GPU-Accelerated Solution of Many Small ODE Systems for High-Temperature Component Lifetime Modeling

Accurate lifetime prediction of critical components in steam and gas turbines is essential for optimizing maintenance strategies and ensuring safe operation under demanding conditions. Traditional engineering methods, though quick, lack the flexibility and precision required for complex scenarios. Advanced continuum damage mechanics models, such as the Chaboche model, offer greater accuracy but are computationally intensive. This thesis investigates the potential of GPU acceleration to significantly reduce the computational cost of lifetime estimation using the Chaboche model. By leveraging the parallel processing capabilities of GPUs, the study enables efficient simulation of numerous load cases and parameter configurations, addressing uncertainties inherent in such models. We evaluate the suitability of SUNDIALS, a state-of-the-art ODE solver, for GPU-accelerated simulations, highlighting its limitations. Simpler numerical methods, including the Forward Euler method, are also explored to balance accuracy and computational efficiency. Performance enhancements are achieved by optimizing data types and step sizes, improving the efficiency of both SUNDIALS and the Forward Euler method on GPU platforms. The results demonstrate substantial computational speedups of 40 to 60 times with GPU-accelerated solvers compared to CPU-based implementations. Additionally, optimizing the simplified Chaboche model with tailored data types and step sizes yields a further 23% performance improvement without compromising accuracy. This research offers key insights into the effective application of GPU acceleration for lifetime prediction, contributing to more efficient and reliable design and maintenance practices for critical components in energy systems.