SINGLE WOOD PARTICLE DRYING: A COMPARATIVE STUDY OF PHYSICS-BASED AND PHYSICS-INFORMED NEURAL OPERATOR MODELS
TL;DR
A physics-informed deep learning approach for simulating coupled heat and mass transfer during the drying of wood particles, a critical unit operation in modern biomass energy processing, offers a fast, reliable tool for real-time monitoring and optimization of wood drying systems.
Abstract
To support the optimized control for the modern drying system, this paper presents a physics-informed deep learning approach for simulating coupled heat and mass transfer during the drying of wood particles, a critical unit operation in modern biomass energy processing. Firstly, a physics-based reference model was developed using Fick’s second law with moisture-dependent effective diffusivity to capture internal evaporation kinetics. Afterward, to overcome the computational bottleneck of standard partial differential equation (PDE) solvers, a Physics-Informed Neural Operator (PINO) framework was evaluated against the physical benchmark. Despite significant temperature PDE losses due to strong nonlinearities in surface heat balances and coupled thermal dynamics compared to moisture transport, the PINO model exhibits excellent predictive agreement with the reference model across parameter space, with the deviation in drying time lower than 5%. It physically resolves both fundamental drying periods: the constant-rate drying period governed by the adiabatic saturation temperature and the falling-rate period following surface moisture depletion. The proposed PINO surrogate offers a fast, reliable tool for real-time monitoring and optimization of wood drying systems.