Session: 02-05-01: Artificial Intelligence and Data-Driven Design 2

Paper Number: 193405

A Prompt-Engineered Ai-to-Cfd Workflow for Early-Stage Aerodynamic Design Validation 

Introduction
Existing CAD-to-CFD workflows assume parametric CAD geometry and do not account for unstructured or defective mesh outputs from generative AI tools. Existing efforts addressing AI-based geometry generation and CFD automation remain largely fragmented, with limited integration into a unified and validated workflow. This work introduces a prompt-engineered AI-to-CFD workflow that integrates generative geometry tools, structured geometry preparation, and high-fidelity CFD validation for early-stage aerodynamic design exploration.
Contribution of the work
This study addresses two primary research questions: (1) How can AI-generated geometry be systematically prepared and validated for CFD simulation? (2) Can generative AI tools, when guided by geometric parameter inputs, produce designs with physically meaningful aerodynamic results? The technical innovation lies in two areas: (1) A structured prompt parameter mapping approach for converting geometric inputs such as fineness ratio, chord length, camber, and frontal dimensions into repeatable and domain-agnostic generative AI geometry outputs. (2) A formalized two-pathway geometry preparation process: CadXStudio exports directly as STEP to ANSYS Fluent, while Meshy AI outputs are imported into Autodesk Fusion 360 for mesh repair and face reduction before simulation. Together, these define a replicable workflow for rapid aerodynamic exploration without full parametric CAD reconstruction.
Methodology
The pipeline works in three phases. In Phase 1, reference images along with geometric parameters from literature studies are used as explicit prompting inputs for CadXStudio software, which generates parametric solid geometry reviewed and exported as a STEP file format. The output from Meshy AI is subjected to face reduction below 50,000 faces and geometry simplification in Fusion 360 software before import into ANSYS Fluent. In Phase 2, a rectangular enclosure boundary condition with a velocity inlet and a pressure outlet is used to construct the domain in ANSYS Fluent 2025 R2. All surfaces in the domain are assigned a named selection. The turbulence model used is the K-ω SST model with a production limiter enabled to enforce steady-state RANS assumptions, as it provides greater accuracy near walls in adverse pressure-gradient flows. An unstructured tetrahedral mesh is used, with local body-surface refinement. The mesh used for the nose cone problem consisted of 823,764 cells with a minimum orthogonal quality of 1.58×10³, a maximum aspect ratio of 38.1, with convergence at residual values of 1×10³. In Phase 3, drag coefficient, pressure distribution, and velocity profiles are obtained from the results for comparison with reference
literature for similar geometries.
Preliminary results and conclusions
Results from three CFD simulations show physically consistent aerodynamic behavior. In the case of the rocket nose cone with an L/D ratio of approximately 1.67, the CFD results show a Cd value of 0.253, consistent with the published subsonic aerodynamic benchmark results for the range 0.20 < Cd < 0.30 for moderate values of the fineness ratio. In addition, the pressure coefficient contour plots show the expected stagnation point at the nose tip and the sign change in the pressure coefficient at the shoulder point, as predicted by classical external aerodynamics.
For the automotive sedan, the AI-generated geometry yielded a Cd of 0.645 under the specified simulation conditions. While benchmark simplified vehicle models under comparable flow regimes typically report significantly lower drag coefficients, the discrepancy reflects geometric simplification rather than CFD inconsistency. The simulated flow field exhibits the
expected bluff-body characteristics, including wake formation and pressure recovery, confirming that the AI-generated geometry produces physically consistent aerodynamic trends. These findings highlight the sensitivity of drag prediction to geometric fidelity and underscore the importance of structured parameter-guided prompting for benchmark-aligned design exploration.

Presenting Author: Francisco Lopez San Jose State University

Presenting Author Biography: Francisco Lopez is a master student from the department of Mechanical Engineering at San Jose State University.

Authors:

Yunjian Qiu San Jose State University
Francisco Lopez San Jose State University
Feruza Amirkulova San Jose State University