Physics-informed machine learning for process modelling and optimization

Physics-informed neural networks (PINNs) enforce physical laws that are described by general nonlinear partial differential equations during training. This approach drastically reduces the data demand and prevents overfitting. We explore the potential of physics-informed neural networks in bioengineering.

Digital technologies such as artificial intelligence (AI) and machine learning (ML) can play a pivotal role in advancing the development of innovative chemical process design. However, data-driven methods often lack the domain knowledge that is well-established in engineering and scientific fields. To improve the reliability and interpretability of the models, this research endeavours to bridge the gap between powerful data-driven methodologies rooted in deep learning and domain-specific knowledge derived from fundamental principles in the context of chemical engineering.

We aim to apply hybrid methodologies founded on advanced artificial intelligence methods and first principle knowledge to the modelling and optimization of complex chemical processes. We employ cutting-edge techniques such as physics-informed neural networks (PINNs), constrained learning and hybrid modelling to describe the multiscale behaviour of chemical systems, such as from microscopic physical phenomena, to macroscopic unit operation and plant level.

Systematic process design based on rigorous mathematical (superstructure) optimization is often computationally tedious and still relies on engineering experience and heuristics. We propose to address the challenge by embedding robust high-fidelity physics-informed surrogate models into optimization routines with the objective of reducing problem complexity and accelerating computational speed of available solvers.