Comprehensive Molecular Representation from Equivariant Transformer

We implement an equivariant transformer that embeds molecular net charge and spin state without additional neural network parameters. The model trained on a singlet/triplet non-correlated \ce{CH2} dataset can identify different spin states and shows state-of-the-art extrapolation capability. We found that Softmax activation function utilised in the self-attention mechanism of graph networks outperformed ReLU-like functions in prediction accuracy. Additionally, increasing the attention temperature from $\tau = \sqrt{d}$ to $\sqrt{2d}$ further improved the extrapolation capability. We also purposed a weight initialisation method that sensibly accelerated the training process.