Predicting grain size in extruded AA6063 profiles: A unified approach based on finite element analysis and machine learning

The evolution of grain size in AA6XXX extruded profiles is a critical factor for enhancing mechanical, thermal and surface properties. Traditional methods for microstructure control rely on extensive experiments requiring significant time and resources. To address this issue, the present work proposes a method for microstructure prediction combining numerical data from Finite Element Method (FEM) simulations with experimentally acquired microstructure data to train an Artificial Neural Network (ANN) capable of predicting grain size. Data was acquired for three distinct AA6063 aluminum alloy profiles extruded under various process conditions in terms of profile and tool geometry, ram speed, billet pre-heating temperature and extrusion ratio, representing a diverse and heterogenous dataset comprising grain size (55-228 mu m), strain (2.8-28), maximum strain rate (2-190 s-1), exit temperature (480-580 degrees C), Zener-Hollomon parameter (4 x 1015-4 x 1017) and Stored Energy (170-480 kJ/mol*K) for training and testing different ANN configurations. The final trained ANN was able to accurately predict grain size in regions of normal grain growth but was less reliable at foreseeing formation of the largest and smallest grains due to limited data points within this range. A Mean Absolute Percentage Error (MAPE) of 13.9% was achieved for predictions in the test set with an ANN comprising two fully connected layers with 9 and 19 neurons, respectively, Rectified linear unit (ReLU) activation functions and a ridge L2 penalty term of 10-6 for regularization. The presented methodology provides a foundation for the development of new data-driven approaches aimed at facilitating microstructure prediction in industrial settings.