Construct Diverse, High-Quality Benchmark Test Sets for U-MLIPs

Develop systematic procedures for constructing diverse and high-quality benchmarking test sets for universal machine-learned interatomic potentials (U-MLIPs) that enable evaluation of physically meaningful properties such as surface energies, elastic moduli, and defect energetics, while mitigating risks of overfitting to implicit targets during model optimization.

Background

The paper discusses recent efforts to benchmark universal machine-learned interatomic potentials (U-MLIPs) and emphasizes the need to evaluate models beyond energy and force errors, focusing on physically meaningful properties including surface energies, elastic moduli, and defect energetics. Despite these efforts, the authors note that building suitable test sets is difficult and caution against overfitting when test datasets become implicit optimization targets.

This problem is central to establishing reliable and fair assessments of U-MLIPs across diverse materials and tasks, ensuring that benchmark evaluations capture generalization and physics-informed behavior instead of narrow metric optimization.