Active Learning Strategies for Efficient Machine-Learned Interatomic Potentials Across Diverse Material Systems

TL;DR

The paper tackles data efficiency in training machine-learned interatomic potentials by conducting a rigorous, multi-system evaluation of active learning strategies across Materials Project and OQMD. It introduces a neural network ensemble with four query strategies (random, uncertainty, diversity, hybrid) and demonstrates that diversity sampling most consistently yields strong formation-energy predictions, notably achieving a $10.9\%$ improvement for the Ti–O system with $p = 0.008$, while reducing labeled data by $5$–$13\%$. The approach achieves these gains within a practical 4-hour Google Colab workflow and with less than 8 GB RAM, enabling broader participation in MLIP development; the authors also provide open-source code and experimental configurations. The results offer practical guidelines for data-efficient MLIP training and point to future directions such as integrating symmetry-aware neural architectures to further reduce data requirements. Overall, the work establishes a robust, transferable framework for AL-driven MLIPs across diverse materials, with concrete evidence of when and why diversity-based selection excels.

Abstract

Efficient discovery of new materials demands strategies to reduce the number of costly first-principles calculations required to train predictive machine learning models. We develop and validate an active learning framework that iteratively selects informative training structures for machine-learned interatomic potentials (MLIPs) from large, heterogeneous materials databases, specifically the Materials Project and OQMD. Our framework integrates compositional and property-based descriptors with a neural network ensemble model, enabling real-time uncertainty quantification via Query-by-Committee. We systematically compare four selection strategies: random sampling (baseline), uncertainty-based sampling, diversity-based sampling (k-means clustering with farthest-point refinement), and a hybrid approach balancing both objectives. Experiments across four representative material systems (elemental carbon, silicon, iron, and a titanium-oxide compound) with 5 random seeds per configuration demonstrate that diversity sampling consistently achieves competitive or superior performance, with particularly strong advantages on complex systems like titanium-oxide (10.9% improvement, p=0.008). Our results show that intelligent data selection strategies can achieve target accuracy with 5-13% fewer labeled samples compared to random baselines. The entire pipeline executes on Google Colab in under 4 hours per system using less than 8 GB of RAM, thereby democratizing MLIP development for researchers globally with limited computational resources. Our open-source code and detailed experimental configurations are available on GitHub. This multi-system evaluation establishes practical guidelines for data-efficient MLIP training and highlights promising future directions including integration with symmetry-aware neural network architectures.

Figures (3)

• Figure 1: Active Learning Methodology Pipeline showing the iteration sequence. The workflow begins with data retrieval and feature engineering, followed by an iterative active learning loop: (1) Initialize with 30 labeled samples, (2) Select query strategy, (3) Train ensemble of 5 neural networks on current labeled set, (4) Evaluate on held-out test set (MAE, $R^2$), (5) Query batch: select 15 most informative samples using chosen strategy, (6) Augment labeled set. This loop repeats 6 times, growing labeled set from 30 to 105 samples.
• Figure 2: Learning curves comparing active learning strategies across four material systems. Each panel shows MAE (eV/atom) versus number of labeled samples. Four strategies are compared: Random (pink), Uncertainty (orange), Diversity (green), and Hybrid (teal). Shaded regions indicate $\pm 1$ standard deviation over five random seeds. Carbon: Diversity achieves the lowest final MAE. Silicon: all methods converge to similar performance. Iron: Diversity outperforms Random. Ti--O: largest strategy separation, with Diversity providing substantial advantage.
• Figure 3: Cross-Database Validation Results for Carbon system. Left: Models trained on Materials Project (MP) and tested on OQMD. Right: Models trained on OQMD and tested on Materials Project. The asymmetric transfer performance highlights domain shift between databases, with MP$\rightarrow$OQMD transfer showing better generalization as the labeled set grows.