Fig. 1: Details of VQCrystal model.
From: Massive discovery of crystal structures across dimensionalities by leveraging vector quantization
a Overview of the VQCrystal model. The model is divided into four main parts: local, global, quantization, and output. The local part consists of a simple transformer encoder that extracts local features. The global part captures global features using CSPNet (detailed in (d)) and a GCN block applied to the local features. After pooling, these blocks generate the global feature. The quantization part is elaborated in (b). The output part includes two decoders: the output decoder (detailed in (c)) and the property decoder on the left, which predicts the corresponding property from the extracted features. The red line means the procedure is used in sampling. b Quantization process for global and local features. Both the global and local features are passed through their respective codebooks, where they are quantized by a lookup-based replacement approach. The codebooks map the input features to their closest codebook entries. c The details of the decoder component. The decoder consists of a classic Transformer block, which includes LayerNorm (LN), multi-head attention with query-key-value (QKV) attention mechanism, and a lattice net to predict a lattice of crystals. d The details of the CSPNet component, which includes LayerNorm (LN) and intersection blocks to enhance feature interaction and improve representation learning.
