High-fidelity Computational Fluid Dynamics (CFD) solvers remain the gold standard for aerodynamic analysis but impose prohibitive computational costs — typically 4 to 24 hours per simulation — that render large-scale design space exploration and multi-objective evolutionary optimization practically infeasible on conventional hardware. This paper introduces AeroML, a production-grade, full-stack aerodynamic intelligence platform that addresses this barrier by integrating four synergistic deep learning architectures: (i) a Physics-Informed Neural Network (PINN) trained to satisfy the steady, incompressible two-dimensional Navier-Stokes equations via automatic differentiation through embedded PDE residuals; (ii) a custom multi-layer perceptron primary predictor operating in the 18-dimensional Kulfan Class-Shape Transformation (CST) latent space with symmetry fusion, delivering aerodynamic coefficient predictions (CL, CD, CM) in under 50 ms with R² > 0.99; (iii) a Multi-Head Deep Operator Network (DeepONet) trained on the AirfRANS CFD dataset to predict complete two-dimensional aerodynamic flow fields at arbitrary spatial resolution; and (iv) an AeroGraphSAGE Graph Neural Network operating over unstructured aerodynamic mesh graphs. The platform further embeds a fully vectorized implementation of the Non-Dominated Sorting Genetic Algorithm II (NSGA-II) for Pareto-optimal airfoil discovery, a gradient-based inverse design engine exploiting PyTorch autograd differentiability, a structural aeroelastic solver based on thin-walled beam theory and Green's theorem, SHAP explainability analysis, and a WebGPU Lattice-Boltzmann Method real-time fluid visualizer. The complete system is deployed as a REST API-backed Next.js web application with JWT authentication and PostgreSQL persistence. Experimental validation demonstrates aerodynamic coefficient prediction errors below 2% for CL and 5% for CD versus XFOIL reference data, NSGA-II Pareto fronts achieving 51.5% improvement in lift-to-drag ratio over NACA 2412 baseline, and end-to-end optimization campaigns completing in under 5 seconds — representing a speedup of multiple orders of magnitude over conventional CFD-based workflows.
Sajid et al. (Sun,) studied this question.