Bayesian inference for resin transfer moulding using a neural network surrogate

The permeability of the fibre preform designed for resin transfer moulding is a critical parameter for predicting resin flow and defect detection, but it can vary within the mould and between samples. We propose a Bayesian inference methodology for predicting the permeability of the preform based on pressure measurements made during fill time. We present two sets of numerical experiments: one in which we consider a homogeneous domain and estimate race tracking strengths, and another in which we estimate the permeability field that can vary throughout the domain. We use a neural network surrogate to emulate the permeability-pressure relationship typically modelled by a computationally expensive and discontinuous moving boundary finite element model. Incorporating the approximation error between the neural network and the finite element model allows for Bayesian posteriors that incorporate the truth but can be computed at a fraction of the cost.