A physics-informed neural network modelling methodology to analyse integrated thermofluid systems

Physics-informed neural networks (PINNs) were developed to overcome the limitations of acquiring large training datasets that are commonly encountered when using purely data-driven machine learning methods. This study explores a PINN modelling methodology to analyse steady-state integrated thermofluid systems based on the mass, energy, and momentum balance equations, combined with the relevant component characteristics and fluid property relationships. The PINN methodology is applied to three thermofluid systems that encapsulate important phenomena typically encountered in integrated thermofluid systems modelling, namely: (i) a heat exchanger network with two different fluid streams and components linked in series and parallel; (ii) a recuperated closed Brayton cycle containing various turbomachines and heat exchangers, and (iii) a simplified boiler consisting of a furnace and radiative convective superheater. The predictions of the three PINN models were compared to benchmark solutions generated via conventional, physics-based thermofluid process models. The largest average relative error across all three models is only 0.93%, indicating that the PINN methodology can successfully be implemented to generate accurate solutions using the non-dimensionalised forms of the balance equations. Furthermore, it was shown that the trained PINN models provided a significant increase in inference speed compared to the conventional process models. The PINN modelling methodology was then extended to develop a surrogate model for the heat exchanger network. An additional surrogate model was developed for comparison using a data-driven multilayer perceptron (MLP) neural network. The MLP surrogate model was able to interpolate accurately. However, its predictive performance declined when making predictions for samples that fell outside of the range of training data. Despite various refinements, the PINN surrogate model could only be trained successfully for datasets that contained up to five unique samples. This limitation could not be resolved within the scope of the present study and should be investigated further. The accuracy of the PINN surrogate model degraded significantly when used to extrapolate beyond the training envelope. Due to the constraint on the number of training samples, it is impossible to draw a general conclusion regarding the extrapolation ability of the PINN concept. In spite of its current limitations, the significant increase in computational speed offered by the PINN modelling methodology when used to analyse integrated thermofluid systems suggests that this is a promising modelling technique that should be explored further.

Grantor dc:publisher.institution

Department of Mechanical Engineering

Year dc:date.issued

2024

Author dc:creator

• Laugksch, Kristina Karli

Advisors dc:contributor.advisor

• Rousseau, Pieter
• Laubscher, Ryno

• Engineering

Language dc:language.iso

en

Handle dc:identifier.uri

http://hdl.handle.net/11427/41003

OAI identifier oai:identifier

oai:open.uct.ac.za:11427/41003