The chemistry lab is closing the loop — design, synthesis and measurement in one pass
Four preprints posted this week cover ligand design, autonomous electrosynthesis, structure identification from NMR, and lab resource tracking. Together they describe a lab that runs itself.
- In a preprint, the ReConLiG system recommended ligands with a 93.3% top-5 hit rate. Self-reported m:…3371
- In a preprint, the eRoboChem platform ran autonomous electrosynthesis across paired, oxidative and reductive modes, using Bayesian optimisation rather than one-factor-at-a-time testing. Self-reported m:…5940
- In a preprint, a transformer trained on 1 million experimental NMR spectra reached 56.6% top-15 accuracy on molecules up to 19 heavy atoms, beating the NMR2Struct model by four times on 86 spectra. Self-reported m:…7584
- In a preprint, the LabSense system tracked electricity, water, fume-hood activity and chemical inventory in materials chemistry labs using low-cost Raspberry Pi hardware. Self-reported m:…9405
Four tools, one direction
Four preprints posted this week come from different corners of chemistry — ligand design, electrosynthesis, structure identification and lab operations — but they point the same way. Each one replaces a step that used to need a chemist's judgment with a system that proposes, runs, measures or tracks on its own. None of them claims the chemist is gone. What they describe, taken together, is a lab where the human decides which question to ask and the loop of design, synthesis and measurement runs with less hand-holding at every step. m:…3371 m:…5940 m:…7584 m:…9405
Designing ligands from almost no data
The ligand-design preprint, ReConLiG, tackles a familiar problem: good ligands for organometallic reactions are hard to find, and the reaction-specific data needed to train a model is scarce. The system separates ligand generation from reaction guidance. It recommended ligands with a 93.3% top-5 hit rate, and its generated ligands matched experimentally preferred ones far more closely than a baseline fine-tuned decoder — a mean Tanimoto similarity of 0.82 against 0.29. Across gold, iridium and palladium-catalysed reactions, the system produced yields of 70 to 99% with selectivity up to 94% enantiomeric excess or better than 20:1 diastereomeric ratio. The point is data efficiency: a specialised corner of chemistry can now get a working generative model without a large labelled dataset. m:…3371
A robot that runs the reaction
The eRoboChem preprint describes a modular autonomous platform for synthetic electrochemistry that pairs Bayesian optimisation with both flow and batch reactors. It handles paired, oxidative and reductive transformations, runs closed-loop flow operation for gas-evolving reactions, and keeps a human in the loop for batch electrode screening, where electrode passivation is a recurring problem. The researchers report that Bayesian optimisation navigated the complex reaction space better than one-factor-at-a-time and design-of-experiments strategies. The practical claim is flexibility: one platform, rather than a bespoke setup per reaction type, that can be pointed at new electrochemical syntheses. m:…5940
Reading a structure from its spectrum
The third preprint attacks the other end of the workflow: figuring out what you made. A transformer model trained on 1 million experimental NMR spectra takes 1D proton and carbon peak lists plus molecular weight and proposes molecular structures. It reached 56.6% top-15 accuracy for molecules up to 19 heavy atoms, and on a benchmark of 86 experimental spectra it was more than four times as accurate as the NMR2Struct model. Training on experimental rather than simulated spectra also produced more chemically valid candidates, and the model offered structurally relevant starting points even when it missed the exact structure. For a working chemist, a system that narrows identification to a shortlist is useful even when the exact structure is not in its top 15 — provided the shortlist is checked, not trusted. m:…7584
And someone has to watch the meter
The fourth preprint is the least glamorous and possibly the most telling about where labs are headed. LabSense is a modular, open-source system built on Raspberry Pi hardware that continuously monitors electricity, water, fume-hood activity and chemical inventory in materials chemistry labs, turning those streams into real-time data for operational decisions. Autonomous labs run more experiments, and more experiments consume more resources; a lab that cannot measure its own footprint cannot manage it. All four preprints are unreviewed, so their numbers should be treated as claims to verify, not settled facts — but the direction is consistent: the scarce resource in chemistry is shifting from hands at the bench to judgment about which experiments are worth running. m:…9405
What is still unresolved
- Whether ReConLiG's hit rates hold up on reaction types outside the gold, iridium and palladium cases tested.
- How eRoboChem performs on reactions beyond the transformations reported in the preprint.
- Whether the NMR model's accuracy holds for larger molecules beyond 19 heavy atoms.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.