A graph-convolutional neural network for addressing small-scale reaction prediction.

Area of Science:

• * Computational chemistry
• * Artificial intelligence in chemistry
• * Machine learning for chemical reactions

Background:

• * Machine learning models are increasingly used for predicting chemical reactions.
• * Graph-convolutional neural networks (GCNs) and transformer models are prominent architectures in this field.
• * Performance differences between these models can be significant, especially under data-limited conditions.

Purpose of the Study:

• * To compare the reaction prediction performance of GCN and transformer models.
• * To investigate the impact of data availability on model performance.
• * To evaluate model efficacy using the Baeyer-Villiger oxidation reaction as a case study.

Main Methods:

• * Development and application of a graph-convolutional neural network (GCN) model.
• * Comparative analysis with a transformer model for chemical reaction prediction.
• * Experimental validation using the Baeyer-Villiger oxidation reaction dataset.

Main Results:

• * GCN models demonstrate comparable reaction prediction capabilities to transformer models when sufficient data is available.
• * Under limited data conditions, the GCN model achieved a top-1 accuracy of 90.4%.
• * The transformer model achieved a top-1 accuracy of 58.4% under the same limited data conditions.

Conclusions:

• * GCN models offer a robust alternative for reaction prediction, particularly in data-scarce scenarios.
• * The study highlights the superior performance of GCNs over transformers when training data is limited.
• * Findings suggest GCNs are a valuable tool for accelerating chemical discovery and process optimization.