Abstract
The research into the use of machine learning methods for predicting the fatigue life of materials illustrates the great potential of current methods to improve the prediction of the performance of materials. However, deriving reliable confidence intervals for the predicted values remains a challenge because of the very significant difficulties in modeling materials. For these reasons, this thesis proposes a new solution for uncertainty quantification (UQ) in fatigue life prediction of metallic materials using an approach which is based on the knowledge of fatigue life modeling in physics and then combined with traditional machine learning models. The experimental validation of this method on the fatigue life of titanium and carbon steel alloys demonstrates the effectiveness of the above-mentioned method. The experimental results show that the addition of the physically based model to the data-driven model increases the consistency of the predicted values very significantly and improves the estimation of the uncertainty intervals to a great extent.At the same time, as the use of data-driven methods in materials science is now becoming more common, the importance of reliable UQ of predictor variables for informed decision making is a very important aspect. UQ methods that do not use throughput models based on physical information, such as Gaussian Process Regression (GPR), have very significant limitations when modeling functions with spurious covariates or anisotropic smoothness. Recently, Bayesian neural networks (BNN) have emerged as a new UQ method that provides a probabilistic framework for capturing uncertainty in neural networks. This study applies Bayesian neural networks to data-driven modeling for fatigue life prediction. Based on the research results on fatigue life of metallic materials, it is shown that BNN can produce accurate point and uncertainty estimates that outperform traditional methods such as GPR. The proposed BNN framework, especially the one based on Markov chain Monte Carlo approximation, provides more reliable results than the variational inference-based BNN.