Online Gate-Driven Flow Control in Resin Transfer Moulding Using a Neural-Network Surrogate

What the math gives to ML

The paper contains a constructive reduction for learning controlled free-boundary dynamics: under quasi-static Darcy filling, the filled region at a fixed time depends on the cumulative, rather than pointwise, gate-pressure history. This converts a high-dimensional time-dependent control input into a lower-dimensional path descriptor and can simplify surrogate-model training, especially when controls are piecewise constant or rapidly varying. The transferable asset is the exact sufficient-statistic construction: replace a control trajectory by its time integral and test whether the resulting representation preserves the network target. This is best implemented as an integral-control encoder or cumulative-control tokens, with an explicit ablation against raw time-series conditioning.