Gradient-enhanced neural networks for model parameter estimation applied to flow chemistry automated platforms

Abstract

The acceleration of chemical process development through flow chemistry depends on obtaining reliable kinetic models. Model-based design of experiments (MBDoE) has been successfully applied with flow chemistry platforms for estimating reaction rate parameters for low-complexity models. However, its use with dynamic experiments involving computationally expensive models remains challenging, particularly when experimental conditions must be suggested in real time. Surrogate models can approximate complex models, but often lack the ability to reproduce the derivatives of the original model, which are essential for MBDoE as a gradient-based approach. This work investigates the use of gradient-enhanced neural networks as surrogate models for parameter estimation within an MBDoE framework. By incorporating gradient information, the surrogate model is able to reproduce the local sensitivity structure of the original first-principles model, ensuring both predictive accuracy and reliable inverse behaviour during parameter estimation. A case study on competitive-consecutive reactions demonstrates that artificial neural networks (ANNs) that were trained with gradient information can be used for parameter estimation and yield confidence regions comparable to those of the original model, while substantially reducing computational cost by a factor of approximately 200,000. This enables closed-loop, sequential MBDoE suitable for real-time applications.