Neural network potentials trained by repeatedly checking where models disagree on hydrogen sticking to copper

The work

Technical · from the record

The study builds machine-learning interatomic potentials for molecular hydrogen scattering on four copper facets — Cu(111), Cu(100), Cu(110) and Cu(211) — using an adaptive sampling loop in which committees of three models flag configurations they disagree about during scattering dynamics, those configurations are clustered and labelled with SRP48 density functional theory, and the models are retrained. Four iterations grew the training set from 2530 to 4230 data points, and sticking probabilities were obtained from 10,000 trajectories per model, facet, rovibrational state and collision energy. The invariant SchNet model required four iterations before its ground-state sticking probabilities approached published quasi-classical reference data, and its potential energy surface cuts remained non-smooth; the equivariant PaiNN model, trained on the same datasets, matched the reference probabilities closely from the initial dataset and gave smooth energy landscapes. For the best final models the paper reports PaiNN energy errors three times lower in MAE and more than four times lower in RMSE, and force MAE and RMSE more than five times lower, than SchNet.

How AI was used

Two message-passing neural network interatomic potentials, SchNet (SchNetPack v1.0.0) and PaiNN (SchNetPack dev branch), were trained from scratch on SRP48 DFT total energies and forces for 6-layer copper slabs with and without hydrogen, with a combined energy–force loss (final weights 0.05 and 0.95) and hyperparameters chosen by five-fold cross-validation, settling on 7 interaction blocks, 512 features and a 4 Å cutoff. The initial dataset was assembled from ab initio molecular dynamics in FHI-aims plus H2 scattering trajectories run with a SchNet potential trained on an externally supplied dataset. Training data were then grown adaptively: reactive scattering molecular dynamics in NQCDynamics.jl was run with a committee of three models differing in train/test split, the standard deviation between their energy predictions was evaluated every fifth step, and structures exceeding a manual threshold (0.03 eV, then 0.025 eV) were collected, described by inverse H–H and H–surface distances, reduced by principal component analysis, grouped with k-means clustering in scikit-learn, manually screened, labelled with DFT and added to the dataset before retraining. The same model committees were run as the force engine for the scattering trajectories from which sticking probabilities and their epistemic uncertainties were computed, and for potential energy surface cuts and minimum energy path calculations.