(532ec) Incorporation of Covariance of DFT Energy Ensembles into Gnn Models for the Trustworthy Catalyst Design Method

The cancellation of error successfully mitigates the systemic errors aroused by the approximations in the computational chemistry methods. In recent work, we showed that the GemNet-T model can predict the relative adsorption energies of the Open Catalyst 2020 (OC20) dataset with high accuracy of ~0.35 eV MAE. Despite the high accuracy, the correlation between the systems is not preserved. An analogous idea to the error-cancellation in density functional theory (DFT) can sophisticate the graph neural network (GNN) model for catalysts design. The adsorption energies across a wide range of catalyst compositions in the OC20 dataset are calculated by using the Bayesian error estimation functional with van der Waals correlation (BEEF-vdW) functionals, which offers the estimate of uncertainty by generating the energy ensemble. The covariance of pairwise energy differences is encoded into the loss function of the GNN model so as to give emphasis to the systems whose electronic structures are highly correlated. Furthermore, a pair of neural network models mainly based on the pre-trained GemNet-T model are combined to establish a twin neural network architecture. Our approach enables the model to capture the similarity of the electronic structures by preserving the known covariance from the DFT. We also substantiate that our covariance-incorporated neural network model outperforms the previous GemNet-T model in terms of the uncertainty calibration and the preservation of the known correlation, while not losing its accuracy.