To predict the moisture desorption behavior of a composite structure in outer space, it is necessary to know the through-thickness distribution of moisture absorption rate in the composite material at the time of launch. However, it is impossible to identify the moisture distribution state by measuring the weight change of the assembled large-scaled structure. In this study, we proposed a method for estimating the moisture distribution state of a composite space structure from the information obtained during the moisture desorption test of a witness sample. This sample had the same moisture distribution as the target structure as it was exposed to the same hygrothermal environment. We used a physics-informed neural network (PINN), which is a neural network with physics-based constraints, for estimation. This neural network model takes time and through-thickness location as input and returns the local moisture concentration. PINN trains the model by enforcing that the output of the network fulfills the diffusion equation and the change in moisture content during the moisture desorption test of the witness sample. First, a numerical experiment using finite element analysis was performed to confirm the effectiveness of the proposed approach. Furthermore, a demonstration experiment was conducted using carbon-epoxy laminates to confirm the practicality of the method.
