Weight initialization in physics-informed neural networks to enhance consistency of mass-loss predictions of plant cells undergoing drying

This study aimed to highlight the significance of weight initialization towards the consistency of Physics-Informed-Neural-Network-based (PINN-based) predictions for spatiotemporal problems in engineering and science. Accordingly, here a PINN was developed for mass transfer analysis (i.e., PINN-MT) for a single plant cell undergoing drying. While solving Fick's laws of diffusion for a cell domain and predicting mass loss and moisture concentration based on convective mass transfer at the cell wall boundary, PINN-MT utilizes moisture concentration at fresh state of the cell as an initial condition. The governing equations, boundary conditions, and initial conditions were incorporated to the corresponding PINN through the loss function. Residuals of these equations and initial-and-boundary conditions were minimized during the training process of PINN, which can predict moisture concentration variations in time and space domains. However, spatiotemporal problems typically involve many tunable hyperparameters that can make the training process more complicated, leading to inconsistent predictions and loss-convergence problems. This is uncommon in the context of traditional computational approaches. To address this complexity associated with PINNs, pre-trained weight initialization can be adopted, enhancing the ability of PINN-MT to provide consistent solutions via automatic differentiation. In this context, this study assessed the effectiveness and efficiency of PINN-MT coupled with weight initialization to address training complexities and to provide consistent solutions for spatiotemporal problems in engineering and science.