Robust Data-Driven Turbulence Modeling for RANS Closures Using a SciML Approach for Validation

Scientific Machine Learning is revolutionizing scientific domains crucial for national security by enhancing analytical capabilities. The Reynolds-averaged Navier–Stokes (RANS) equations, essential for simulating compressible fluid flows, suffer from model-form errors that limit their applicability. Addressing this, Parish et al. [AIAA 2023-2126] introduced a data-driven turbulence modeling strategy to improve RANS models. Utilizing multi-step training on eight diverse datasets, (channel flows at different Reynolds numbers, duct flow, periodic hill, and hypersonic boundary layers) the study demonstrated the model's efficacy. The research focuses on predicting Reynolds stress term discrepancies, emphasizing hyperparameter sensitivity and out-of-distribution dataset performance across various combinations of training datasets. Extensive validation efforts ensure the reliability of the machine learning models in capturing elusive model-form errors. The findings show robust improvements in wall-bounded flows, jet flows, and hypersonic boundary layers, advancing turbulence modeling comprehension. Future phases will further test dataset and hyperparameter sensitivity to ensure credibility and applicability. This study significantly enhances RANS simulations' accuracy and reliability, impacting broader fluid dynamics and machine learning fields. This work is supported by the DOE-NNSA ASC program.