Explainable and trustworthy artificial intelligence for correctable modeling in chemical sciences

Area of Science:

• Multidisciplinary research at the intersection of data science, physics, chemistry, and engineering.
• Development of advanced computational and mathematical frameworks for scientific modeling.

Background:

• Traditional data science and machine learning methods are inadequate for small, correlated, and noisy datasets common in physics, chemistry, and engineering.
• Existing approaches lack systematic frameworks for integrating expert knowledge into physics-based models under uncertainty.

Purpose of the Study:

• To develop a mathematical and computational framework for probabilistic artificial intelligence (AI)-based predictive modeling.
• To combine data, expert knowledge, multiscale models, and information theory using uncertainty quantification and probabilistic graphical models (PGMs).
• To provide predictive guarantees for PGMs and demonstrate their application in chemistry.

Main Methods:

• Development of a novel framework integrating AI, uncertainty quantification, and probabilistic graphical models (PGMs).
• Incorporation of expert knowledge, multiscale models, and information theory.
• Application of the framework to a microkinetic model of the oxygen reduction reaction.

Main Results:

• A demonstrated framework for probabilistic AI-based predictive modeling applicable to small, correlated datasets.
• Development of predictive guarantees for probabilistic graphical models (PGMs).
• Successful application to a specific chemical reaction (oxygen reduction reaction).

Conclusions:

• The proposed framework offers explainable results, leading to correctable and trustworthy AI models.
• This approach effectively combines diverse information sources (data, expert knowledge, physics-based models) for enhanced predictive power.
• The framework addresses critical limitations of current methods in scientific data analysis and modeling.