Cross-material stress-strain prediction: A Seq2Seq transfer approach with small punch data

Accurate determination of stress-strain relationships in metallic materials has traditionally relied on extensive, costly experiments, especially when dealing with cross-material scenarios where data for certain materials are limited. To address these challenges, this study proposes a cross-material sequence-to-sequence (Seq2Seq) framework that predicts stress-strain relationships from Small Punch Test (SPT) data. The model transfers knowledge from high-strength steel to aluminum alloy with limited data. This domain selection reflects industrial practice, where high-strength steel has long been widely used with abundant data available, whereas aluminum alloys often suffer from data scarcity. To enhance feature extraction, raw SPT data are transformed into Gramian Angular Field (GAF) representations, and integrated with a high-order moment matching (HoMM) algorithm and a cross-attention algorithm to mitigate domain discrepancies between source and target materials. Experimental validation on both simulated datasets and real-world samples confirms the robustness and accuracy of the proposed approach. The model achieves an average mean absolute error (MAE) of 4.53±3.43 MPa and a mean absolute percentage error (MAPE) of 0.91%±1.04% on high-strength steel. For real aluminum alloys, the prediction errors are as low as 2.86 %, 8.63 %, and 8.66 %. Additional ablation studies highlight the critical contributions of each model component to improving generalization across materials. The proposed framework offers a practical and accurate solution for cross-material stress-strain prediction, even when only limited data are available.