Adaptive physics-informed neural operator for coarse-grained non-equilibrium flows

Area of Science:

• Computational fluid dynamics
• Chemical kinetics
• Machine learning

Background:

• Simulating non-equilibrium reacting flows is computationally intensive.
• Accurate modeling of chemical kinetics is crucial for applications like hypersonic flight.

Purpose of the Study:

• To develop a machine learning (ML)-based surrogate model to enhance computational efficiency in non-equilibrium reacting flow simulations.
• To ensure the ML model adheres to fundamental physical principles.

Main Methods:

• A hierarchical and adaptive deep learning strategy combining dimensionality reduction and neural operators.
• Physics-informed neural operator blocks with soft and hard constraints.
• Transfer learning for simplified training and adaptive prediction for accelerated evaluations.

Main Results:

• Accurate prediction of chemical kinetics for nearly thirty species in 0-D scenarios with a maximum relative error of 4.5%.
• Achieved 1-4.5% accuracy in 1-D shock simulations, with a tenfold speedup over conventional methods.
• Demonstrated adaptive prediction capabilities based on local non-equilibrium conditions.

Conclusions:

• The proposed ML framework provides a foundation for efficient, physics-compliant surrogates for reactive Navier-Stokes solvers.
• Enables accurate characterization of non-equilibrium phenomena in complex multi-dimensional simulations.
• Offers significant speedup for simulating hypersonic flow chemical kinetics.