Enhancing constitutive modeling and workability analysis via deformation history-informed recurrent neural networks: A case study on 2024 aluminum alloy

Abstract

Accurate prediction of the thermo-mechanical processing (TMP) response of metals is essential for optimizing their industrial manufacturing processes. However, existing constitutive models are limited in capturing the non-linear flow behavior arising from complex deformation histories, resulting in poor predictive accuracy and generalization ability. This work presents a novel machine learning (ML)-based end-to-end constitutive modeling framework that directly incorporates the nonlinear effects of non-costant deformation history on future flow stress. Using the hot forming of 2024 aluminum (Al) alloy as a case study, its TMP behavior was systematically investigated through hot compression tests under temperatures of 300–450 °C and strain rates of 0.01–10 s⁻¹. Two ML models—artificial neural network (ANN) and long short-term memory (LSTM)—were trained and benchmarked against the traditional Arrhenius-type model. Owing to inherent ability to encode sequential data, the LSTM model achieves significantly improved predictive accuracy and generalization ability, especially under large-strain and non-constant thermo-mechanical loading conditions. This approach also enables a more reliable hot workability characterization of metals, providing new insights into the causal relationship between TMP parameters and microstructural evolutions, including flow instability and various dynamic recrystallization. The proposed framework represents a significant step forward in data-driven modeling of complex TMP responses, offering practical guidance for the design and control of advanced metals processing.