What Does the Machine Learn? Knowledge Representations of Chemical Reactivity.

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry

Background:

• Data-driven machine learning (ML) models are increasingly used for chemical reactions, achieving statistical success.
• However, these ML models often function as "black boxes," limiting their ability to provide fundamental chemical insights.
• This lack of interpretability hinders the advancement of chemical understanding, unlike traditional chemistry approaches.

Purpose of the Study:

• To deconstruct black-box machine learning models for chemical reactions.
• To investigate the knowledge acquisition process within these models.
• To identify the origins of statistical accuracy in ML models lacking physical principles.

Main Methods:

• Analysis of diverse chemical reaction datasets using various machine learning techniques.
• Experimentation with different chemical representations to understand model learning.
• Systematic classification of reaction types and application of Evans-Polanyi relationships.

Main Results:

• Statistical accuracy in ML models can arise even without explicit incorporation of physical chemical principles.
• A minimal, chemically intuitive model was developed, independent of machine learning.
• This intuitive model relies on reaction-type classification and Evans-Polanyi relationships.

Conclusions:

• Black-box ML models for chemical reactions can be deconstructed to reveal underlying chemical logic.
• A simplified, interpretable model based on chemical principles offers deeper understanding than complex ML models.
• Expert interaction with simplified models can enhance their reliability and chemical interpretability.