59295 - High-Performance Computing Probabilistic Fracture Mechanics Implementation for Gas Turbine Rotor Disks on Distributed Architectures Including Graphics Processing Units (Gpus) 

We present the implementation of a Monte Carlo based probabilistic fracture mechanics simulation method ---for heavy duty gas turbine rotor disk utilized in the energy sector--- on high performance distributed heterogeneous computing architectures equipped with General Purpose Graphics Processing Units (GPUs). In these simulations, millions of individual fracture mechanics simulations have to be performed. In order to speed-up these time-consuming computations, we utilized the message passing interface (MPI) in order to distribute the computational tasks evenly to available CPU cores on a distributed or shared memory system. Furthermore, the most computationally intensive tasks, i.e. the numerical integration of crack sizes with cycles, have been efficiently deployed to GPU cores by utilizing the CUDA paradigm provided by NVIDIA.

This enables us to utilize combined compute resources of any number of CPUs and GPUs over the network. We detail our experiences in porting this algorithm to GPUs such as numerical correctness of our results and running and seeding an efficient parallel random number generator. We also microbenchmark the key computation of Runge-Kutta integration step and show over two orders of magnitude speedup on a typical GPU compared to a single threaded CPU version. On strong and weak scaling computational experiments we show that our GPU implementation is as scalable as the CPU implementation. We evaluate the performance of our complete implementation on seven different GPUs spanning four different architectures and achieve speedups ranging from 10x to 48x compared to a single threaded CPU version.

We will discuss details of the implementation strategy, and showcase application examples of probabilistic fracture mechanics of heavy duty steel rotor disks for the energy sector[1,2].

The demonstrated scalability of the approach on architectures including several hundred of CPUs and GPUs for Monte Carlo based probabilistic applications can be adapted to other probabilistic lifing applications relevant for the energy sector. This paves the way for a fast and reliable holistic risk quantification of power plant components. These probabilistic digital twins allowing the support of customer needs for more flexible individualized operational profiles due to the emergence of renewable energy sources. 

1)     Kadau, Kai, Gravett, Phillip W. and Amann, Christian. “Probabilistic Fracture Mechanics for Heavy-Duty Gas Turbine Rotor Forgings.” Journal of Engineering for Gas Turbines and Power Vol. 140, No. 6 (2018): p. 062503. DOI 10.1115/1.4038524.

2)     Radaelli, Francesco, Kadau, Kai, Amann, Christian and Gumbsch, Peter. “Probabilistic Fracture Mechanics Framework Including Crack Nucleation of Rotor Forging Flaws.” Proceedings of ASME Turbo Expo 2019: pp. GT2019–90418. 2019. DOI 10.1115/GT2019-90418.