Modeling Hardness in Industrial C-Mn Cast Steels with Neural Network Models

초록

This study develops an Artificial Neural Network (ANN) model to predict the hardness of industrial C-Mn and low-alloy cast steels using chemical composition and heat-treatment parameters. The optimized 19-20-20-1 architecture achieved about 95% accuracy for training and 90% for testing datasets. Error analysis confirmed high reliability, with 78.5% of training and 74% of testing samples showing prediction errors below 2%, and only 6% exceeding 6%. Although the testing R-2 value was relatively low due to repeated hardness values (148-155 BHN dominating most samples), the model achieved low mean absolute (3.30 BHN) and percentage (2.15%) errors, indicating strong predictive agreement. Sensitivity analysis identified carbon, soaking time, and cooling time as the key factors affecting hardness. The proposed ANN provides a cost-effective, data-driven framework for virtual experimentation, alloy design, and process optimization in industrial steel casting and heat-treatment applications.

키워드