General Chemically Intuitive Atom- and Bond-Level DFT Descriptors for Machine Learning Approaches to Reaction...

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry
• Materials Science

Background:

• Predicting reaction conditions is crucial for chemical synthesis and process optimization.
• Current methods often rely solely on structural information, limiting predictive power.
• Density functional theory (DFT) provides valuable atom- and bond-level insights.

Purpose of the Study:

• To evaluate the effectiveness of general atom- and bond-level DFT descriptors in enhancing machine learning models for reaction condition prediction.
• To compare the performance of models using hybrid (DFT + structural) descriptors against purely structural models.

Main Methods:

• Treated reaction condition prediction as a multiclass classification task.
• Utilized neural networks and random forests trained on a large dataset (69,935 reactions, 296 condition classes).
• Compared models with varying input embedding compositions, including structural and hybrid DFT descriptors.

Main Results:

• Hybrid models achieved comparable or superior performance (weighted precision, top-1/top-3 accuracy) with up to 71% fewer parameters than structural models.
• Improvements of 5-11% in weighted precision, top-1 accuracy, and F1 score were observed for hybrid neural networks.
• The best hybrid model outperformed the best purely structural model, despite the latter using a much larger unsupervised embedding dataset.

Conclusions:

• General DFT descriptors are highly effective in enhancing machine learning models for reaction condition prediction.
• Hybrid representations combining DFT and structural information offer a more efficient and accurate approach.
• This strategy significantly advances the capabilities of predictive modeling in chemistry.