Derivative-Free Domain-Informed Data-Driven Discovery of Sparse Kinetic Models

Area of Science:

• Chemical Engineering
• Applied Mathematics
• Data Science

Background:

• Data-driven kinetic models are crucial for process design but are sensitive to experimental noise.
• Existing methods for learning dynamical systems from data struggle with noisy reaction kinetics.

Purpose of the Study:

• To develop a robust method for inferring interpretable kinetic models from noisy reaction data.
• To improve the accuracy and reliability of data-driven dynamical models by incorporating domain knowledge.

Main Methods:

• Introduced a derivative-free sparse identification technique (DF-SINDy) that approximates integrals instead of derivatives.
• Incorporated domain information, including mass balance and chemical principles, into the model discovery process.
• Validated the method using synthetic data with varying noise levels, sampling frequencies, and experimental counts.

Main Results:

• DF-SINDy identified models with lower errors compared to the standard SINDy method.
• The inclusion of domain knowledge significantly improved the recovery of correct kinetic terms.
• Demonstrated improved reliability in interpreting discovered models.

Conclusions:

• DF-SINDy offers a more robust approach to learning kinetic models from noisy data.
• Integrating domain knowledge enhances the accuracy and interpretability of data-driven kinetic models.
• This work advances the development of chemistry-agnostic, interpretable kinetic models for complex reaction networks.