Model Weights for FlowGen

This repository contains the trained neural network weights and normalization parameters for FlowGen, a hybrid Diffusion and Fourier Neural Operator (FNO) framework designed for modeling weakly compressible laminar flows. The provided checkpoints allow users to replicate the inference pipeline, which transitions from coarse generation via a diffusion process to physical refinement via a dedicated correction network. Repository Contents best_model_3_channels.pt: The trained weights for the denoising U-Net block. This model operates over a 500-step noise schedule (DDPMScheduler with a squaredcos_cap_v2 variance schedule) during training and is optimized for accelerated 50-step DDIMScheduler coarse inference. best_correction_unet_fno.pt: The trained weights for the Correction U-Net. This network embeds Fourier Neural Operator (FNO) blocks in parallel with standard convolutional U-Net layers to enforce structural, boundary, and fluid-dynamic coherence on the raw diffusion outputs. minmax_dict.pt: A PyTorch dictionary containing the pre-calculated global minimum and maximum values for all physical fields (e.g., velocities, pressure, boundary conditions). These parameters are strictly required to normalize raw physical inputs before feeding them into the networks, and to denormalize the final outputs back into physical space.