Data-driven prediction and optimization of high-strength austenitic steels for cryogenic applications

초록

An artificial neural network (ANN) model was developed as a predictive framework for estimating the mechanical properties of cryogenic austenitic steels. A comprehensive dataset comprising 82 literature-derived steel compositions and their corresponding tensile properties at cryogenic temperatures was employed to establish nonlinear relationships between alloy composition, testing temperature, and mechanical performance. A singlehidden-layer ANN with five neurons achieved the best predictive performance. Following hyperparameter optimization (momentum = 0.6, learning rate = 0.6, 1800 iterations), the optimized model achieved high prediction accuracy (R2 = 0.9004, Pearson's r = 0.9486) for the test dataset, demonstrating excellent agreement between predicted and experimental yield strength (YS) and ultimate tensile strength (UTS). Ten-fold crossvalidation confirmed the robustness of the ANN model, yielding average test R2 and Pearson's r values of approximately 0.90 for both YS and UTS predictions, along with low prediction errors. A comparison with several widely used machine learning models also confirmed the superior predictive performance of the optimized ANN model. Sensitivity analysis using the trained ANN revealed both individual and synergistic effects of alloying elements, highlighting the dominant role of interstitial elements such as carbon and nitrogen, whereas copper and titanium exhibited comparatively lower contributions. The model also enables quantitative evaluation of stepwise alloying additions, providing insights into composition-property relationships. The generalization capability of the model was further assessed using an independent validation steel that was not included during model development, where prediction errors remained below 5% at both 77 K and 4.2 K. Furthermore, the ANN framework was employed for virtual alloy design and identified a candidate cryogenic steel composition with predicted values of 1660 MPa for YS and 2004 MPa for UTS at 4.2 K, exceeding the highest-strength alloy in the experimental dataset. Overall, the developed ANN model provides a reliable and efficient tool for the prediction and preliminary design of cryogenic austenitic steels with tailored mechanical properties.

키워드