An automated framework for optimising aluminium extrusion dies using data-driven surrogate models

Abstract Traditional extrusion die design methods rely heavily on empirical design rules, trial-and-error corrections, and finite element (FE) simulations. These methods are typically iterative and time-consuming, leading to long design cycles and high costs. To address these inefficiencies, this research proposes a data-driven optimisation framework for aluminium extrusion dies—demonstrated on a T-bar profile—which utilises machine learning-based surrogate models to predict and enhance the die performance by minimising the curvature factor ( CF ), an innovative metric for assessing extrusion flow uniformity. Die designs were systematically parametrised, generating a comprehensive dataset of design variables and corresponding performance metrics through a combination of design of experiments and FE simulations. Gradient-based, population-based, and reinforcement learning optimisation algorithms were implemented to iteratively adjust die parameters, guided by the optimisation objective, CF . Among evaluated regression models, a scalar-to-scalar multi-layer perceptron, optimised through hyperparameter tuning, demonstrated the highest accuracy in predicting CF , with further performance improvements achieved with active sampling strategies. All investigated optimisation methods, although with different computational efficiency, successfully reduced CF values as verified by FE simulations, with genetic algorithms delivering the optimal die design for flow uniformity. The approach used in this study, i.e. data-driven surrogate models integrated with advanced optimisation techniques, offers an efficient solution for automating extrusion die design.