Turbulence Model Prompts: How to Ask AI Better CFD Questions

If you ask an AI tool, "What turbulence model should I use for my simulation?", it will likely recite the textbook definitions of k-epsilon or k-omega SST. It will tell you that SST is "good for adverse pressure gradients." But it won't ask if you have the mesh resolution to support $y^+ < 1$, nor will it know if your solver uses automatic wall functions.

AI cannot automatically choose the correct turbulence model for you. Model selection requires engineering context, mesh constraints, solver documentation, and validation evidence.

However, AI can help compare assumptions, structure model-selection reasoning, identify missing context, and prepare validation questions. This guide is for CFD engineers, CAE engineers, students, simulation managers, researchers, and engineering developers who want to interrogate AI about turbulence safely.

[!CAUTION] Never ask AI to decide alone. Do not rely on AI for your final turbulence model choice, wall treatment choice, $y^+$ strategy, mesh adequacy, validation acceptance, or regulatory/safety-critical conclusions.

What to provide before asking AI

Before pasting a prompt, establish your context. AI will hallucinate less if you constrain its choices.

• Flow type: Internal or external? Steady or transient?
• Expected physics: Separation? Strong curvature? Swirl? Impingement?
• Reynolds number range: (If known, use our Reynolds Number Calculator).
• Relevance: Compressibility, heat transfer, or buoyancy?
• Near-wall strategy: Resolving the viscous sublayer or using wall functions?
• $y^+$ target: (If known, use our y+ Calculator).
• Mesh constraints: Hexahedral, polyhedral, or tetrahedral? Maximum cell count limit?
• Solver/software context: various CFD solvers?
• Validation/reference data: Do you have experimental data, analytical correlations, or DNS data?
• Target outputs: Are you predicting a global force (drag) or a local scalar field (heat transfer coefficient)?

1. The Generic Turbulence Prompt (And How to Fix It)

Risk Level: Low (Explanation / Comparison)

Bad prompt: "What do you know about the k-epsilon model?" Why it fails: Yields Wikipedia-level trivia without actionable engineering context.

Better prompt (Copy-ready):

I am setting up a steady-state RANS simulation for [describe flow]. I need to choose between standard k-epsilon, RNG k-epsilon, and realizable k-epsilon.
My solver is [solver name]. 
My primary quantity of interest is [e.g., global pressure drop, local heat transfer].
Please compare these three variants based on:
1. Suitability for this specific flow physics.
2. Sensitivity to inlet boundary conditions.
3. Typical convergence behavior.
List the known weaknesses of each variant for this application.

Why this works: It restricts the AI to comparing specific variants and asks for known weaknesses rather than generic praise. What to verify: Check the solver manual to ensure the described variants and their constants actually match what the software implements.

2. Comparing Major RANS/DES Families

Risk Level: Low (Comparison Table)

Better prompt (Copy-ready):

I am evaluating standard k-epsilon, k-omega SST, Spalart-Allmaras, and DES for an external aerodynamic flow over a [describe geometry].
Create a comparison table with the following columns:
- Turbulence Model
- Near-wall resolution requirement (y+ range)
- Computational cost (relative to standard k-epsilon)
- Strengths for external aerodynamics
- Weaknesses (e.g., predicting separation, free-shear flows)
Do not recommend a model. Just provide the comparison matrix.

3. External Aerodynamics

Risk Level: Medium (Model Suitability Review)

Better prompt (Copy-ready):

I am simulating external aerodynamics over a [vehicle/wing/building] at a Reynolds number of [X]. 
I care primarily about predicting the onset of separation and the total drag coefficient. 
My mesh uses prism layers targeting a y+ of [Y].
I am considering the k-omega SST model.
What are the known limitations of k-omega SST for predicting separation in this regime, and what specific solver settings (e.g., curvature correction, production limiters) should I review in the [solver name] documentation before trusting the results?

4. Internal Pressure Drop

Risk Level: Medium (Missing-Context Analysis)

Better prompt (Copy-ready):

I am simulating internal flow through a complex piping manifold with multiple bends and a sudden expansion. 
My goal is to calculate the total pressure drop. 
I am currently using the Realizable k-epsilon model with standard wall functions.
Play the role of a senior CFD analyst reviewing my setup. Ask me 3 to 5 highly specific questions about my mesh, near-wall treatment, and flow assumptions that could expose why Realizable k-epsilon might underpredict or overpredict the pressure drop in the sudden expansion.

5. Separated Flow

Risk Level: High (Choosing a Final Model - Use for Discovery Only)

Better prompt (Copy-ready):

My simulation involves massive, unsteady flow separation behind a blunt body (Reynolds number = [X]). 
Steady RANS (SST) failed to converge, showing limit-cycle oscillations in the residuals.
I am moving to a transient simulation. I am deciding between URANS (SST) and DES.
Provide a technical breakdown of the risks of using URANS for massively separated flows (the "gray area" problem). 
What specific criteria should I use to justify the increased computational cost of DES over URANS to my manager?

6. Heat Transfer

Risk Level: Medium (Suitability Review)

Better prompt (Copy-ready):

I am performing a conjugate heat transfer (CHT) simulation in [solver name] where resolving the thermal boundary layer is critical.
The flow is [natural convection / forced convection] at Prandtl number [X].
I am planning to use the [k-omega SST / k-epsilon] model.
What specific considerations must I make regarding the turbulent Prandtl number assumption ($Pr_t$) and the near-wall thermal formulation in this solver? What literature or validation benchmarks should I seek out for this regime?

7. Rotating Machinery

Risk Level: Medium (Missing-Context Analysis)

Better prompt (Copy-ready):

I am simulating a [pump / compressor / fan] using a [MRF / sliding mesh] approach.
I want to capture strong swirl and secondary flows.
I am considering the RNG k-epsilon model with swirl modification.
What are the theoretical limitations of two-equation eddy-viscosity models in regions of strong streamline curvature and system rotation? 
Should I be evaluating a Reynolds Stress Model (RSM) instead, and what convergence penalties should I expect if I do?

8. Wall Treatment and y+

Risk Level: High (y+ Strategy - Do Not Ask AI to Decide)

Better prompt (Copy-ready):

I am struggling to maintain my target y+ across a complex geometry. 
In [solver name], I am using the k-omega SST model with automatic/blended wall treatment.
My y+ values range from [Min] to [Max], with a significant portion in the buffer layer ($5 < y^+ < 30$).
Explain mathematically how blended wall functions attempt to bridge the buffer layer.
What are the documented risks of having cells in the buffer layer for predicting [skin friction / heat transfer], and how can I verify if the blended wall function is introducing unacceptable errors?

9. Validation and Sensitivity

Risk Level: High (Validation Acceptance)

Better prompt (Copy-ready):

I have run a baseline simulation using k-omega SST. 
I need to perform a turbulence model sensitivity study to quantify uncertainty before presenting the results.
What are two alternative turbulence models (one simpler, one more complex) that would provide a meaningful bound on the modeling error for [describe physics]?
Outline a rigorous procedure for comparing the results of these three models against my experimental data.

Prompts for Managers and Team Leads

Use these prompts to structure a design review or evaluate an analyst's justification.

Manager Prompt 1: The Credibility Check

My CFD analyst has presented results for [application] using the [Model Name] turbulence model. They justify this choice by saying "it is standard practice."
Provide a checklist of 5 technical questions I should ask the analyst during the design review to ensure they understand the model's limitations, the required near-wall mesh resolution, and the specific sensitivities of this model to the inlet boundary conditions.

Manager Prompt 2: Validation Strategy

We are relying on a RANS simulation (using [Model Name]) to sign off on a safety-critical design for [application].
Act as an external auditor. Draft a critique of why trusting a single RANS model for this application is risky, and propose a validation matrix that includes mesh sensitivity, model sensitivity, and required experimental anchor points.

Prompts for Researchers and Developers

Use these to dive into equations, code implementation, or literature gaps.

Developer Prompt 1: Code Implementation Comparison

I am reviewing the implementation of the k-omega SST model in [Open-Source Solver A] versus [Commercial Solver B].
Identify the common variations in the SST formulation (e.g., 1994 vs. 2003 Menter papers, production limiters, curvature corrections, Kato-Launder modification). 
What specific terms or constants in the source code or solver documentation should I check to ensure I am comparing "apples to apples" between the two solvers?

Developer Prompt 2: Literature Assumptions

I am researching the use of the [Turbulence Model] for [Highly specific physics, e.g., supercritical CO2 heat transfer].
Summarize the foundational assumptions made in the derivation of this model that might be violated by strong density gradients. 
Suggest keywords and authors I should search for in the literature regarding modifications to this model for variable-density flows.

The Master Prompt: Turbulence Selection Review

If you are finalizing your setup, use this master prompt to perform a holistic review of your turbulence strategy.

Master Prompt (Copy-ready):

I am finalizing my turbulence modeling strategy for a CFD simulation. 
I will provide you with the exact flow physics, mesh constraints, and my proposed model.
You will act as a strict CFD peer reviewer.

Here is my context:
- Solver: [Solver Name]
- Flow Regime: [e.g., Incompressible, Ma = 0.8, Transient]
- Key Physics: [e.g., massive separation, swirl, heat transfer]
- Target Output: [e.g., total drag, maximum wall temperature]
- Mesh Strategy: [e.g., Hexahedral, targeting y+ < 1]
- Proposed Model: [e.g., k-omega SST with curvature correction]

Do NOT validate my choice. Instead, perform a risk assessment:
1. Flow Physics Mismatch: What specific phenomena in my "Key Physics" is this model notoriously bad at predicting?
2. Near-Wall Risks: Given my "Mesh Strategy", what happens if my y+ assumptions are wrong?
3. Solver Specifics: What default limits or blending functions in "Solver Name" could artificially stabilize this model and hide physical instabilities?
4. Required Sensitivities: What specific parameter (e.g., turbulent length scale at the inlet) must I test for sensitivity before trusting the result?