[Paper Review] Efficiency, Accuracy, and Transferability of Machine Learning Potentials: Application to Dislocations and Cracks in Iron

Lei Zhang, Gábor Cśanyi+2 more|arXiv (Cornell University)|Jul 19, 2023

Machine Learning in Materials ScienceMaterials Science3 citations

TL;DR

This study evaluates the efficiency, accuracy, and transferability of state-of-the-art machine learning interatomic potentials (ML-IAPs) for modeling dislocations and cracks in body-centered cubic iron. Using a three-step validation framework—benchmarking against DFT, uncertainty quantification, and large-scale simulations—it demonstrates that GAP and PACE-FS ML-IAPs achieve DFT-level accuracy in predicting dislocation core structures, Peierls barriers, and fracture mechanisms, with PACE-FS offering up to 100× speedup while maintaining transferability when trained on optimized databases.

ABSTRACT

Machine learning interatomic potentials (ML-IAPs) enable quantum-accurate, classical molecular dynamics simulations of large systems, beyond reach of density functional theory (DFT). Yet, their efficiency and ability to predict systems larger than DFT supercells are not fully explored, posing a question regarding transferability to large-scale simulations with defects (e.g. dislocations, cracks). Here, we apply a three-step validation approach to body-centered-cubic iron. First, accuracy and efficiency are assessed by optimizing ML-IAPs based on four state-of-the-art ML packages. The Pareto front of computational speed versus testing root-mean-square-error (RMSE) is computed. Second, benchmark properties relevant to plasticity and fracture are evaluated. Their average relative error Q with respect to DFT is found to correlate with RMSE. Third, transferability of ML-IAPs to dislocations and cracks is investigated by using per-atom model uncertainty quantification. The core structures and Peierls barriers of screw, M111 and three edge dislocations are compared with DFT. Traction-separation curve and critical stress intensity factor (K_Ic) are also predicted. Cleavage on the pre-existing crack plane is found to be the zero-temperature atomistic fracture mechanism of pure body-centered-cubic iron under mode-I loading, independent of ML package and training database. Quantitative predictions of dislocation glide paths and KIc can be sensitive to database, ML package, cutoff radius, and are limited by DFT accuracy. Our results highlight the importance of validating ML-IAPs by using indicators beyond RMSE. Moreover, significant computational speed-ups can be achieved by using the most efficient ML-IAP package, yet the assessment of the accuracy and transferability should be performed with care.

Motivation & Objective

• To assess the efficiency, accuracy, and transferability of ML-IAPs in modeling extended defects such as dislocations and cracks in bcc iron.
• To determine whether ML-IAPs trained on standard databases can reliably predict complex defect structures and fracture mechanisms at scale.
• To evaluate the impact of training database design, ML framework choice, and model uncertainty on predictive performance.
• To benchmark computational speed versus accuracy across multiple ML-IAP packages (GAP, PACE-FS, MTP, qSNAP) using a Pareto front analysis.

Proposed method

• Trained ML-IAPs using two independent DFT databases containing point defects, surfaces, and bulk configurations, with active learning to improve data efficiency.
• Optimized hyperparameters across four ML packages—GAP, MTP, SNAP, qSNAP—using extensive hyperparameter tuning to maximize accuracy and efficiency.
• Computed the Pareto front of computational speed versus root-mean-square error (RMSE) to identify optimal trade-offs between speed and accuracy.
• Performed large-scale molecular statics/dynamics (MS/MD) simulations of screw, M111, and edge dislocations to assess transferability beyond DFT supercells.
• Used per-atom model uncertainty quantification to identify regions of high prediction error and validate ML-IAP performance against consistent DFT calculations.
• Conducted nudged elastic band (NEB) and traction-separation curve analyses to evaluate fracture properties, including critical stress intensity factor $K_{\rm Ic}$.

Experimental results

Research questions

• RQ1 Can ML-IAPs trained on standard databases accurately predict the core structures and Peierls barriers of multiple dislocation types in bcc iron?
• RQ2 How does the choice of ML-IAP framework (e.g., GAP vs. PACE-FS) affect the prediction of dislocation glide paths and fracture mechanisms?
• RQ3 To what extent does the size and composition of the DFT training database influence the transferability of ML-IAPs to large-scale defect simulations?
• RQ4 Can model uncertainty quantification reliably identify regions where ML-IAP predictions deviate from DFT benchmarks?
• RQ5 Is cleavage on the pre-existing crack plane the dominant fracture mechanism in bcc iron at zero temperature, independent of the ML-IAP used?

Key findings

• The PACE-FS ML-IAP achieved up to 100× faster computation than other frameworks while maintaining high accuracy, placing it on the Pareto front for efficiency.
• GAP emerged as the most accurate ML-IAP, achieving DFT-level accuracy in predicting the energetic hierarchy of screw dislocation core structures (easy, hard, split).
• Both GAP and PACE-FS successfully reproduced the Peierls barrier and kink-pair nucleation in screw dislocations, confirming their suitability for large-scale plasticity simulations.
• Cleavage on the pre-cracked plane was consistently predicted as the zero-temperature fracture mechanism across all ML-IAPs and databases, independent of model or training data.
• The same predictive accuracy and transferability were achieved with only 1/5 of the DFT data when using active learning and optimized database design, significantly reducing computational cost.
• Small energy differences between competing core structures (e.g., edge and M111 dislocations) were sensitive to both ML-IAP choice and training database, highlighting the role of DFT accuracy in model discrimination.