Preprint: Hybrid Physics-Informed Neural Networks for Automated Kinetic Parameter Calibration in Bio Manufacturing: A Human-in-the-Loop Case Study

Abstract

Simulations of bio manufacturing processes such as microbial fermentation are governed by kinetic parameters. For example, the maximum rate at which a microorganism consumes sugar must be re-calibrated for every substrate composition before a mechanistic process model can predict how a culture will behave. Today this calibration is often performed manually by domain experts and may take hours. This paper presents a human-in-the-loop hybrid physics-informed neural network (PINN) that couples a fixed mechanistic bioprocess model with a lightweight network mapping sensor trajectories to five kinetic parameters. A central finding is that the data scientist alone, despite access to state-of-the-art optimisation tools, is systematically blocked at every critical juncture by incorrect domain assumptions i.e., initial conditions, feed concentrations, time-unit conventions, and substrate assignments. These remain invisible to the training algorithm. Two training approaches are compared: a piecewise model that resets at each measurement and a continuous model trained on the full trajectory. The piecewise model achieves low training error but performs worse than the original mechanistic model. The continuous model performs better, improving the baseline by up to 31.4% on unseen experiments. Final results from the expert interactions lead into a dual-role checklist for data scientists and domain experts undertaking hybrid PINN development on physical systems.