Moreira Minete E, Immertreu M, Teichmann F, Müller S (2026)
Publication Language: English
Publication Status: Submitted
Publication Type: Journal article, Review article
Future Publication Type: Journal article
Publication year: 2026
Article Number: 113310
URI: https://arxiv.org/abs/2510.25697
DOI: 10.1016/j.rineng.2026.113310
Open Access Link: https://arxiv.org/abs/2510.25697
In metal casting, mold-filling hydrodynamics strongly influence defect formation and final component quality. However, systematic exploration of gating configurations remains limited by the computational cost of repeated transient Computational Fluid Dynamics (CFD) simulations. This study investigates whether a neural operator can provide faster short-horizon field predictions for parametrized mold-filling problems. We develop a non-autoregressive Fourier-Graph neural operator that extends the Geometry-Informed Neural Operator (GINO) framework to transient multi-field prediction through spatio-temporal latent construction, inlet and gating parameter conditioning, multi-field decoding, and causal rollout weighting. The model is evaluated on two controlled 2D isothermal water-air Cahn-Hilliard-Navier-Stokes (CHNS) datasets of increasing parametric complexity. For held-out parameter combinations within the sampled fixed-cavity design grid, the mean relative L2 errors are 29.79% on DS1 and 35.92 % on DS2. Matched-horizon inference is 90.55 × faster than CFD over 0.1 s and 40.52 × faster over 0.6 s. Qualitative comparisons show broad agreement in the velocity, pressure, and volume-fraction fields, although pressure transients and localized gradients remain challenging to resolve. Moreover, performance deteriorates under spatial or temporal subsampling and reduced training-data availability. Overall, the results demonstrate computational acceleration for controlled in-grid interpolation, while the field-wise relative L2 errors make the current accuracy limitations explicit. The present model is not a stand-alone substitute for high-fidelity CFD in demanding settings and has not yet been validated on unseen cavity geometries. Nevertheless, with further development, neural operators of this type could support early-stage mold design by rapidly screening designs and reducing the number of final high-fidelity CFD evaluations.
APA:
Moreira Minete, E., Immertreu, M., Teichmann, F., & Müller, S. (2026). Fourier Neural Operators for Two-Phase, 2D Mold-Filling Problems Related to Metal Casting. Results in Engineering. https://doi.org/10.1016/j.rineng.2026.113310
MLA:
Moreira Minete, Edgard, et al. "Fourier Neural Operators for Two-Phase, 2D Mold-Filling Problems Related to Metal Casting." Results in Engineering (2026).
BibTeX: Download