Automated Discovery of Physically Interpretable Subgrid Length Scales for Detached Eddy Simulation via Agentic AI

Abstract

Traditional turbulence-model development relies heavily on expert knowledge and is time-consuming. This study develops an agentic AI framework, DEScover (automated Discovery of DES closures), to explore subgrid length scales for detached eddy simulation (DES). The user provides only a modeling idea: the length scale is the grid width normal to the local vortex axis. Four large language model (LLM) agents then propose, review, implement, and evaluate candidate formulations and generate scripts for CFD simulations. The Discovery Agent infers the vortex axis from various vortex-identification methods. The candidates are tested on four cases spanning canonical and engineering flows: a flat plate, a backward-facing step, a free shear layer, and an iced airfoil. For all test cases, the best candidate consistently outperforms the standard length scale based on the maximum cell size. Its accuracy is similar to, or slightly better than, that of the expert-designed shear-layer-adapted length scale. Each agentic discovery run takes about 25 minutes on average, excluding the CFD simulations. These results show that LLM-based reasoning can significantly accelerate the discovery of physically interpretable turbulence models and reduce reliance on manual expert formulation.