Fig 1

Overview of T2MAT's three primary modules: ⅰ) Capturing material design requirements from user-input (blue background); ⅱ) Generating a variety of novel material structures based on user-specified properties (orange background); ⅲ) Implementing a fully automated DFT validation process (green background)."

Fig 2

Architecture of CGTNet. The self-attention layers are designed to capture dynamically and efficiently capture long-range interaction. Graph features are constructed from atom features, interatomic distances, and angles. Edge features, processed through a fully connected layer, are integrated into the self-attention mechanism for enhanced information exchange."

Fig 3

Performance of CGTNet. MAE of DimeNet++, CGCNN, SchNet, PaiNN, GemNet, and CGTNet on testing data using 25%, 50%, 75% and 100 % training data for the prediction of (a) bulk modulus, (b) HSE band gap, (c) maximum phonon spectrum frequency, (d) electrical conductivity, (e) SLME, and (f) exfoliation energy."

Fig 4

Necessity, architecture and performance of contrastive learning. (a–c) Linear relationship between the InfoNCE loss function and the predictive performance on testing data. (d) Architecture of contrastive learning for minimizing distances between positive structures and maximize those between negative ones. (e) Performance comparison of CGTNet with and without contrastive learning."

Fig 5

Example of the GNN explainer for interpreting GNNs in T2MAT. (a)Visualization of the node and edge importance in predicting the SLME of MAPbI3. (b) Contribution of individual elements in improving the SLME, which is summarized based on 9770 GNN-based SLME predictions. The bluer an element's color, the greater its positive contribution to the SLME."

Fig 6

GUI of T2MAT. The interface includes the submission of material design requirements, modifications of the input file, display of available generative models, property-prediction models and generated structures."