Predicting residual wall thickness in water-assisted injection molding

Task

Water injection technology (WIT) is a special variant of injection molding used to produce parts with complex hollow sections, such as bicycle frames, handles or fluid lines. Once the cavity is filled, injected water displaces the still-molten core.

The residual wall thickness left behind defines the part's properties. It depends on a large number of process and material parameters. Process simulations are time-consuming, and finding suitable boundary conditions takes many runs.

The project addressed two questions. How can the wall thickness distribution be determined for new parameter sets without running another simulation? And which process setting produces a required wall thickness — the inverse problem?

Approach

The study used a generic demonstrator part with an overflow cavity. Plastic Innovation ran process simulations with varying mold wall temperature (60–125 °C), fluid flow rate (150–550 cm³/s), fluid temperature (20–40 °C) and delay time (0.1–0.9 s).

Each result was evaluated for residual wall thickness. The part was split into segments, and the distance between the hollow channel and the outer wall was measured. This yields the minimum and maximum wall thickness, their locations, and the thickness at defined positions.

Digital Physics AI trained a multilayer perceptron (MLP) on these simulation data using the no-code software dAIve Predict. For the inverse prediction, a network with 3 hidden layers of 32 neurons each proved effective. Its inputs are the required wall thicknesses and the fixed process parameters; its output is the delay time.

For the forward prediction, the same data set was restructured. The four process parameters became the inputs, and the wall thickness at 300 evenly spaced positions along the part became the output. No additional simulations were needed.

Result

The predicted wall thickness distribution was compared with a reference simulation for a previously unseen parameter set. According to NAFEMS Magazin 1/2024, predicted and simulated wall thicknesses agreed to approx. 98 %, measured by the mean absolute percentage error (MAPE).

The required wall thicknesses at the two specified positions were met with an accuracy of approx. 95 %, both in simulation and in the AI prediction (ibid., also MAPE).

The process setting for a target wall thickness can therefore be determined directly, without solving an optimization problem across many simulation runs. The results are based on simulation data; validation on real parts or with machine data is still pending.