Physics-guided deep learning framework for predictive modeling of the Reynolds stress anisotropy

TL;DR

This paper introduces a physics-guided, interpretable deep learning framework for turbulence modeling that combines data and domain knowledge, improving accuracy and robustness in Reynolds stress predictions across different flow conditions.

Contribution

It presents a universal, interpretable machine learning framework with domain knowledge integration, novel input features, and constraints to enhance turbulence modeling accuracy and robustness.

Findings

Achieved invariant, realizable, unbiased, and robust turbulence model.

Demonstrated good generalization across various flow configurations.

Compared neural network architectures for module performance.

Abstract

Despite a cost-effective option in practical engineering, Reynolds-averaged Navier-Stokes simulations are facing the ever-growing demand for more accurate turbulence models. Recently, emerging machine learning techniques are making promising impact in turbulence modeling, but in their infancy for widespread industrial adoption. Towards this end, this work proposes a universal, inherently interpretable machine learning framework of turbulence modeling, which mainly consists of two parallel machine-learning-based modules to respectively infer the integrity basis and closure coefficients. At every phase of the model development, both data representing the evolution dynamics of turbulence and domain-knowledge representing prior physical considerations are properly fed and reasonably converted into modeling knowledge. Thus, the developed model is both data- and knowledge-driven. Specifically, a version with pre-constrained integrity basis is provided to demonstrate detailedly how to integrate domain-knowledge, how to design a fair and robust training strategy, and how to evaluate the data-driven model. Plain neural network and residual neural network as the building blocks in each module are compared. Emphases are made on three-fold: (i) a compact input feature parameterizing the newly-proposed turbulent timescale is introduced to release nonunique mappings between conventional input arguments and output Reynolds stress; (ii) the realizability limiter is developed to overcome under-constraint of modeled stress; and (iii) constraints of fairness and noisy-sensitivity are first included in the training procedure. In such endeavors, an invariant, realizable, unbiased and robust data-driven turbulence model is achieved, and does gain good generalization across channel flows at different Reynolds numbers and duct flows with various aspect ratios.