Subject workflows · Updated 14 September 2026

How to Study Engineering Thermodynamics With AI (Safely)

The tool is good at the numbers now. That is exactly what makes it a dangerous study partner in this course — because the numbers were never the part you were being graded on.

Ask a current chatbot for the enthalpy of steam at 5 MPa and 400 °C and it will almost certainly be right — the strongest models tested in a 2026 benchmark scored between about 96% and 98% on property lookups. But it is reciting a table it has seen thousands of times, not computing a property, and where the table has no printed row the same confident tone delivers an answer that can be 27% wrong. Use it for drilling and checking, get property values from a real calculator, and keep what it is genuinely weak at — deciding what your system is and which relation applies — on your own side of the desk.

The course grades state determination, not arithmetic

Before any equation is useful you have to know where you are on the phase diagram, and that step is the one most marks are lost on. Claire Yu Yan's free open textbook Introduction to Engineering Thermodynamics states the requirement plainly: "To determine the fluid properties at a certain state, we will need to know two intensive properties from this list: P, T, v, u, h, s, and x."

Two properties — but they have to be independent ones, and inside the saturation dome temperature and pressure are not. That is the trap the whole first half of the course is built around: between the saturated liquid and vapour lines, the pair that fixes your state has to include quality, x = (v − v_f) / (v_g − v_f).

So the graded sequence is: identify the region, pick the right table, then read or interpolate. The reading is clerical. The identifying is the engineering, and it is the step a chatbot deletes when you paste a problem and ask for the answer.

What was actually measured

In March 2026 a benchmark called ThermoQA tested exactly this: 293 open-ended engineering thermodynamics problems in three tiers — 110 property lookups, 101 component analyses and 82 full cycle analyses — with answers generated programmatically from the CoolProp property library and cross-checked to better than 0.01% against NIST reference data. Six models were run three times each, inside a ±2% tolerance.

The headline is the gap between the tiers. The best models sat around 96–98% on the lookups, and every model did worse as the problems got longer, by anything from under 3 percentage points to more than 30. The author's summary is the sentence to take away: "property memorization does not imply thermodynamic reasoning."

Two caveats before leaning on any of this. It is a single-author preprint, not peer-reviewed work, and the models tested are the generation available in early 2026. The author also discloses that the benchmark itself was built with help from Claude, which is why nothing here quotes the leaderboard order — the failure patterns are the useful part, and those showed up across every model tested.

Where the memorised table runs out

The most striking result is the one nobody would predict from ordinary use. On ten questions about supercritical water — above 373.95 °C and 22.064 MPa, where the liquid and vapour distinction disappears — scores fell to between 45.0% and 89.5%, the widest spread anywhere in the benchmark. The author's explanation of why is worth reading twice: models "have memorized discrete steam table entries at textbook pressure intervals (typically 1, 5, 10, 15, 20 MPa) but have not internalized continuous equations of state."

The worked example is sobering. Asked for the enthalpy of water at 402 °C and 25.3 MPa, one model interpolated between values it remembered and returned 1,887 kJ/kg. The correct figure is 2,585.8 kJ/kg — an error of 27%, which the paper calls "not a rounding issue but a fundamental failure of the interpolation strategy."

Read that as a rule about when to distrust it: strongest on the round numbers that fill worked examples, weakest between them — and your homework, which names 402 °C precisely so that you have to interpolate, lives in the gaps.

Three mistakes it makes that look just like yours

The benchmark catalogues three error patterns in the multi-step problems and calls them "systematic reasoning failures that persist across runs" rather than random slips. All three are also classic student errors, and all three are invisible in a tidy-looking solution.

It flips the compressor efficiency. Models were found applying isentropic efficiency to a compressor as work = (h_2s − h₁) × η, when the correct form divides instead. The diagnosis is that they had learned a generic "efficiency times ideal work" template from turbine problems "without understanding the physical asymmetry between work-producing and work-consuming devices." A turbine gives you less than the ideal; a compressor costs you more.

It reaches for constant specific heats when told not to. Models applied the constant-c_p isentropic relation "even when the problem explicitly specifies variable specific heats and instructs use of air tables" — pattern-matching to the case that appears more often in training data instead of selecting the method the question asked for.

It contradicts its own numbers. In cycle problems, models were caught "reporting a turbine work output that does not match h₁ − h₂ given their own stated enthalpy values." That one is a gift, because it means you can catch a bad solution using nothing but the solution itself and a calculator.

The failure you are least likely to notice

One finding changes how you should read any AI-written solution. The dominant error mode was not a wrong number but a missing one: models "that attempt a calculation step almost always produce an answer within 2% tolerance, but frequently omit steps entirely." The chain breaks when the model cannot pin down an intermediate state, and it quietly moves on.

A wrong value is something you might spot; an absent step is not — and if you are reading the solution to learn the method, that step is precisely the piece you needed. You can only see what is missing if you already know what should be there, which is the argument for working the problem first and opening the tool second.

Refrigerants, and the 200 kJ/kg that is not an error

If your course reaches vapour-compression refrigeration, expect the tool to get noticeably worse. On the component-analysis tier, models scored 75–98% on water and collapsed to 44–63% on R-134a. The reason is simply how much data exists: steam tables are in every textbook and every solved example online, refrigerant tables are not. So the most widely deployed cycle in the world — every fridge, air conditioner and heat pump — is the one the models handle worst.

There is a second problem stacked on top, and it will confuse you even when nothing is wrong. Enthalpy has no absolute zero, so every table picks an arbitrary baseline, and the conventions disagree. The CoolProp documentation lists them: IIR sets "h = 200 kJ/kg, s = 1 kJ/kg/K at 0C saturated liquid", ASHRAE sets "h = 0, s = 0 @ -40C saturated liquid", and NBP sets both to zero at the normal boiling point. Two tables can therefore give enthalpies about 200 kJ/kg apart for the same state, and both be right.

Differences cancel the baseline out, so this never bites you — as long as every number in one calculation comes from one source. It bites the moment you take h₁ from your textbook and h₂ from a chatbot. ThermoQA found models "trained predominantly on ASHRAE or NBP conventions produce systematically offset property values even when the IIR convention is stated explicitly in the problem," so asking nicely does not fix it.

Split the job

The workflow, step by step

Four prompts. Paste them into ChatGPT, Claude, Gemini or whatever you already use. Each one keeps the judgement with you and pushes the recall onto something that can be checked.

1. Fix the state before anything else. No numbers yet — this is the step that decides whether the rest is even possible.

You are my thermodynamics state coach. Here is a problem:

[paste the problem]

Do NOT give me any property values, do NOT solve anything, and do
NOT tell me which table to open.

Ask me one question at a time, waiting for my answer each time:
1. What is the substance, and what two properties am I actually
   given at this state?
2. Are those two properties independent of each other here? If the
   substance might be saturated, how would I check?
3. Which region is the state in - compressed liquid, saturated
   mixture, superheated vapour, or supercritical?
4. What evidence in the problem settles that, and what would I
   compare it against to be sure?

If I get one wrong, say only that it is wrong and ask again a
different way. Do not correct me.

2. Make it show the lookup instead of remembering it. This is the prompt that turns a recall problem into a checkable one.

I need properties for [substance] at [state].

Rules:
- If you can run code, use CoolProp or an IAPWS formulation and
  show me the code and its raw output.
- If you cannot run code in this chat, say so plainly and stop.
  Do not give me remembered values.
- If you do give any value, state whether it sits exactly on a
  standard table row or whether it required interpolation, and
  between which two rows.
- For any refrigerant, state which reference convention the value
  uses: IIR, ASHRAE or NBP.

Then stop. Do not solve the problem.

The last two rules do most of the work: interpolation near the critical point is where that 27% error came from, and a refrigerant enthalpy on an unstated baseline is not comparable to your textbook's.

3. Audit a worked solution against itself. Use this on any solution you did not write, including your own.

Here is a worked solution:

[paste it]

Do not rewrite it and do not tell me whether the final answer is
right. Do this instead:

1. List every state in a table: state number, region, and each
   property value the solution asserts.
2. Using ONLY those listed values, recompute each energy balance
   and each efficiency, and show your arithmetic. Tell me where a
   recomputed number disagrees with the number the solution gave.
3. List every step the solution assumed but never showed.
4. For each efficiency used, say whether the device produces or
   consumes work, and whether the ideal value was multiplied or
   divided.

Then stop.

Steps 2, 3 and 4 target the three documented failure patterns in turn: the inconsistent work output, the omitted step, and the compressor efficiency applied the turbine way round.

4. Drill the assumptions, which is what the exam is really testing.

Quiz me on modelling assumptions, one question at a time.

The system is: [describe the device or cycle]

Ask me these one at a time, waiting for my answer each time:
1. Where exactly am I drawing the system boundary, and is it a
   closed system or a control volume?
2. Which terms of the energy balance can I drop here, and what
   physical claim am I making by dropping each one?
3. Is this process steady-state? What in the problem tells me?
4. Am I treating specific heats as constant? What does the problem
   say, and what changes if I am wrong?

Do not answer for me. After each answer, ask one follow-up
beginning "what would change in your result if that were false?"

Where the line is

The split above sits comfortably inside most academic integrity policies, and the FAQ below covers the ordinary cases. The risk specific to thermodynamics is one the accuracy creates rather than removes: a correct-looking solution you did not form is unusually easy to submit and unusually hard to reproduce in a closed-book exam. Read your syllabus, and if the wording is vague, our guide to homework help without cheating and the class AI policy checklist cover how to read it and how to ask in writing.

Related reading

• The courses either side. Heat transfer, which this one is the prerequisite for, and fluid mechanics.
• The chemistry version. Physical chemistry and Carnot cycles and Gibbs free energy — the same laws, a different set of questions.
• The exam. The FE exam workflow, where the property tables are handed to you and speed is the whole skill, plus STEM problem sets.
• The maths underneath. Differential equations and multivariable calculus.
• How the same tool fails elsewhere. Immunology, where it cannot hold rare facts at all — the opposite problem to this one — and the hallucination checklist.

FAQ

Can I trust ChatGPT to read a steam table for me?

On a value printed in your book, usually yes. In the 2026 ThermoQA benchmark the strongest models scored between about 96% and 98% on property lookups. The catch is that they are reciting the table rather than computing the property, so accuracy falls away where the table has no row. The benchmark's supercritical questions produced its widest spread of all, from 45.0% to 89.5%, because models "have memorized discrete steam table entries at textbook pressure intervals ... but have not internalized continuous equations of state." Look the value up yourself at the free NIST Chemistry WebBook, which takes about ten seconds.

Why do AI answers go wrong on refrigerant problems but not on steam?

Two reasons, and the benchmark separates them. The first is how much data exists: on component analysis the models scored 75–98% on water but collapsed to 44–63% on R-134a, because steam tables appear everywhere in textbooks and refrigerant data does not. The second is the reference state. Enthalpy has no absolute zero, so every refrigerant table picks a baseline, and the conventions disagree — CoolProp documents IIR as h = 200 kJ/kg at 0 degrees C saturated liquid and ASHRAE as h = 0 at minus 40 degrees C. A model trained mostly on one convention returns values offset from your table by a constant.

What is the fastest way to check an AI solution to a thermodynamics problem?

Check its numbers against its own numbers. The ThermoQA authors found models "reporting a turbine work output that does not match h1 − h2 given their own stated enthalpy values," which they describe as a failure to self-verify against conservation laws. So take the enthalpies the model listed, redo the energy balance on a calculator, and see whether the answer it gave comes back out. Then count the steps: the dominant error mode was a missing step rather than a wrong value, so look for what is absent, not only for what looks odd.

Should I let the AI use a property calculator or code?

Yes for the numbers, and do not expect it to fix the thinking. The ThermoQA authors report that enabling the CoolProp property library through function calling "raised Tier 1 scores to near 100% for all models while leaving Tier 3 scores largely unchanged," which is a clean demonstration that lookup errors and reasoning errors are different problems. A tool removes the recall guesswork. Choosing the system boundary, the assumptions and the right form of the energy balance is still yours.

Is it cheating to use AI for thermodynamics homework?

That depends on your course policy, and thermodynamics has one risk worth naming separately. Because the tool usually gets the arithmetic right, a solution you did not form can look completely convincing to you, so it is unusually easy to submit work you cannot defend in an oral check or reproduce in a closed-book exam. Being quizzed on which region a state sits in, having your assumptions challenged, or asking why one relation applies and another does not is ordinary studying. Read your syllabus, and ask your instructor in writing if the wording is unclear.

Bottom line

Thermodynamics is the course where the usual warning about hallucinated numbers has stopped being the point. The tool will hand you an accurate enthalpy and a clean-looking solution, and still cost you the marks — by skipping the step you needed, by reaching for the relation that appears most often rather than the one the question specified, or by interpolating confidently in the one region where interpolation does not work. Fix the state yourself, get the properties from NIST or CoolProp, and make the chatbot prove its answer against its own numbers before you believe a line of it.