We propose a new score-based approach to generate 3D molecules represented as atomic
densities on regular grids. First, we train a denoising neural network that learns to map from …