Scientific AI and Digital Twins

We combine physical models with machine learning to predict chemical and biological process behaviour. Our work investigates how hybrid models can be developed with limited data, adapted between systems and scales, and used as a foundation for process digital twins.

The challenge

Reliable process models support the design of experiments, prediction of product quality and selection of operating conditions. Building them is difficult when reaction mechanisms or biological behaviour are only partly understood. Industrial measurements can also be noisy, and experiments at larger scales are expensive.

We investigate how physical knowledge and machine learning can be combined to build useful models under these conditions. A central question is how much mechanistic detail to retain, and where a data-driven component can add information that the physical model does not capture.

Our approach

Combining engineering knowledge with data

Our hybrid models combine engineering relationships, such as material balances and kinetic expressions, with statistical or machine-learning components. These components can represent uncertain reaction rates, correct missing behaviour or provide predictions that support a mechanistic simulation. The structure is chosen around the available knowledge, measurements and intended use of the model.

Adapting models between processes and scales

We also examine how models can be adapted between related bioprocesses and reactor scales. Rather than constructing each model independently, transfer methods reuse information from an existing system and update it with measurements from the new setting. Published work investigates this approach for bioprocess digital twin development and for industrial-scale prediction using limited laboratory and pilot-scale data.

Building a foundation for digital twins

These models provide a foundation for digital twins: computational representations that can support monitoring, prediction and operational decisions. Our research addresses model development and adaptation, including the balance between mechanistic structure and data-driven flexibility.