Parameter identification for turbulence models in shock wave boundary layer interference using an improved Bayesian neural network

Abstract

Driven by the limitations of current turbulence models in predicting Shock Wave Boundary Layer Interference (SWBLI) flows with high accuracy, this study introduces an enhanced Bayesian Neural Network (RBNN)-assisted optimization framework designed to rapidly and accurately identify SST turbulence model parameters. To explore various SWBLI conditions under different shock intensities, a compressed corner flow configuration was selected for investigation. The parameter identification capabilities of Artificial Neural Networks (ANN), Residual Networks (ResNet), and traditional Bayesian Neural Networks (BNN) were systematically compared and analyzed. The results demonstrated that, across different training sample sizes, the RBNN consistently delivered superior identification accuracy, achieving a coefficient of determination (R²) value of over 0.997 for the test set and maintaining a root mean square error (RMSE) for wall pressure predictions below 0.013. When applied to conditions of Ma = 2.85 and a 24° compression corner, the RBNN-optimized turbulence model parameters significantly enhanced prediction accuracy, achieving a 65.5 % reduction in RMSE for wall pressure (from 0.6029 to 0.2080) compared to standard parameter calculations, and realizing a 40 % decrease in wall friction coefficient error (from 0.0010 to 0.0006), with results showing the highest agreement with experimental data. Furthermore, the model parameters calibrated for the 24° compression corner condition demonstrated excellent transferability, successfully extending to other ramp angles (20°, 16°, and 8°) under the same Mach number. This research provides an efficient and reliable intelligent algorithm framework for optimizing turbulence models under complex flow conditions, highlighting its broader applicability and robust performance in addressing challenging fluid dynamics problems.