Surrogate model

Neural network based surrogate models offer great potential for accelerating expensive combustion simulation with acceptable loss of accuracy ...

Neural network (NN) based regression on real fluid properties

The update of real fluid properties takes about 50% (reactive case) [1] or 75% (non-reactive case) of the computation time in supercritical flow simulation. Fully connected (or multi-Layer perceptron) type NN was trained to predict the thermo-physical properties of pure nitrogen. The network was optimized to 69 parameters and with prediction error lower than 1% for most conditions.

Comparison between reference results (line) and prediction of NN (symbol).

The trained NN was coupled with an AI-empowered platform, DeepFlame, to accelerate CFD simulation. Several tests were done using nitrogen convection and supercritical nitrogen injection as examples. The NN significantly reduced the computation time by several folds.

Simulation results obtained with explicit calculation and NN prediction.

References

[1] P.C. Ma, Y. Lv, and M. Ihme, “An entropy-stable hybrid scheme for simulations of transcritical real-fluid flows,” Journal of Computational Physics 340, 330–357 (2017).
[2] W. Mayer, J. Telaar, R. Branam, G. Schneider, and J. Hussong, “Raman Measurements of Cryogenic Injection at Supercritical Pressure,” Heat and Mass Transfer 39(8), 709–719 (2003).
[3]  https://github.com/deepmodeling/deepflame-dev

Neural operator

coming soon