Neural Surrogate-Based Operating-Window Identification for Slit-Die Extrusion

Resumen

Slit-die extrusion requires the simultaneous control of pressure drop, wall shear rate, apparent viscosity, and post-die swell. This work presents a neural surrogate-based workflow for identifying feasible operating windows in slit-die extrusion. A structured synthetic dataset was generated using a reduced-order physics model that combines generalized-Newtonian slit flow, temperature-dependent Carreau–Yasuda rheology, and an empirical area-swell closure. An ensemble of independently initialized multi-output neural networks was trained to predict pressure drop, apparent viscosity, wall shear rate, and area swell ratio, while a separate ensemble of neural classifiers was used to classify feasible and non-feasible operating conditions under pressure, shear-rate, and swell constraints. The regression ensemble achieved R² values above 0.9988 for all outputs, and the feasibility classifier reached an accuracy of 0.9838 and an F1-score of 0.9884. The trained models screened 50000 candidate conditions, identified feasible operating windows, and verified the top recommendations against the reduced-order physics solver. The results show that neural surrogate learning can provide a fast decision-support tool for constrained operating-window exploration in extrusion process design.