Abstract

Completed geometry optimization histories may support early prediction of final electronic energies, but improvements over current energy can overstate the value of structural machine learning. Seven previously generated adsorption trajectories were reconstructed into 423 unique states. Endpoint priors, energy-based ridge regression, compact geometry features, and interface-centered smooth overlap of atomic positions were compared using nested validation that excluded complete sites. Over records 3–20, the structural model reduced the mean absolute error from 12.340 to 4.262 kilocalories per mole, but the training-endpoint median achieved 3.053 kilocalories per mole. The structural model outperformed this median in only one of seven sites. A common-model evaluation of unseen site pairs yielded 46.3% ordering accuracy, compared with 47.4% for current energy. Prediction-stability triggers occurred in all seven structural-model trajectories, but none of the triggered geometries met all recorded convergence criteria. Entry into a one-kilocalorie-per-mole endpoint band preceded full geometric convergence by 13–85 records. The results of these retrospective measurements show that endpoint-prior controls, common-model ranking tests and separate geometric-convergence tests are important. They are not shown to offer an additional advantage of the tested structural representation to endpoint priors, reliable determination of the most suitable structural representation, or validated improvement in density functional theory computation time.