Report 081 · AI in the Lab
Getting a molecular structure out of a spectrum is what I did for years, by hand, one molecule at a time. So when a model claims to do it, I read the metric before the headline. The most-cited number in this field is 69.6%, and two things about it are worth knowing: it is not a single answer, it is fifteen, and it was measured on spectra a computer generated. Fed 106 real ones, the same framework scored 33.0%.
Structure elucidation is the part of chemistry that never got easier. You have a vial of something. You do not know what it is. You put it through an instrument, get back a pattern of peaks, and your job is to work backwards from that pattern to an arrangement of atoms in space.
My own version of this is microwave rotational spectroscopy, which is a stricter dialect of the same language. You measure how a molecule tumbles, you get rotational constants, and then you propose a geometry and check whether it reproduces the numbers. The check is brutal and it is the entire point. A wrong geometry does not almost work. It misses.
NMR is the workhorse version, the one running in every synthetic chemistry building on earth. It is less absolute than rotational spectroscopy and enormously more general, and interpreting it is a skill people spend a decade getting good at. So the prospect of a model that reads a routine spectrum and hands back a structure is not hype-bait to me. It is genuinely interesting, and it is the kind of thing I would use.
Which is exactly why the number matters.
The paper that anchors this field is Hu and colleagues, from Stanford, in ACS Central Science, November 2024. They built a multitask machine-learning framework that predicts a molecule's formula and connectivity from ordinary one-dimensional 1H and 13C NMR spectra, with no prior chemical knowledge supplied, not even the molecular formula.
That last constraint is a real one, and I want to give the work its due. Withholding the molecular formula makes the problem much harder than the way a working chemist usually meets it, and the paper is upfront about the scale of what it is searching. Their headline sentence, verbatim:
Without relying on any prior chemical knowledge such as the molecular formula, we show that our approach predicts the exact molecule 69.6% of the time within the first 15 predictions, reducing the search space by up to 11 orders of magnitude.
Read the middle clause again: within the first 15 predictions.
That is not an error and it is not hidden. It is stated plainly in the abstract, in the same sentence as the number, which is more than a lot of papers manage. But it is the first thing that falls off when the number travels. "69.6% accurate" is what survives the trip to a slide deck, and 69.6% accurate is not what was measured. What was measured is that the correct molecule appeared somewhere in a list of fifteen, about seven times in ten.
Those are different products. One is an answer. The other is a shortlist, and a shortlist is only useful to someone qualified to work through it.
Now the part that decides whether any of this reaches your bench.
The 69.6% was measured on simulated spectra. The molecules span up to 19 heavy atoms, a space the paper notes contains trillions of possible structures, and the spectra were computed rather than recorded. The same paper then took the framework and pointed it at real data, and reported the result, verbatim:
the accuracy of the framework drops from 69.6% to 33.0% when used to predict the structures from 106 experimental 1H and 13C NMR spectra
Same model. Same fifteen-guess allowance. Half the performance.
I have a lot of respect for the decision to publish that sentence. It would have been trivially easy to report the simulated benchmark, call it state of the art, and leave the experimental test for someone else to run in three years. They ran it themselves and put the disappointing half in the abstract.
It is also the most important sentence in the paper, because it quantifies the gap between the benchmark and the bench.
A simulated spectrum is what the physics says a lone, ideal molecule would produce. A real spectrum is what an instrument recorded from an actual sample in an actual solvent on an actual Tuesday.
The NMRTrans group, whose work I come to in a moment, states the difference compactly: simulated spectra "inevitably deviate from experimental measurements due to complex solvent effects, impurities, and magnetic interactions."
Unpack that from the operator's side, because each item is a specific thing that goes wrong. Solvent shifts peaks, and shifts them by different amounts depending on which solvent and what else is dissolved in it. Impurities add peaks that belong to no part of your molecule, and a residual solvent peak or a trace of grease is not labeled as such in the data. Magnetic interactions mean real coupling patterns, real linewidths, real second-order effects that a fast simulation approximates or ignores. Add baseline drift, phasing, shimming, water suppression, and concentration effects.
A trained chemist handles all of this without consciously noticing. They know that peak is chloroform, that hump is the baseline, that shoulder is a rotamer. The knowledge is so automatic it is nearly invisible, which is precisely why it is easy to forget it is not in the training data when the training data is synthetic.
The model was trained on a world where none of that happens, and then asked to work in a world where all of it does. A 69.6 to 33.0 drop is what that costs.
The field knows this, and the response has been the right one: stop training on simulated data. A preprint posted to arXiv on 10 February 2026, NMRTrans, describes what its authors call "the first NMR Transformer trained solely on large-scale experimental spectra," built on a corpus of real 1H and 13C spectra mined from the chemical literature.
That is the correct fix and it works. Their reported results on experimental benchmarks are the current high-water mark, and I am going to give you all four numbers rather than the flattering one:
Top-1 accuracy: 42.81%, against 37.33% for the strongest prior model.
Top-10 accuracy: 61.15%, against 43.33%.
The paper's own summary sentence leads with the second pair: state-of-the-art performance, "improving Top-10 Accuracy over the strongest baseline by +17.82 points (61.15% vs. 43.33%)." A 17.8-point jump is a substantial result and I do not want to undersell it. Real progress happened here.
But hold the two framings side by side, because the distance between them is the whole subject of this report.
The optimistic reading: on real spectra, the right structure is in the model's top ten about 61% of the time. That is a tool.
The literal reading: hand this system a real spectrum and its single best guess is correct 43% of the time. Better than a coin flip. Not something you would put in a paper without checking.
Both are true. Only one of them is a headline, and it is not the one that tells you how to use the thing.
Top-k accuracy is not a trick. It is the honest way to score a system that produces a ranked list, and structure elucidation genuinely is a ranking problem: you narrow, then you discriminate. Every chemist doing this manually also carries a shortlist in their head.
But top-k quietly relocates the hard part. A model reporting top-10 has not identified your compound. It has cut a space of trillions down to ten, which is a genuinely enormous reduction and is worth the eleven orders of magnitude the Stanford paper claims for it. Then it hands the ten to you.
And discriminating among ten plausible candidates that all roughly match the spectrum is not the easy remainder of the job. It is frequently the hard part. Ten candidates that survived a spectral filter are, by construction, ten structures whose predicted spectra look similar. Telling them apart is where regiochemistry, stereochemistry, and two-dimensional experiments come in. That work needs the expert the automation was supposed to replace.
So the deliverable is real, and it is triage. That is worth having. It is not what "AI determines molecular structures" puts in a reader's head.
Here is the thing that would keep me careful, and it is not the accuracy number.
These systems learn a mapping from spectral patterns to structures that appeared in their training corpus. NMRTrans mined its spectra from the published literature, which is the best available source of real data and also a specific, non-random slice of chemical space: the molecules people made, succeeded with, and published.
Ask what you actually reach for structure elucidation to do. Usually it is because you have something unexpected. An unknown natural product. A degradant. A side product from a reaction that misbehaved. A contaminant. Something, in other words, that is more likely than average to sit outside the distribution of published compounds.
The routine known compound, where these models score best, is the case a chemist often resolves in twenty minutes anyway. The genuinely novel scaffold, where you most need help, is where a distribution-matching model is weakest, and nothing in the output announces which case you are in. A ranked list of ten looks identical whether the true answer is in it or not.
This is the same shape I described when structure predictors returned a confident wild-type fold for a protein that had measurably changed. The output format does not vary with the model's competence. Confidence is presentational, and the interface offers no boundary marker for the edge of the training distribution.
I would use these tools. Specifically:
Dereplication. You have a spectrum and you suspect it is a compound already known. Narrowing to ten candidates in seconds is an excellent use of the technology and a real time saving.
Getting unstuck. Staring at an unfamiliar pattern, a ranked list of plausible structures is a decent generator of hypotheses, provided you treat every entry as a hypothesis rather than a conclusion.
Screening at volume. Where you have far more spectra than chemist-hours, a 43% top-1 hit rate on the easy ones is a lot of triage you were not otherwise going to do.
And the boundary: not as the last step before you write down a structure. Something else has to be able to say no. A two-dimensional experiment, a mass spectrum, an independent synthesis, a comparison with an authentic sample. If nothing in your workflow can contradict the model, your workflow cannot detect that the model was wrong.
Said plainly, because the alternative is letting a claim ride that I did not check.
NMRTrans is a preprint. Posted to arXiv on 10 February 2026, not peer reviewed at the time I read it. I have quoted its reported numbers as reported numbers, and self-reported benchmark results from a preprint are the weakest evidence in this report. Treat the 42.81% and 61.15% as the authors' claim, not as an established fact. The 2024 Stanford result is peer reviewed and open access, and I read its abstract verbatim in full from the PubMed Central deposit.
Two 2026 papers stayed shut. A multi-spectral model in Angewandte Chemie and a multimodal structure-elucidation paper in Nature Communications both appeared in my search and both refused my requests, one behind a paywall and one behind a login. Neither informs a single sentence above. There may be stronger numbers in them, in either direction.
I have not run any of these models. Everything here is read off published benchmarks. The published benchmark is not the same instrument as your instrument, your solvent, and your sample.
The useful habit is small: when a model reports accuracy, ask two questions before you believe the number.
How many guesses? Top-1 and top-15 are different products, and the difference is who does the remaining work. A shortlist is a service. An answer is a claim. Only one of them ends the task.
Measured on what? Simulated data measures whether the model learned the simulator. Real data measures whether it learned the problem. Where a paper reports both, as this one honorably does, the second number is the one that describes your Tuesday.
None of which is a case against AI in the lab. The 2024 result is good work, the 2026 preprint is better work, and training on experimental spectra is exactly the right correction. Six months from now the numbers will be higher.
The reason to hold the metric straight is not skepticism about the tools. It is that a spectrum is a physical measurement of a real sample, and a structure is a claim about what that sample is. Between the two sits an inference, and the entire discipline of structure elucidation consists of refusing to let that inference go unchallenged. A model that narrows trillions to ten has done something remarkable. It has not done the refusing.
U.S. Army combat veteran (Counter-IED / Electronic Warfare), peer-reviewed researcher in microwave spectroscopy, and founder & CEO of Shroombiosis. Consults on laboratory operations, AI, and supplement formulation.