Mean absolute error (meV/atom) of an ML model's predicted formation energy or total energy compared to DFT reference calculations. The primary accuracy benchmark for ML property prediction models. State-of-the-art models achieve ~20-30 meV/atom on Materials Project datasets.