ChefNMR2

ChefNMR2 is a generative model for small-molecule structure elucidation built by Princeton University (the Ellen D. Zhong lab in Computer Science and the Mohammad R. Seyedsayamdost lab in Chemistry) together with AI-for-science company Edison Scientific. Given a molecular formula plus one-dimensional ¹H and ¹³C NMR peak lists, it generates candidate molecular structures directly.

The paper, Small molecule structure elucidation from NMR spectra with generative and reasoning models, appeared on bioRxiv on October 29, 2026. On November 2, 2026 (US time) Edison announced that ChefNMR2 is now available inside Kosmos, its autonomous AI-scientist platform, so researchers can call it without running the model themselves.

How it works

• Atomic flow matching: structure elucidation is framed as generating atomic coordinates conditioned on the formula and NMR peaks. The architecture is a peak encoder plus a multimodal diffusion transformer with joint attention between peak tokens and atom tokens. At inference, noise is transformed into coordinates, then bonds are assigned to obtain candidate structures.
• Calibrated confidence: repeated sampling yields an empirical confidence (how often each planar structure appears). After an initial 100 samples, the sampling budget adapts to that confidence.
• ChefNMR2-Agent: an LLM reasoning agent (running on Kosmos) wraps the generator, adding a natural-language interface, chemical reasoning, computational tools and literature search, and digging deeper only when the model is uncertain.

Key numbers

Metric	Value
Training data	2.52M molecules with experimental and simulated spectra
Evaluation	2,031 molecules across five test sets
ChefNMR2 top-1	82.3%
Best specialized de novo model (NMRTransformer)	35.4%
Retrieval method (NMR-Solver)	60.1%
Best LLM (Claude Opus 5)	52.0%
Speed / cost	up to 2,170x lower median runtime, 360x lower median cost per molecule

The team used ChefNMR2-Agent to solve the structures of four novel natural products, matching human-chemist accuracy in a fraction of the time.

Caveats

The work is a preprint and not yet peer reviewed. Some baselines had training overlap with test molecules, and accuracy falls to roughly 57% on molecules that are both large and dissimilar from the training data. The authors position it as ranked candidates for expert review, not an autonomous certificate of molecular identity.

Who it is for

• Researchers in natural products, medicinal chemistry and synthesis who need to infer unknown small-molecule structures from 1D NMR quickly.
• Labs and CROs with large spectral archives that want batch first-pass structure proposals.
• AI-for-science practitioners interested in the "specialized generator + general reasoning agent" pattern.

Getting started

1. Read the project page at chefnmr.cs.princeton.edu and the bioRxiv preprint for input format (formula + ¹H/¹³C peak lists) and limitations.
2. Use ChefNMR2 / ChefNMR2-Agent inside Edison Scientific's Kosmos platform: provide peak lists and sample context.
3. Review the ranked candidates and confidence: confirm high-confidence answers quickly, and let the agent (or a chemist) investigate low-confidence ones.
4. Validate key conclusions experimentally (2D NMR, mass spectrometry, etc.).