Acceleration of Structural Analysis Simulations using CNN-based Auto-Tuning of Solver Tolerance

With the emergence of AI, we observe a surge of interest in applying machine learning to traditional HPC workloads. An example is the use of surrogate models that approximate the output of scientific simulations at very low latency. However, such a black-box approach usually suffers from significant accuracy loss. An alternative method is to leverage the large amount of data generated at simulations’ runtime to improve the efficiency of numerical methods. However, there is still no clear solution to apply AI inside HPC simulations. Thus, we propose to incorporate AI into structural analysis simulations and develop an auto-tuning of the iterative solver tolerance used in the Newton-Raphson method. We leverage residual data to train a performance model that is aware of the time-accuracy trade-off. By controlling the tuning using AI softmax probability values, we achieve 1.58x acceleration compared to traditional simulations and maintain accuracy with 1e-02 precision.