ML Modelling on Prediction of Residual Strength of RC Column Exposed to Fire by FEM Numerical Data

According to national fire agency of Korea, there had been over thirty thousand of fire cases for an year only in Korea. Despite of this nonnegligible figure of fire cases and the vulnerability of structure to fire exposure, fire analysis in structure is frequently marginalized. So far fire analysis for structure were conducted under FEM approach, which has been verified to simulate the behavior of structure element under fire with high accuracy. One of the reasons FEM fire analysis is not regarded as essential one is the low accessibility to fire analysis, since it requires a lot of models and coefficients to be determined. This research was devoted to enhance the accessibility problem of fire analysis, by modelling the residual strength of fire damaged Reinforced Concrete (RC) column with ML regressors, utilizing FEM generated data. Before generate FEM data, the ground RC column sample set need to be defined. Following Korean design code and practice, 1770 RC column sample set was defined. Each of those RC column samples were parametrized with 4 sectional variables of B, H, BN, HN, representing width, height, and number equally spaced rebar along width and height span, respectively. Then established fire analysis was conducted on this RC column sections, resulting residual strength of each RC column sections from 0 minute to 190 minutes fire exposure conditions with every 10 minutes. The residual strength of each time step was reduced to 4 points on P-M interaction diagram. Thus total of 1770 * 20 = 35540 data points were used as dataset, 80% of them for train set and remaining for test set. For ML modelling, 5 most representative ML regression model was chosen, that is Support Vector Machine (SVM), Artificial Neural Network (ANN), Random Forest (RF), eXtreme Gradient Boost (XGB), and Light Gradient Boosting Machine (LGBM). Performance of each model was compared with respect to both accuracy and reliability. It is concluded that LGBM with partial monotone constraints worked the best, and XGB followed.