Jernej Klemenc

A surrogate neural network model for fatigue assessment and optimisation of an exhaust system is presented in this study. The approach is based on established fatigue analysis tools and employs a parameterised sample geometry. Conventional numerical methods were employed to generate a sufficiently large sample set and to provide stress field data. Fatigue damage calculations were subsequently performed as a post-processing step to prepare input data for the surrogate model. The resulting surrogate model is capable of predicting both the maximum fatigue damage value and its spatial location directly from geometric parameters. Due to the complexity of the prediction task and the wide range of damage values, extensive effort was devoted to model optimisation. Proprietary data pre-processing techniques proved essential for effective neural network training, and the network hyperparameters were tuned to achieve satisfactory predictive performance. To further address the wide range of damage values, a linear regression step was applied following inverse data transformation. As sample generation is computationally expensive, the influence of sample size on prediction accuracy was also investigated. The proposed surrogate methodology enables efficient fatigue assessment and is suitable for iterative design and optimisation workflows.

The study reports on the essential level of details in simulations during the development of structural components if reliability-based design is used to ensure their quality and operational safety. A general method, which is initially introduced, is then applied to an indicator spring of a fuse element during assembly and operation stages. First, it is proven that design of simulations based on orthogonal arrays which includes variations of form, material properties and operating conditions within expected scatter limits provides a comparable determination of the scale parameter for the two-parameter Weibull distribution as the experimental observations of the same process. The shape parameter of the distribution tends to be underestimated by the simulations resulting in a higher scatter of the expected properties than experimentally measured. Next, it is shown that the maximum likelihood method to determine representative parameters of the scatter of assembly and operation stages provides a better match with experimental data than the median rank regression. Finally, a high reliability of the indication has been calculated for the fuse element if both the scatter of the assembly and the operation conditions were considered.

This paper focuses on optimisation of material parameters to describe the elastoplastic stress–strain relationship in finite element solvers. Two new methods are introduced to minimise the numerical error that occurs in the interspace between the experimental cyclic stress–strain curve and its representation using multilinear interpolation. Specifically, both methods are integrated into a Prandtl operator approach, which can be used to simulate the elastoplastic response of mechanical components subjected to variable thermomechanical loadings. The improvement as compared to standard interpolation is most substantial when the number of yield planes is limited, especially in the case of bilinear stress–strain curves. The innovation of this study is an algorithm that optimises positions of the stress–strain points across the temperature range of interest considering several input temperatures. It is shown that these methods are especially applicable for optimisation of material parameters when the stress–strain curves are available for a range of test temperatures that are needed for simulating thermomechanical fatigue. The improvement in the interpolation using these methods is exhibited for two materials with available experimental results: stainless steel EN 1.4512 and polyamide PA12.