Neural-Network and Reduced-order Modeling Workflows for AI-Driven CFD: Fast Response Surfaces, Reduced Dynamics and Jet in Cross-flow Examples

Abstract

Highly resolved computational fluid dynamics (CFD) simulations are essential for design but too expensive for dense design-space sampling. This chapter presents an AI-driven CFD workflow that combines scalar-response modeling and reduced-order dynamics using jet-in-cross-flow examples. A reacting hydrogen jet-in-cross-flow study is first used to train a multilayer perceptron (MLP) mapping injector spacing to unburnt hydrogen throughput, wall heat transfer, and bulk temperature concentration, with shape-preserving interpolation as a baseline. The CFD samples show a non-monotonic spacing response, and the MLP identifies an intermediate-to-wide favorable region near eight jet diameters. Leave-one-sample-out validation shows strong dependence on the predicted quantity: the bulk temperature concentration is robust, while heat transfer and unburnt hydrogen throughput are substantially harder to predict. Sparse Identification of Nonlinear Dynamics (SINDy) is then used as a reduced-order modeling framework for field-derived Reynolds-stress statistics. The parametric SINDy model provides compact field-level predictions at simulated and out-of-sample spacings, though its aggregate spacing-mean Reynolds-stress error is $6.7\%$ higher than the POD-basis reconstruction because of degradation at selected cases. The broader conclusion is that AI-driven CFD is not a single-model prescription: MLPs are effective for fast scalar responses, while POD--SINDy is better suited when transient reduced dynamics and field-derived statistics are central to the question.