Constructing and explaining machine learning models for chemistry: example of the exploration and design of boron-based Lewis acids

Interpretability can outperform complexity in specialized domains: The study found that explainable ML models using chemically meaningful descriptors achieved better accuracy (mean absolute error < 6 kJ/mol) than conventional "black-box" deep learning models when working with limited data, suggesting that simpler, interpretable approaches may be more practical for specialized chemistry applications.

Explainable AI reveals actionable chemical insights: By using interpretability analyses, researchers identified specific, actionable ways to modify boron-based Lewis acids—such as changing the nature and positioning of substituents—demonstrating that explainable models can directly guide molecular design rather than just making predictions.

Bridging human expertise and machine learning improves outcomes: The research explicitly demonstrates how ML models designed to align with "the chemist's way of thinking" can enhance scientific understanding of chemical reactivity, suggesting that responsible AI development should incorporate domain expert knowledge rather than treating it as separate from algorithmic design.

Transparency enables scientific validation: By using chemically meaningful descriptors (ab initio computed features and Hammett parameters) rather than opaque numerical representations, the models' predictions could be understood and validated by chemists, supporting the principle that AI systems in science should be interpretable to domain experts.