Online gate-driven flow control in resin transfer moulding using a neural-network surrogate

In resin transfer moulding, the preform must be fully saturated before resin reaches the vents, or dry spots form. Race-tracking, which varies between nominally identical parts and is unobservable in a closed mould, makes the flow non-uniform. We estimate the race-tracking strengths sequentially from pressure-sensor data with an iterated extended Kalman filter, and re-optimise the pressure at an auxiliary injection gate at each measurement to prevent premature vent arrival. For real-time operation, neural network surrogates replace the finite element-control volume model for the sensor pressures and the terminal unfilled area, and the Bayesian approximation error framework accounts for the error of the pressure surrogate. A result proved here makes surrogate training tractable: for quasi-static Darcy flow, including heterogeneous and anisotropic permeability, the resin domain depends on the injection history only through the time-averaged gate pressures, so constant pressures reproduce the terminal state of any admissible policy. On a forked geometry with six race-tracking regions, control raises successful fills from 22 to 43 of 100 scenarios, and from 85 to 213 of 500 in an independent test set. High-fidelity control gives 46, not significantly more, and control with the true race-tracking strengths 49, against 52 scenarios that some constant pressure fills: the gate and vent layout limits any controller, and within that limit the loss is mainly due to estimation. A measurement every second over-contracts the posterior, because the surrogate error is correlated in time; longer intervals restore much of the calibration without significantly affecting control.