Pareto-Optimized Physics-Constrained Autoencoders (PO-CAE) for Energetically and Spectrally Consistent Reduced-Order Fluid Modeling

This repository provides the full software implementation of the Pareto-Optimized Physics-Constrained Convolutional Autoencoder (PO-CAE) framework for reduced-order modeling of unsteady incompressible flows. The method augments a standard convolutional autoencoder with physics-informed soft constraints enforcing consistency in: global kinetic energy, vorticity field structure, enstrophy evolution, temporal energy increments. A multi-objective hyperparameter sweep is performed over the physics-loss weights, and Pareto-front analysis is used to identify optimal trade-offs between reconstruction accuracy and physical consistency. The software includes: Baseline convolutional autoencoder training Physics-constrained training with configurable loss weights Automated lambda sweep and Pareto selection Latent-space spectral analysis tools Performance logging (training time, memory usage, inference speed) Reproducibility utilities for dataset preparation The framework is designed for structured-grid velocity data stored in HDF5 format and was validated on the two-dimensional cylinder wake at Reynolds number 100. This software accompanies the manuscript: “Pareto-optimized physics-constrained autoencoders for energetically and spectrally consistent reduced-order fluid modeling” All experiments can be reproduced using the provided scripts and configuration files. License: MIT