Engineering AI Safety Checklist

Generative AI and Large Language Models (LLMs) are powerful tools for accelerating engineering workflows, but they do not understand physics. They are pattern-matching engines. When applied to computational fluid dynamics (CFD) and computer-aided engineering (CAE), unchecked AI outputs can lead to critical design failures.

This checklist provides a framework for knowing when to use AI and when human verification is strictly required.

---

1. What AI is Good For (Low Risk)

You can confidently use AI to accelerate these tasks, provided you perform a basic review:

• ✅ Code Scaffolding & Automation: Writing Python scripts for Paraview, PyVista, or pandas to parse logs and extract data.
• ✅ Documentation Search: Summarizing syntax for specific solvers (e.g., open-source CFD software dictionary formats) based on provided documentation.
• ✅ Checklist Generation: Creating pre-processing and post-processing checklists tailored to your specific project needs.
• ✅ Report Structure: Drafting outlines, summaries, and formatting data tables for engineering reports.
• ✅ Log Parsing: Searching large solver logs for specific warning or error keywords.
• ✅ Comparison Frameworks: Generating tables comparing the theoretical pros and cons of different turbulence models or numerical schemes.

2. What AI is Risky For (High Risk)

You should never accept AI output for these tasks without independent engineering verification, source-checking, or direct intervention:

• ❌ Selecting Constants Without a Source: AI will confidently invent coefficients for material properties or empirical models. Always demand a verifiable literature citation.
• ❌ Replacing Validation: AI cannot validate a simulation against experimental data. It can only format the data you provide.
• ❌ Choosing Turbulence Models Without Context: AI will often default to "use k-epsilon," regardless of whether you are simulating a massive atmospheric boundary layer or a tiny microfluidic device.
• ❌ Claiming Convergence from Residuals Alone: AI does not understand mass conservation or physical stability. It will see dropping residuals and falsely claim the solution is converged.
• ❌ Inventing Mesh Targets: AI will hallucinate $y^+$ values or cell counts that may be completely inappropriate for your chosen wall function or boundary layer physics.
• ❌ Interpreting Safety-Critical Results: AI cannot determine if a structural load or thermal stress is "safe."

---

3. Human-in-the-Loop Decision Checklist

Before copying and pasting AI-generated engineering advice, run through this checklist:

Step 1: Source Verification

• [ ] Did the AI provide a specific textbook, paper, or documentation link for the equation, constant, or model it recommended?
• [ ] Have I manually verified that the source actually says what the AI claims it says?

Step 2: Math and Logic Verification

• [ ] If the AI performed a calculation, did I recalculate it myself or use a verified engineering calculator? (LLMs are notoriously bad at arithmetic and unit conversions).
• [ ] Are the assumptions underlying the AI's recommendation (e.g., incompressible flow, steady-state) actually valid for my specific problem?

Step 3: Workflow Verification

• [ ] Does this AI suggestion alter the fundamental physics or boundary conditions of my simulation? (If yes, document the change and justify it with engineering principles).
• [ ] Have I run a small-scale, simplified test case to verify the AI's suggested solver settings before applying them to a production run?

Step 4: Final Sign-Off

• [ ] If this project fails due to this AI recommendation, am I prepared to take professional responsibility for the error? (If no, do not use the output).

---

Conclusion

AI should be treated as an extremely fast, enthusiastic, but dangerously overconfident junior engineer. It can save you hours of scripting and structuring, but it cannot replace your engineering judgment.

Explore the Prompt Library: Use our structured Prompt Library to learn how to frame questions that force AI to stay within safe bounds.