Fig. 2. The framework of SDEGen. At training time, given the graph G and conformation R, we: (1) sample the time from [0,1] uniform distribution and utilize the Gaussian random feature to encode the time information to the model, then this temporal feature is mapped together with the perturbed distance to form the d̃ (2) Map the edge(E) and atom(A) features from molecules to form corresponding embeddings (light blue and slightly darker blue arrows) (3) utilize the GNN model to encode the graph structure to the model (d̃|G) and train the SDEGen network with denoising score matching (dark blue arrows). The procedure amounts to learning the evolutionary state of the molecule in the stochastic dynamics system at the given time.