Researchers present a physics-constrained machine learning framework that replaces direct evaluation of detailed chemical source terms with a surrogate model to accelerate the direct numerical simulation (DNS) of turbulent reacting flows. The approach incorporates the second law of thermodynamics as a training constraint by enforcing non-negative entropy generation, which restricts the thermochemical state evolution to physically admissible directions and improves stability during time integration.
- Demonstrated on DNS of a two-dimensional planar lean premixed methane-air flame interacting with a turbulent flow field.
- Reproduces detailed-chemistry results with high fidelity while achieving more than an order-of-magnitude reduction in computational cost.
- Utilizes a residual-based synthetic data augmentation strategy to construct new training data from the original dataset.
- Enables accurate simulation at new inlet conditions without requiring additional detailed-chemistry CFD runs.
The authors consider this significant because thermodynamically constrained machine learning provides reliable and computationally efficient surrogates for detailed chemistry in high-fidelity combustion simulations.