Fig. 1: Overall training pipeline of SpbNet.

SpbNet consists of four main parts: (1) Descriptors. SpbNet adopts raw material structure and potential energy surface (PES) basis functions as input. PES basis functions are regarded as a multi-channel three-dimensional (3D) image and linearly combined through a linear layer to obtain a single-channel 3D image. (2) Model architecture. A Graph Neural Network (GNN) extracts material structure features from the raw material structure. i, j, k represent the atoms in the material. The combined PES is processed with a Vision Transformer (ViT). Structural features are incorporated into the ViT via the cross-attention mechanism. (3) Pre-training. The model is pre-trained via the prediction of global geometric features and atom-level properties including number of atoms and Signed Distance Function (SDF). A Class⁴⁹ ([CLS]) token is appended to the PES patches to predict the global features via dedicated prediction heads, including Topology type (Topo), Pore Limited Diameter (PLD), Largest Cavity Diameter (LCD), Void Fraction, etc. Features of PES patches are used to predict SDF and number of atoms through separate output heads, including SDF Head and ANP (Atom Number Prediction) Head. (4) Fine-tuning. After pre-training, the model retains the backbone weights while discarding all pre-training heads. A new task-specific prediction head is added to predict the target property. Downstream tasks cover guest molecule-related properties, intrinsic material properties, and generalization to out-of-distribution material systems. Trainable parts are marked in blue background.