Data-efficient stress–strain prediction using optimized explainable LSTM network

Accurately predicting the stress–strain curves in hot stamping steels is crucial for process optimization but remains challenging as collecting sufficient data of high-temperature tensile tests is time-consuming. While machine learning techniques appear promising, they typically require extensive datasets, limiting their practicality. Hence, we developed a comprehensive data-efficient framework that enables accurate flow curve prediction for 1.5-GPa-grade hot stamping steel using feature engineering and long short-term memory network architecture optimizations. The developed models were validated through prediction tasks involving both interpolation and extrapolation conditions, with the optimized model achieving high accuracy (R 2 = 0.9798). Shapley additive explanations (SHAP) analysis was conducted to identify the mechanisms through which each methodology enhances model performance. The integration of explainable AI throughout the modeling process reveals how feature engineering and architecture modifications enhance prediction performance under data constraints. The proposed framework provides an interpretable and generalizable approach for data-efficient modeling of material properties.