Induction self-heating insert for mold temperature control evaluated with experiments, simulation, and neural network model

Injection molding of thin-walled parts lives or dies by how fast the mold cavity surface can be heated and cooled. A new study in PLOS One reports an experimental and computational evaluation of an induction self-heating insert that couples a thin heating zone directly to a cooling layer, combining laboratory measurements, a sequentially coupled electromagnetic-thermal simulation, and an Artificial Neural Network (ANN) surrogate model.

The research team, led by Thanh Trung Do, Tran Minh The Uyen, and Pham Son Minh, built a design space of 25 Taguchi L25 geometries defined by three parameters: T, W, and L. They then added X and Y coordinates so that the ANN could predict not just a single temperature value but the entire surface temperature field of the insert.

Three cooling states were tested: no water (NW), stagnant water (SW), and flowing water (FW). Experimental targets for the neural network were taken as the mean of five repeated tests at each condition, a step intended to reduce measurement noise in the training data.

The multiphysics simulation reproduced the location of the hot zone on the insert surface. Among the four cases with reported quantitative differences, three differed from experiment by no more than about 10 degrees Celsius, while one case showed a peak difference of about 20 degrees Celsius, indicating generally good but not perfect agreement.

To prevent data leakage during model validation, the researchers kept all 846 data points from each geometry together in a geometry-wise validation scheme, so that the network was never trained and tested on points from the same insert design.

The ANN’s predictive accuracy varied with cooling state. The coefficient of determination (R²) was 0.746 for the no-water case, 0.653 for stagnant water, and 0.493 for flowing water, with root mean square errors of 17.03, 12.97, and 12.62 degrees Celsius, respectively.

Interestingly, the effects of the three geometric parameters T, W, and L depended on which cooling state was active and were not consistently monotonic. This means that an insert geometry optimized for one cooling configuration may not perform best under another, underscoring the value of a surrogate model that can explore the combined geometry-cooling design space quickly.

As a proof of concept, the team ran a molding test with low-density polyethylene using a Chip 2 part. The insert achieved complete filling, with a mean feature spacing of 28.6 micrometers against a nominal 30 micrometers, and a mean feature height of 106.6 micrometers against a nominal 100 micrometers.

The authors conclude that combining insert geometry, cooling-water state, and a neural network surrogate can support local thermal design in injection molding, with the potential to delay frozen-layer formation and improve filling in thin-wall parts. The work remains at the proof-of-concept stage, and the moderate ANN accuracy under flowing-water cooling suggests room for further model refinement before industrial deployment.

Subject of Research: Induction self-heating mold inserts and neural network modeling for rapid temperature control in injection molding

Article Title: Geometry and cooling effects on an induction self-heating insert for rapid mold temperature control: Experimental, Multiphysics, and artificial neural network-based analysis

Article References: Do, T. T., Uyen, T. M. T., & Son Minh, P. (2026). Geometry and cooling effects on an induction self-heating insert for rapid mold temperature control: Experimental, Multiphysics, and artificial neural network-based analysis. PLOS One, 21(10), e0359849. https://doi.org/10.1371/journal.pone.0359849

Image Credits: AI Generated

DOI: 10.1371/journal.pone.0359849

Keywords: injection molding, induction heating, mold temperature control, artificial neural network, multiphysics simulation, Taguchi design, thin-wall molding, frozen layer, cooling water, surrogate modeling, LDPE, PLOS One