Deep Learning for Inverse-Problem-Based Calibration of the γ Transition Model

This study introduces a novel framework that leverages deep neural networks to optimize the coefficients of a Reynolds-averaged Navier–Stokes (RANS) transition model, achieving precise alignment with experimental data. The approach is validated using the well-established one-equation intermittency [Formula: see text]-transition model and the comprehensive ERCOFTAC T3 flat plate experiment series. By systematically varying the coefficients of the transition model, an extensive database is constructed to train a fully connected neural network capable of predicting skin friction coefficient distributions as a function of the transition model coefficients. Once trained, the neural network is employed to solve the inverse problem, identifying tailored sets of coefficients for each individual experiment, as well as a global set optimized for all experiments considered together. To ensure robust solutions, the study also provides essential techniques for regularizing the inverse problem, enabling the derivation of solutions that effectively capture the transition region. The results demonstrate the framework’s potential to enhance predictive accuracy and automate the calibration process for models struggling with generalization due to a limited understanding of transition physics. This automated methodology not only reduces the reliance on manual parameter tuning but also establishes a foundation for streamlining turbulence model calibration, marking a significant advancement in the field.