Automating the modelling and simulation process benefits all related areas, such as research, design, manufacturing and especially the training of machine learning models. This approachsignificantly improves the modelling efficiency and speed, providing scalability in data generation, which motivates the discovery of new patterns. In addition, consistency and reproducibility through unified modellingmethods improve accuracy by eliminating manual mistakes in repetitive processes. Machine learning provides an efficient approach foraccessing large-scale simulation data and providing insights into unknown circumstances. Three-sided protection steel beams arecommon yet rarely investigated section types in offshore oil platforms. With the top side of the upper flange exposed to fire, large convective and radiative heat fluxes are induced, leading to a rapidly descending temperature diagram along the web. The increase in thelever arm caused by the decrease in the elastic neutral axis under this temperature distribution and the decrease in the elasticity of the compressed flange cause earlier lateral torsional buckling failure compared to that of the 4-sided protection beam. To provide a close temperature profile, an integrated framework automating modelling, simulation, data processing and machining learning sequence using the ABAQUS kernel scripting method, Python and MATLAB is proposed. The modelling method automates the model generation process with inputs from a parameter text file and establishes restraints using the edge contact-detection algorithm for unusual shapes. Second, the model files (.inp) are submitted to ABAQUS, and MATLAB controls the simulation process. The output data are extracted and written into .csv files. Third, the extracted data are divided by Python code into data batches and fed to machine learning models for training. All IPE, universal beam sections and 54 weldedsections from a realistic oil processing structure with different protection limits are tested with the protection thicknesses calibrated with an optimisation program with a 5 °C allowance. The gradient boosting method achieves a root mean square error of 1.34 °C compared to the simulation results. The calculation time of the developed software with a graphical user interface is also tested with various numbersof temperature points and output intervals.