Prediction of fatigue life of 6082-T6 aluminum alloy using machine learning

Aluminum alloys are among the materials of interest in many sectors, particularly the automotive and aerospace sectors. One of the primary reasons for their popularity is their light weight compared to steel. One of the key reasons why aluminum alloys, such as 6082-T6, are particularly popular is their adequate mechanical properties for many automotive components. Machine parts operating under cyclic loads are subjected to fatigue and eventually fracture. Knowing the number of cycles within which aluminum alloys, commonly used in the aerospace and automotive sectors, will undergo fatigue fracture is crucial for life safety. Fatigue tests are routinely performed in industry to determine the number of cycles before fatigue fracture occurs. Experiments conducted with realistic loads to simulate the operating environments of machine parts take a significant amount of time, resulting in wasted time and increased energy consumption. The primary objective of this study is to reduce the duration of repeated fatigue tests in industry. The 6082-T6 aluminum alloy, frequently used in the automotive industry, was selected for fatigue tests. Axial force, axial elongation, and fatigue cycle count data were collected from fatigue tests conducted using 6082-T6 aluminum alloy. Additionally, electrical data was simultaneously collected from the sample using the DCPD (Direct Current Potential Drop) method. With this method, a constant 1 ampere current was applied to the sample, and voltage and current values were recorded over time at four points on the sample. In the second phase of the study, machine learning models were developed using the collected data, and predictions were made using the developed models. Decision trees, extra trees, random forest, KNN, and XGBoost algorithms were used in the first phase of the study. The first phase's predictions were sample-based. The cycle count of the sample, determined as the target variable, was estimated based on the selected input variables. One of the samples was then designated as the test sample. Using the remaining four samples, an ensemble model created with an LSTM and LGBM-based approach was trained. To simulate the operating environment, the first 50% of the test sample data was fed to the trained model and asked to estimate the sample's lifespan. In the final stage of the study, an LSTM and PINN model was developed using a physics-based approach. The Paris-Erdoğan crack propagation law was used for physical constraints. With this approach, the developed model was forced to comply with the laws of physics while making data-based predictions. When the results were evaluated, the decision tree algorithm performed best in sample-based predictions, while the KNN algorithm performed worst. In the remaining stages of the study, the LSTM and LGBM ensemble model yielded an R2 value of approximately 0.896. The model developed with a physics-based approach achieved an R2 value of 0.942, an improvement of approximately 4%. In conclusion, data collected using fatigue testing and the DCPD technique on aluminum alloys is shown to be an effective method for predicting the number of cycles to fatigue fracture using machine learning methods, and this method can shorten fatigue tests. Furthermore, the success of physics-based machine learning models with noisy and sparse data is highlighted.