Residual Neural Networks for Transient Surrogate Modeling using HPCMP CREATE™ Sage

Computational physics-based simulations, while pivotal in the aerospace field, can be expensive and infeasible for applications requiring fast inference rates, such as optimization or trade space exploration. This paper presents a novel approach for real-time reduced-order approximation modeling of transient vorticity fields generated by fluid flow over an oscillating cylinder. Machine learning and model-order reduction techniques are used to generate a surrogate model to approximate the time evolution of the vorticity field. The HPCMPTM CREATE Sage surrogate modeling software is utilized to enable automated creation, evaluation, and refinement of high-quality surrogate models. The Sage surrogate modeling tool uses a residual neural network (resnet) to approximate the vorticity field given the position of the cylinder and the previous vorticity field. The resnet predicts the change in the field and uses a closed-loop feedback architecture to approximate the vorticity field at the future timestep. Sage was used to train, validate and evaluate the resnet models. The surrogate approximates the vorticity field with maximum percent errors less than 5%.