Physics-Informed Machine Learning Modelling of Hydrokinetic Energy Harvester

Hydrokinetic energy converters are environmentally friendly and hold significant potential for bridging the energy gap in remote villages where mainstream electrical grid infrastructure is unavailable. However, the design of these energy conversion systems is complicated by the Multiphysics nature of fluid–structure interactions. Bluff bodies, such as cylinders commonly used in these converters, often exhibit complex fluid–structure interactions known as vortex-induced vibrations (VIV). Traditional low-order models can capture the basic physics of these systems but often struggle to account for nonlinearities and real-world complexities. On the other hand, purely data-driven models may suffer from overfitting or require large volumes of data to perform effectively. This paper presents a physics-informed machine learning (PIML) approach that integrates a simplified VIV model with a neural network trained to learn the unmodelled dynamics from experimental heave displacement measurements. We demonstrate the implementation of this hybrid model using PyTorch and validate its performance with experimental data of a cylinder’s heave response in fluid flow. The proposed PIML model effectively compensates for the unmodelled dynamic characteristics, achieving an 80% improvement in performance. The mean squared error (MSE) decreased from 149.92 in the physics-only model to 25.74 in the PIML model.