The ability to reasonably predict the response of steel structures under fire effects is of great importance in structural design for fire safety. This paper presents the predictions of neural networks for the axial load capacity of steel columns in fire. Using ANN and LSTM. The goal is to develop ANN and LSTM models to predict the axial load capacity of steel columns in a fire. In addition to the prediction by (ANN, LSTM), using MATLAB to perform a parametric study on some input parameters that affect the axial load capacity of steel columns in fire, such as ((T- (temperature), B- (width dimensions), H- ( Dimensions (depth), t- (thickness), L- (dimensions for length), stress rate, e- (Salenderness), Fu, Fy, Ex103)) and compare the results from the developed models with the existing ones. This paper presents the use of long-term memory (LSTM) as well as an artificial neural network (ANN) model with three algorithms (LM), (B-R), and (S-C) to approximate the fit of the samples. To learn the empirical data associated with the prediction of the axial load capacity of fire steel columns loaded into a model using 320 samples, a backpropagation-trained neural network was added to the model. 10 input parameters form the basis of the proposed ANN and LSTM model. The proposed sample ratio design is expected to reduce the number of laboratory and field experiments, save labor and material costs, as well as save time as it provides improved accuracy. Engineered columns are expected to be more robust and efficient, and Microsoft Excel Solver was used to create a neural network model analysis; The accuracy of these models, such as those generated by LSTM, and ANN is ultimately determined by a realistic estimated value. The proposed architecture is adopted, which is the optimal neural network architecture, with minimal error. The input layer of the proposed neural network consists of 10 variables, and the results obtained from the neural networks indicated the success of this technique in predicting the axial load capacity of steel columns in fire. Both ANN and LSTM were reasonably successful in estimation. The best version of the advanced neural version is represented by (LSTM) where the accuracy reached (99.237%), followed by the results obtained from normal neural networks in the (Bayesian normalization algorithm). The ANN (II) and(I) models were able to achieve average accuracy ratios of (99.897%) and (99.913%), and after retraining (99.911%) and (99.91%), respectively.
