Attention-based multi-head feature-fusion network: a generalised method for hot deformation behaviour prediction in low-alloy steels

Abstract

Modelling hot deformation of low alloy steels is important for optimising processing efficiency and cost reduction. Existing approaches lack generalisation as they primarily focus on single steel grades, ignoring chemical composition. To address this, a dataset comprising 58 distinct low-alloy steels and an Attention-Based Multi-Head Feature Fusion Network (AMHFnet) has been established and proposed. AMHFnet uses multi-head residual modules and a head weighting mechanism to adaptively learn high-dimensional representations of both chemical composition and processing variables, which are then fused and processed via a feature filtering and repeat step decision mechanism to predict hot deformation response. It demonstrated superior accuracy and generalisability in 10-fold cross-validation compared to existing decision tree-based models and state-of-the-art deep learning models. Ablation studies demonstrate the effectiveness of individual components of the network. An investigation of the effect of carbon demonstrated that AMHFnet could reliably reflect elemental effects on hot deformation behaviour, further validating its reliability and applicability. The model successfully captures work hardening behaviour and dynamic recrystallisation for most compositions. By accurately predicting hot deformation behaviour across diverse low-alloy steels, this work can simplify the alloy-composition design process and target the experimental testing.