fig1

Figure 1. Architecture of GNN used in this work consists of a main module responsible for predicting energy and forces, and a loss prediction module that estimates the predicted loss, $$ \hat{\mathcal{L}} $$. In each message passing block, the invariant component, the “0e” channel, is extracted and subsequently concatenated after passing through a FC layer. The loss prediction module employs standard nonlinear functions, ReLU and Softplus, which introduce nonlinearity and enforce stable positive outputs. GNN: Graph neural network; FC: fully connected; ReLU: rectified linear unit; GT: ground truth.