Prediction of accelerated cooling rates in H-section steels using artificial neural networks

This study presents a model utilizing artificial neural networks (ANN) to predict cooling rates during the accelerated cooling heat treatment of section profiles. The experimental studies focused on S275HEA120, S275HEB120, and S275HEB140 profiles to assess three cooling rates: Web (W- o C/s), Web + Flange (W + F o C/s), and Flange (F o C/s). Variables included pressure (4 bar, 8 bar, 12 bar), distance to web (60 mm–210 mm), and application time (10 s, 20 s, 30 s, 40 s), leading to the acquisition of experimental results. Each of the three cases for the ANN was analyzed for cooling rates ( W , W + F , F ). The training and testing experimental data were separated by percentage, ranging from 67.53 % to 73.26 % for training and from 26.74 % to 41.03 % for testing. We investigated hidden layer neuron numbers (HLNN) of 5, 10, 20, 30, 40, and 50. R 2 , MSE, and MEP (%) were used to evaluate ANN training performance, with R 2 ranging from 0.9453 to 0.9984, MSE from 11.57 to 0.05511, and MEP from 0.548 % to 0.0682 %. The ANN obtained R 2 values of 0.9463–0.9981, MSE values of 11.33 to 0.053, and MEP (%) values of 1.54 %–0.0093 % during testing. The HLNN set to 30 or 40 performed best for prediction. Out of range HLNN decreased performance.