TRACER: a reliability-first GemNet baseline for trustworthy computational materials discovery.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Graph neural networks (GNNs) are crucial for computational materials discovery.
• Existing GNN models face challenges in robustness and uncertainty quantification (UQ).
• A standardized framework for reproducibility, fairness, and dependability is lacking.

Purpose of the Study:

• Introduce TRACER, a transparent and repeatable reliability-first pipeline for GNN-based materials discovery.
• Evaluate TRACER's robustness, accuracy, and UQ capabilities.
• Provide a framework for dependable and confidence-aware computational materials discovery.

Main Methods:

• Developed TRACER, a pipeline built on a GemNet-based GNN.
• Investigated robustness through sensitivity analysis (graph cutoff, depth, data fraction).
• Benchmarked UQ using deep ensembles and compared heuristics for error identification.

Main Results:

• Achieved competitive accuracy with a single GemNet model, showing a 25.8% MAE reduction vs. ALIGNN on JARVIS-DFT.
• Deep ensembles provided informative uncertainty estimates, correlating strongly with error.
• TRACER demonstrated operational utility, outperforming random selection in identifying high-error cases under budget constraints.

Conclusions:

• TRACER offers a reproducible and confidence-aware approach to computational materials discovery.
• The pipeline generalizes well, confirmed by performance on Matbench Perovskites.
• TRACER provides a strong predictive framework for reliable materials exploration.