The influence of adversarial training on turbulence closure modeling

Investigation of the generalization capability of a generative adversarial network for large eddy simulation of turbulent premixed reacting flows

“…where the coefficients β " r0.89, 0.06, 0.00006s were previously tuned [20]. The Mean-Squared Error (MSE) is computed between the reconstructed (SR) and the ground-truth (DNS) fields, as indicated in Eq.…”

“…The L1 norm is applied to each component of the loss function to drive the network's optimization. The generator, initialized with random weights, is pretrained using only pixel loss and without the adversarial component to prevent GAN convergence failure [20]. Subsequently, the pre-trained generator is used to initialize the GAN training, and trained until no additional improvement on the reconstructed field is observed.…”

“…The GAN employed in this work, schematically shown in Fig. 1, is composed of two different CNNs, a generator, and a discriminator, already introduced in previous work [20] and designed for physical applications. Based on the PIESRGAN architecture developed by Bode et al [15], both generator and discriminator heavily rely on the use of three-dimensional Convolutional Layers (CL) with leaky rectified linear units as activation functions, in contrast to the original ESRGAN structure [11] which relies on twodimension CLs.…”

“…A mini-batch size per GPU of 8 boxes, an initial learning rate of 10 ´4 and the ADAM optimizer are chosen based on previous optimizations [20].…”

“…In this work, a SR Generative Adversarial Network (GAN) [20], is employed as closure model for the unresolved Reynolds-stress and scalar-flux tensors. The filtered momentum equation reads [2] Bρ…”