Plant growth is a dynamic process affected by genes and growing environment, with all kinds of interactions between them. These complex relationships make the prediction of plant growth challenging. We propose a hybrid modelling framework that combines a logistic ordinary differential equation model with a Long Short-Term Memory (LSTM) neural network, resulting in a Physics Informed Neural Network (PINN). While PINNs have been widely applied to physical dynamical systems, their use in modelling the dynamics of plant growth systems is still largely unexplored. We illustrate the construction of a PINN on plant height data in wheat and compare its performance with alternative models for longitudinal plant data. All temporal prediction models only require time and temperature as input. Among a set of competing models, our PINN had the lowest average root mean squared error (RMSE) of prediction and the smallest standard deviation across multiple random initialisations. Therefore, we conclude that incorporating biological growth constraints into data-driven growth models can enhance prediction accuracy of longitudinal plant traits.
Shao et al. (Sun,) studied this question.