ChatGPT for Chemical Engineering Calculations: What Works, What Doesn’t

Ask a room of engineers whether they use ChatGPT and most hands go up. Ask whether they trust its numbers and the hands come down. Both reactions are correct. ChatGPT has become a genuine part of the engineering workflow — but it is widely misused, because most people treat it as a calculator. It is not one. Used the right way, it is one of the most useful assistants a chemical engineer has ever had. Used the wrong way, it will hand you a confident, wrong answer with a straight face.

This is a practical guide to the difference, from an engineer who uses these tools daily.

What the Research Actually Shows

It is worth being specific instead of hand-waving. Peer-reviewed evaluations published in 2026 tested ChatGPT on undergraduate and professional chemical engineering problems. The pattern is consistent:

• Accuracy on mass transfer and reaction engineering problems sat around 72% — useful, but nowhere near reliable enough to trust unchecked.
• Roughly one answer in five contained a dimensional or unit error.
• A recurring failure mode: the model produces a correct explanation but an incorrect final number. The reasoning reads well; the arithmetic is wrong.
• It struggles most when several equations or correlations must be combined and it has to choose the right one — exactly the judgment-heavy work engineers are paid for.

The takeaway is not “don’t use it.” The takeaway is: use it for what it is good at, and never let it be the calculator.

Where ChatGPT Genuinely Helps

Play to its strengths and it earns its place in your workflow:

• Setting up problems. Describe a system in plain English and it will lay out the governing equations, define variables, and organize a balance table — fast.
• Explaining concepts. As a patient tutor for a rusty topic — azeotropes, NPSH, film theory — it is excellent.
• Writing the code that does the math. This is the key move. It is unreliable at arithmetic but very good at writing a Python function that computes the arithmetic reliably.
• Drafting documentation. Turn your results into a clean first draft of a report paragraph or procedure.

This post focuses on calculations, but ChatGPT is one tool in a larger kit. For the full workflow — data analysis, simulation, and reporting — see my guide to AI tools for chemical engineers.

The One Rule: Don’t Let It Be the Calculator

Because language models predict text rather than compute, raw arithmetic is their weak point. The fix is simple and reliable: have ChatGPT set up the problem and write the code, then let Python (or a tool like Wolfram) do the actual number-crunching. Research backs this up — pairing the model with a real computation engine sharply reduces calculation errors.

So the workflow is: ChatGPT frames it, code computes it, and you verify it.

A Worked Example: Vapor Pressure with the Antoine Equation

Say you need the vapor pressure of water at 100°C using the Antoine equation:

log₁₀(P) = A − B / (C + T)

A weak prompt is “What is the vapor pressure of water at 100°C?” — it invites a single unverifiable number, and the Antoine constants it uses may be hallucinated. A strong prompt looks like this:

“Set up the Antoine equation for water. State the constants, their units, and their valid temperature range. Then write a Python function that returns P for a given T, and show the result at 100°C so I can check it against the known value of 760 mmHg.”

That prompt does three things right: it asks for the setup, it routes the math through code, and it builds in a verification point. Here is the kind of function it should produce:

def antoine_water_mmHg(T_C):
    # Valid ~1-100 C, P in mmHg, T in C
    A, B, C = 8.07131, 1730.63, 233.426
    return 10 ** (A - B / (C + T_C))

print(round(antoine_water_mmHg(100), 1))   # -> 760.0 mmHg

The check is the whole point: water boils at 100°C under 1 atm, which is 760 mmHg. If the function returns 760, the constants are trustworthy for this range. If it returns something else, you caught a bad constant before it reached a design. Always verify AI-supplied property values against a real source — NIST, Perry’s, or a validated property package.

Prompt Habits That Improve Accuracy

• Ask for the setup, not the answer. “Set up the equations and define the variables” beats “give me the number.”
• Make it show its work. Asking for step-by-step reasoning (chain-of-thought) measurably improves correctness over demanding a bare answer.
• Specify the correlation. If a particular method applies — a specific equation of state or friction-factor correlation — name it, rather than letting the model pick.
• Route math through code. “Write a Python function” turns its weakest skill into its strongest.
• State units everywhere. Given that a fifth of answers carry unit errors, being explicit about units in the prompt is cheap insurance.

A Word on Data Security

Do not paste proprietary process data, client information, or trade secrets into a public AI tool. Depending on the service and settings, prompts may be retained or used to improve the model. Treat anything you type as potentially leaving your control, and strip identifying or confidential details before you ask.

The Failure Modes to Watch

• Confident and wrong. A clean, well-formatted answer is not a correct one. Verify anyway.
• Hallucinated constants. Property values and correlation coefficients are frequently invented. Always cross-check.
• Wrong correlation chosen. When several methods are plausible, the model often picks poorly. That choice is yours to make.
• Right method, wrong arithmetic. The reason to push computation into code every time.

For a full worked case that pairs these ideas with real engineering math, see my post on mass and heat balances in the AI era.

Bottom Line

ChatGPT is a superb engineering assistant and a poor engineering calculator. Let it frame the problem, write the code, and draft the words — then verify every number yourself. Do that, and it makes you faster without making you wrong.