Determination of Fatigue Strength of Steel Using Deep Learning Techniques

Steel is one of the primary components of a structure from a strength perspective. Since the load applied to some structures is cyclic, fatigue strength becomes a crucial factor for durability. However, due to the complex nature of fatigue strength formation in steel, it is not yet completely understood. Fortunately, the evolution of new technologies like artificial intelligence and machine learning has led to the development of regression models that can accurately predict the fatigue strength of steel. Although machine learning models have a defined structure, their performance improvement is limited after reaching a certain threshold. However, this limitation can be overcome with neural networks (NN), which are highly customizable and improve with the addition of data. In this study, seven different NN models were constructed with successive enhancements to predict the fatigue strength of steel with high accuracy. Various performance metrics were utilized to evaluate the models. The dataset used in this project is the fatigue dataset for steel from Japan NIMS (National Institute for Materials Science), which is an open-source and the world's largest dataset on steel fatigue strength. It comprises 437 instances with 25 different factors influencing the property, such as chemical composition (9), heat treatment conditions (12), and upstream processing details (4). These instances include 371 carbon and low-alloy steels, 48 carburizing steels, and 18 spring steels. The data covers various heats of each steel grade and different processing conditions. After testing various neural networks and determining the best train-test split, a maximum accuracy of 99.7% was achieved on the test data. Furthermore, the model was deployed as a web application, making it accessible to everyone.