Datasets for training of FlowGen

This repository contains the complete training dataset and intermediate outputs used to train and validate FlowGen, a hybrid Diffusion and Fourier Neural Operator (FNO) framework designed for modeling weakly compressible laminar flows. These files consist of ground-truth numerical solutions generated via finite element simulations and the intermediate uncorrected generative data used for down-stream correction training. (Note: Trained model weights and checkpoints for the networks are hosted separately). Dataset Contents 1. Baseline Simulation Data (Diffusion Training) boundary_conditions.npy: The input boundary condition tensors generated via numerical FEniCS simulations. Each sample contains 4 channels representing physical constraints: inlet velocities, an inlet mask, an outlet mask, and a domain wall mask. output.npy: The steady-state fluid flow "ground truth" fields (velocities and pressure) obtained directly from the FEniCS numerical solver. Paired with boundary_conditions.npy, this constitutes the baseline dataset used to train the primary Diffusion U-Net. 2. Generative Realization Data (Correction Training) results_bc.npz: The dataset generated to train the downstream Correction U-Net. It contains the raw, uncorrected predictions produced by the fully trained Diffusion U-Net alongside their corresponding boundary conditions. This allows the correction network to explicitly map and learn how to fix generative physical discrepancies, boundary mismatches, and structural artifacts.