Gains:
- Ability to use artificial intelligence as a preliminary prediction and prediction method (group contribution) guide for thermophysical properties and request any value.
- Ability to capture hallucinations by verifying each critical feature with a reliable database (DIPPR/NIST) or experiment
- Ability to understand that material selection is a safety-critical decision that requires corrosion data and authorized approval.
Every process calculation ultimately comes down to a number: a fluid's density, viscosity, heat capacity, vapor pressure, boiling point; a resolution; resistance of a material to corrosion. These are collectively called thermophysical properties - physical quantities that determine the heat, flow and phase behavior of a substance. A distillation column, a heat exchanger or a pump is designed correctly only if these features are correct. In this unit, you will learn how to use artificial intelligence as a preliminary prediction and comparison tool in property prediction and material selection (selection of metals, gaskets, coatings to withstand process conditions); but we will learn why these values need to be verified with a reliable database or experiment.
The critical truth is this: feature estimation is the numerical basis of design; An error here spreads to the entire account. AI may confidently give a vapor pressure or a viscosity, but that value may be for another compound, may be at the wrong temperature, or may be entirely hallucinatory. So the golden rule of feature data: every critical feature should be based on a reliable source (DIPPR, NIST, manufacturer data sheet) or experiment.
Step-by-step AI support for feature prediction
1. What feature do you need, under what condition? Temperature, pressure, phase and concentration change the property. The value taken without giving the condition to the AI is misleading.
2. Choosing the estimation method. If there is no experimental data, AI recalls prediction approaches such as group contribution methods — methods that predict properties by breaking the molecule into its parts and adding the contributions of each part; e.g. Joback, UNIFAC — and explains which one is appropriate.
3. Preliminary value and uncertainty. AI can give a preliminary value, but definitely within the framework of "this is a prediction, its uncertainty is high, verify it". Also ask the AI about the typical margin of error of the value.
4. Comparison with the source. Compare the value of AI to a thermodynamics database or manufacturer sheet. If the deviation is large, the value is not used.
5. Material compatibility preliminary screening. YZ recalls known corrosion/incompatibility risks for a fluid-material pair (e.g. pitting corrosion of stainless steel in a chloride-containing environment). This is a forewarning; The final selection is made by compatibility charts and expert evaluation.
6. Reasonableness filter. Does the value make sense in terms of rank? For water, is the density around ~1000 kg/m³? Simple order checking catches most errors.
Tip: When you ask the AI for a feature value, say “specify the source and estimation method, tell me whether it is experimental or estimated.” If AI cannot cite a source, mark that value as a guess and verify it independently.
Material selection: safety and durability
Material selection is both an economic and safety-critical decision. The wrong material can weaken with corrosion, causing a pressure vessel to burst or a toxic fluid to leak. The AI can recall a starting materials list and known incompatibilities; But the combined effects of temperature, pressure, concentration and impurities are complex and the final selection is made by materials engineering data (compatibility charts, corrosion rate data) and testing where necessary.
Attention: AI saying "this material is suitable for this fluid" is not a guarantee. In the case of pressurized equipment and toxic/flammable fluids, material selection requires qualified engineer approval and verified corrosion data.
three mini cases
Case 1 — Quick prediction. An engineer wanted a preliminary estimate of the heat capacity for a new solvent mixture for which data were scarce in the literature. The AI gave a range via a group contribution method. The engineer made a rough energy calculation with this preliminary value, then had it measured in the laboratory; measurement was within 6% of prediction. Forecasting accelerated planning, but design shifted to experimental value.
Case 2 — Hallucination caught. YZ gave the normal boiling point of a compound as 78°C. When the engineer compared it with NIST, he found that the actual value was 118°C; The AI had confused the name with another similar compound. Comparison with a source prevented a 40°C error from entering the column design.
Case 3 — Material warning. YZ flagged the choice of 304 stainless steel for a chloride ion-containing aqueous stream as "risk of pitting and stress corrosion cracking" and cited a more resistant alloy. The materials engineer confirmed with compatibility data and selected a higher alloy material. AI brought the risk to the agenda, the decision was made with data.
Four copyable templates
1) Feature prediction (source required):
Your role: physique expert. Compound: [name/formula], condition: [T, P, phase]. Feature I want: [e.g. vapor pressure]. Give: (1) experimental value and SOURCE (DIPPR/NIST/manufacturer), if available, (2) if not, estimation method and estimated value and margin of uncertainty. Clearly state whether it is experimental or predictive. If you cannot cite a source, mark it as "prediction that needs verification". Do not fabricate the value; If you're not sure, tell me.
2) Group contribution method guide:
I have a compound for which there is no experimental data: [structure/SMILES]. Which group contribution method (Joback, UNIFAC, etc.) is suitable for [property], and why? Explain the steps of the method and the required group parameters. State that the result is an estimate and the typical margin of error. I will verify the final value with experiment/database.
3) Value comparison/verification:
Below are values from different sources for a feature. Compare them in a table, calculate the deviation, and rank which one is more reliable (empirical > evaluated > predicted). If there is a large deviation, mark the possible cause (different condition, different isomer). Sources: [paste]
4) Material compatibility preliminary screening:
Fluid: [composition, concentration, impurities], condition: [T, P].Candidate materials: [list]. Recall known corrosion/incompatibility risks (pitting, stress cracking, general corrosion) for each. This is a PRELIMINARY scan; Final selection requires corrosion data and authorized engineer approval. Clearly mark risky pairs.
Weak prompt / Strong prompt
Weak prompt:
What is the viscosity of toluene?
No heat, no welding. AI gives a number, but under what conditions and how reliable it is is unclear.
Powerful prompt:
Your role: physique expert. Dynamic viscosity for toluene in liquid phase, 25°C, 1 atm. Give: value, unit, SOURCE, and experimental estimate. If possible, give two independent sources and state consistency. If you cannot cite the source, mark it as a guess. Value fitting. I will verify the result with a thermodynamic database.
The difference is clear: the condition, the resource mandate, and the validation framework make the value available.
Role distribution in feature and material decision
business
Role of AI
Human decision/source
Experimental feature
source reminder
Database confirmation
No data, guess
Method + default
Experiment/database validation
Value comparison
Table, deviation
reliability judgment
Material candidates
risk reminder
Compliance data, approval
Rank control
marking
physical reasonableness
Common mistakes
- Using unsourced value. Every critical feature given by the AI is confirmed by the database/experiment.
- Not stating the condition. The value taken without specifying temperature/pressure/phase may belong to the wrong condition.
- Mistaking the estimated value as empirical. Group contribution estimates are uncertain; Take this into consideration in the design.
- Making material selection with a one-sentence confirmation. Corrosion is multivariate; data and confirmation required.
- Bypassing the rank check. A simple “does this number make sense?” control catches most hallucinations.
In summary
AI in property and material prediction; It is a powerful assistant that generates preliminary values, reminds you of the estimation method, compares resource values and flags material risks. But every critical feature must be based on a reliable database or experiment; Material selection is a safety-critical decision that requires corrosion data and authorized engineer approval. AI provides predictions and warnings; The engineer connects to the source, verifies, and selects.
Application task
Select a compound and a property (e.g. vapor pressure of a solvent, 40°C). Get the value and its source from the AI with the "Attribute prediction (resource required)" template; then compare it to a thermodynamic database and note the deviation. Then use the “preliminary material compatibility screening” template for a fluid-material pair and confirm the resulting risks with a compatibility chart.
checklist
- I wanted the [ ] feature along with the condition (T, P, phase).
- [ ] I compared the value given by the AI with a reliable database/experiment.
- [ ] I distinguished estimated values from experimental and noted the uncertainty.
- [ ] I verified the material selection with corrosion data and authorized approval.
- [ ] I applied a rank/plausibility check to each critical number.