Hybrid Learning of Transport Equations with Differentiable Neural Solvers from Experimental Data

Arthur Jessop, Mohammed Alsubeihi, Ashwin Kumar Rajagopalan, Ben Moseley

Keywords: Hybrid Modelling, Experimental Data, Automatic Differentiation, Scientific Machine Learning, Process Engineering, Transport Equations

TL;DR: In this paper, we 1) present a SciML framework for empirical law discovery for transport equations, and 2) demonstrate and release a curated dataset appropriate for SciML and analysis on data suitability for learning empirical laws.

Abstract: Transport equations are used across chemical engineering to model dynamic systems. However, these often rely on hand picked empirical equations to describe underlying physical laws, with no guarantee that these will accurately fit data - resulting in trial and error. We demonstrate an end-to-end differentiable solver that is used to learn underlying physical laws from real experimental data by integrating hybrid components into a mechanistic model. Compared to an industry standard benchmark model and solver, we show that our hybrid modelling approach achieves better model predictions without the need for additional data, and drastically outperforms the benchmark in both forward simulation and training. This approach can lead to improvements in the efficient design, operation, and control of processes in the fine chemicals and pharmaceuticals industry.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 36