Interpretable AI and Knowledge Discovery

We develop methods that discover mathematical relationships and candidate physical explanations from data. By combining symbolic regression, engineering knowledge and experimental design, we investigate how interpretable models can support scientific understanding and process development.

The challenge

An accurate prediction does not necessarily explain why a process behaves as it does. Engineers also need relationships they can inspect, compare with established knowledge and use to guide further experiments. This is especially important when data are limited or when a model will be applied beyond the conditions used to develop it.

We develop methods that turn process measurements into candidate equations and physical explanations. Our work connects interpretable AI with reaction engineering, formulation engineering, and biochemical modelling, with an emphasis on understanding the relationships behind observed behaviour.

Our approach

Discovering equations from small datasets

Symbolic regression searches for mathematical expressions that describe relationships in data. We combine this search with engineering knowledge and kinetic model structures so that the resulting expressions can be assessed as parts of a process model. Recent bioprocess work investigates how individual kinetic terms can be identified and refined when only small datasets are available.

Transferring model structure

We also investigate structural model transfer: modifying the equations of an existing model when applying it to a different biochemical system. Combining neural-network feature attribution with symbolic regression helps identify where the original model needs revision. This approach has been demonstrated through an in-silico case study.

Choosing informative experiments

Experimental design is another part of the discovery process. Our published framework combines symbolic regression with model-based design of experiments to choose measurements that distinguish competing expressions while considering process performance. In simulated formulation studies, this iterative approach recovered the mechanisms used to generate the data.

Connecting kinetics with physical mechanisms

Complementary reaction-engineering work links measurable kinetic characteristics to otherwise inaccessible surface behaviour. Analytical relationships between reaction orders, activation energies and catalyst surface coverages help connect macroscopic measurements with microscopic mechanisms. The framework was tested in simulated water–gas shift reaction studies. Candidate explanations still require validation against independent evidence.