Integrating Molecular Semantics and Three-Dimensional Geometry for Critical Property Prediction: A Multimodal GNN-BERT Framework

Accurate prediction of critical temperature, pressure, and volume is essential for thermodynamic modeling and process safety design, yet remains challenging for complex molecules under data-scarce conditions. Here, we develop a multimodal GNN-BERT framework that integrates SMILES-based chemical semantics with two-dimensional topology and three-dimensional molecular geometry for critical property prediction. BERT captures molecular sequence information, while graph neural networks learn topology- and geometry-aware representations through message passing. Evaluation on 913 chemical compounds demonstrates that the proposed framework consistently outperforms conventional machine-learning models and single-modality baselines. Importantly, comparative analyses among BERT, BERT+2D-GNN, and BERT+3D-GNN reveal that incorporating three-dimensional molecular geometry provides a consistent 5–10% improvement across critical temperature, pressure, and volume prediction. Additional validation using random forest and support vector regression further confirms that the predictive contribution of 3D molecular information is not architecture-dependent. These results highlight three-dimensional molecular geometry as an important structural parameter for data-driven critical property prediction and provide a reliable computational strategy for thermodynamic estimation and chemical process safety applications.