Validation — predicted vs experiment

At a glance

Mean absolute error (MAE) of MoleBench against experiment, across the benchmark sets below:

*excluding one gas-phase carboxylic-acid outlier discussed below. All calculations run on this site; you can reproduce any of them in the Studio.

Geometry — GFN2-xTB bond lengths

Getting the shape right is the foundation of everything else, and here MoleBench's default engine is genuinely strong: bond lengths land within about a hundredth of an ångström of experiment.

Dipole moments — GFN2-xTB

The dipole tests the electronic structure, not just the shape. Trends are captured perfectly (the non-polar molecules come out at zero; the most polar comes out most polar), with a mild systematic over-prediction of a few tenths of a Debye on the carbonyls.

¹³C NMR — GIAO B3LYP/6-31G* (Advanced tier)

The quantum NMR is calibrated against experiment and performs well across a 200-ppm range, from a shielded methyl to a deshielded carbonyl.

¹H NMR — GIAO B3LYP/6-31G*

O–H / N–H protons are omitted: they are dominated by hydrogen bonding and concentration, so a gas-phase value is not comparable to a solution measurement.

pKa — GFN2 deprotonation + per-class calibration

This one has a story. An earlier single global calibration carried a +1.5–2 unit high bias on carboxylic acids — which this very benchmark exposed. The cause: the GFN2 deprotonation energy maps to pKa with a class-dependent slope (carboxylic acids, phenols and alcohols each follow a different line), so no single line can fit them all. We re-fit per functional-group class against a 20-acid set spanning pKa 0–17. The bias is gone, and the error dropped to ~0.3 units:

Held out from the calibration set, then predicted blind: 2-naphthol 9.8 (exp 9.51) and propanoic acid 4.7 (exp 4.87) — so it generalizes, it isn't memorizing. Acidity ranking across 17 units is reliable, and absolute values are now good to roughly ±0.5 unit for the common classes.

Two honest edges, now flagged in the tool itself. Phosphorus/sulfur oxyacids (phosphoric, phosphonic, sulfonic) were originally mis-scored as alcohols (phosphoric came out ~8 instead of ~2); they now use dedicated P- and S-oxyacid classes and carry an "approximate, strong acid" note. Amino acids are detected and labelled: the gas-phase neutral model can't form the zwitterion that dominates in water, so glycine's −COOH reads ≈3.9 rather than the measured ≈2.35 — the tool now says so up front instead of quietly handing you the wrong number.

UV-Vis & why the basis set matters

UV-Vis is a great lesson in how method choice drives accuracy. The strong π→π* absorption of paracetamol (experimental λ_max ≈ 243–249 nm) marches steadily toward experiment as the basis set improves:

This is why the Studio's Quick UV uses 3-21G and Advanced uses 6-31G*. TD-DFT is reliable for ordinary valence (π→π*, n→π*) excitations but should not be trusted for charge-transfer or Rydberg states.

The instant (empirical) NMR — and its honest limits

The Quick NMR returns shifts in milliseconds using substituent-additivity rules. For the chemistry it was built for it is remarkably good — but it knows its limits, and now tells you so.

When the instant estimate is unreliable, MoleBench shows a warning and offers to run the real quantum calculation instead — so a fast estimate never masquerades as a trustworthy one. And the Advanced (quantum) tier genuinely handles them — here it is on the very heteroaromatics the instant tier flags, within ~2–3 ppm of experiment:

The two tiers are complementary by design: the instant estimate for speed on its sweet spot, the quantum calculation for the cases it can't reach — and the tool always tells you which one you should be using.

How to read this

• Trends are more reliable than absolutes. Comparing two similar molecules with the same method cancels most systematic error — relative answers are the safest use of any of these tools.
• The model matters as much as the method. Several of the larger errors above are gas-phase vs. solution effects, not the quantum chemistry being wrong.
• Pick the right tier. Quick tiers are for speed and exploration; Advanced (quantum) tiers are for the numbers you'll quote.
• Everything here is reproducible. Build any of these molecules in the Studio and run the same calculation yourself.

Open the Studio →   Read the Lecture Notes

Benchmark run on the live MoleBench compute service. Experimental values from standard reference compilations (CRC Handbook, NIST, SDBS and the primary literature). Updated 2026-06-25.