level Validation: Beyond static energy/force errors, the study assessed elastic moduli, vibrational properties (phonon dispersions), surface energies, and neck molecular dynamics (RMSD and rupture behavior).

Key Results

Zero-Shot Performance: Pretrained models exhibit pronounced geometry-dependent degradation. While bulk configurations are described relatively accurately, errors increase significantly for neck and wire configurations. The best zero-shot model (ORB-V3) achieved an energy RMSE of 6 meV atom $^{-1}$ and a force RMSE of 197.3 meV Å $^{-1}$ , though force errors remained highest in low-coordination environments.
Adaptation Efficiency: Under a controlled 50-epoch optimization protocol, fine-tuning yielded lower energy and force errors than training from scratch, with comparable wall-clock times. This suggests pretrained initialization provides a superior starting point for adaptation.
Geometry-Specific vs. Mixed-Geometry: Fine-tuning on a single geometry class improved in-domain accuracy but frequently resulted in "negative transfer," degrading performance on other structural classes (e.g., wire-only fine-tuning increased errors on bulk). Conversely, mixed-geometry fine-tuning reduced cross-geometry errors and improved generalization across the structural manifold.
Property-Level Discrepancies: Rankings based on average energy and force errors did not universally predict property-level behavior.
- Elastic/Vibrational: Zero-shot models often retained strong performance for bulk elastic and vibrational properties, whereas fine-tuning sometimes degraded these quantities due to the inclusion of non-bulk data in the training set.
- Surface Energies: Adaptation strategies showed model-dependent optimal choices; fine-tuning benefited MACE, while training from scratch was superior for ORB.
- Dynamics: Molecular dynamics simulations of neck rupture revealed failure modes (e.g., trajectory divergence or rupture) not apparent in static metrics.

Significance and Claims
The paper claims that geometry-diverse target data and independent physical validations are necessary when adapting foundation MLIPs to low-coordination ionic nanostructures. The authors demonstrate that:

Data Diversity is Critical: For processes traversing multiple coordination regimes, mixed-geometry fine-tuning is superior to geometry-specific adaptation, which risks catastrophic forgetting of broader structural representations.
Static Metrics are Insufficient: Average energy and force errors do not guarantee accurate derived properties or stable dynamics. Property-level validation (elastic, vibrational, surface, and dynamic) is essential for assessing model reliability.
Benchmarking Utility: The constructed $ZrO_2$ dataset serves as a benchmark for out-of-distribution structures, complementing existing bulk-dominated datasets by incorporating low-coordination and non-equilibrium geometries.

The authors modestly conclude that while geometry-diverse data can recover much of the practical accuracy needed for stabilizing low-dimensional $ZrO_2$ trajectories, residual errors may still stem from the lack of explicit long-range electrostatic treatments or incomplete optimization. They propose that future work should integrate automated active-learning workflows to systematically explore the potential-energy surface.

More like this