Convergence and Solver Log Prompts: How to Interrogate AI for CFD Debugging

Reviewing CFD convergence is more than just checking if residuals dropped below a certain threshold. It requires analyzing monitor values, mass balances, and solver output to confirm the physics are stable. While Large Language Models (LLMs) can be excellent at parsing text-based solver logs or writing Python scripts to extract data, they have zero intuition for physical reality.

This guide provides structured, verifiable prompts to help you use AI for parsing logs and reviewing convergence without falling victim to hallucinations or blind acceptance.

---

1. The Risk of Blind Acceptance

When dealing with solver logs and convergence, AI models are prone to false confidence. If an LLM sees residuals decreasing, it will almost always claim "the solution has converged," even if engineering monitors (like drag or temperature) are still wildly oscillating, or if the energy residual has plateaued at an unacceptable level.

Why this is dangerous:

• LLMs do not understand mass conservation; they only understand pattern matching in text.
• They will confidently ignore warning messages buried in large log files if not explicitly instructed to find them.
• They cannot inherently distinguish between a physically valid steady-state and a numerically stagnant, incorrect solution.

Rule of Thumb: Use AI to parse and extract log data, or to write scripts that plot convergence. Never use AI to make the final determination of whether a solution is physically valid.

---

2. Parsing Solver Logs

Solver logs can be massive. AI is very good at extracting specific errors, warnings, or the final iteration states.

Bad vs. Better Prompts

❌ Bad Prompt:

"Here is my open-source CFD software log file. Is my simulation converging?"

✅ Better Prompt:

"Attached is a solver log file from a steady-state incompressible RANS simulation in open-source CFD software. Please perform the following:

1. Extract the final residual values for Ux, Uy, Uz, p, k, and epsilon from the last iteration in the log.
2. Search the entire log for the terms 'Warning', 'Error', 'Bounding', or 'Divergence'. List the exact line text and iteration number for every occurrence.
3. Do not evaluate whether the simulation has converged; only provide the extracted data."

Why this works:

• It limits the AI's role to text extraction.
• It forces the AI to look for specific numerical stability red flags (like bounding values in open-source CFD software or commercial CFD software).
• It explicitly prevents the AI from giving unverified opinions on convergence.

---

3. Extracting Residuals and Engineering Monitors

Engineering convergence requires both residuals to drop and engineering quantities of interest (monitors) to plateau. You can ask AI to help you format this data for plotting.

❌ Bad Prompt:

"Write a script to plot my CFD convergence."

✅ Better Prompt:

"I have a CSV file containing iteration history for residuals (continuity, x-velocity, y-velocity, energy) and an engineering monitor (average outlet temperature). Write a Python script using pandas and matplotlib to:

1. Plot the residuals on a logarithmic y-axis against the iteration number.
2. Plot the outlet temperature monitor on a separate graph with a linear y-axis against the iteration number.
3. Calculate the moving average of the outlet temperature over the last 100 iterations and overlay it on the second plot.
4. Print the standard deviation of the outlet temperature over the last 100 iterations to the console. Include comments and error handling for missing files."

Why this works:

• It gives specific constraints on plotting (log scale for residuals, linear for monitors).
• It asks for statistical verification (moving average, standard deviation) to help you assess whether the monitor has truly leveled off.

---

4. Comparing Residuals vs. Engineering Monitors

Sometimes you need to articulate your convergence criteria to a team or client. You can use AI to help draft this justification, provided you supply the hard data.

✅ Master Prompt: Drafting a Convergence Justification

"I need to write a brief convergence justification section for a CFD report. Here is the data I have collected:

• Continuity residual: $10^{-4}$ (plateaued)
• Momentum residuals: $10^{-5}$
• Energy residual: $10^{-7}$
• Drag coefficient monitor: Reached $0.452$ and the variation over the last 500 iterations is less than $0.1\%$.
• Mass imbalance between all inlets and outlets: $0.005\%$.

Draft a professional, one-paragraph summary stating that the simulation is considered converged based on these metrics. Emphasize that while the continuity residual plateaued at $10^{-4}$, the stability of the engineering monitor (drag) and the tight global mass balance satisfy the engineering requirements for this steady-state analysis."

Why this works:

• You provide the engineering judgment; the AI only handles the technical writing.
• It ensures the final report accurately reflects that engineering monitors supersede strict residual drop thresholds when appropriate.

---

5. Quick Reference: The "Do Not" List for Convergence Prompts

To maintain a safe AI workflow, never ask an LLM:

1. "Is this converged?" (It doesn't know).
2. "What should my convergence criteria be?" (Criteria depend entirely on your specific physics and accuracy requirements).
3. "Why did my simulation crash?" based only on a log snippet. (LLMs often guess incorrectly; instead ask "What is the exact error message in this log and what does the documentation say it means?").

---

Further Reading

• Grid Convergence Calculator
• Mesh, y+, and Near-Wall Prompts