LG-Transformer: learned-graph transformer framework enabling diverse physicochemical properties prediction toward...

Area of Science:

• Chemical Engineering
• Computational Chemistry
• Artificial Intelligence

Background:

• Accurate prediction of green fuel properties is vital for decarbonizing transportation.
• Existing artificial intelligence (AI) models struggle with interpretability and utilizing inter-molecular information.
• This limits their generalizability for diverse fuel property predictions.

Purpose of the Study:

• To develop an interpretable deep learning framework for predicting fuel physicochemical properties.
• To enhance the utilization of internal and external molecular information for property prediction.
• To improve the generalizability of AI models for diverse fuel property forecasting.

Main Methods:

• A novel deep learning framework, the learned graph feature fusion Transformer (LG-Transformer), was developed.
• LG-Transformer constructs an inter-molecular relationship graph using contrastive learning, topological descriptors, and property similarity.
• A comprehensive fuel property database with 1850 molecules and 17 properties was utilized.

Main Results:

• LG-Transformer achieved a superior predictive performance with an average R-squared of 0.900.
• The model significantly outperformed existing graph neural networks (GNNs) and other deep learning baselines.
• Interpretability analyses revealed key molecular structure-property relationships.

Conclusions:

• LG-Transformer offers a powerful and interpretable approach for predicting fuel properties.
• This framework advances AI applications in green fuel design and optimization.
• The study highlights the potential for improved engine performance and reduced emissions through accurate fuel property prediction.