Nmr Unknown Identification
search/get_schema ("read JCAMP", "pick peaks", "searchread_jcamp for the vendor JCAMP-DX export, thensearch_chemical_by_formula from the elemental formula.brain_hpc_orca_nmr / brain_hpc_pyscf_nmr permatch_spectra for predicted-vs-measured agreement and residual errorOverview
Modern NMR identification is retrieval + ranking, not solving: candidates are proposed, their shifts are predicted, and the measured spectrum ranks them. This skill takes 1H (and 13C) data of an unknown, retrieves candidate structures from the elemental formula, predicts candidate shifts at QM depth, ranks against the measured shifts, and reports per-candidate evidence — match value, residual error, per-shift agreement, unexplained signals. It is a decision aid; the chemist confirms. The value is the judgment between the tool calls: how wide the candidate net goes, how deep the prediction runs, and when to abstain.
Guidance vs execution: ingestion, peak picking, candidate retrieval, shift prediction, and matching come from tested Paramus tools. Do not hand-roll a ranking or residual calculation in ad-hoc code.
When to Use
- NMR data of an unknown are in hand (plus an elemental formula, e.g. from MS) and the question is "what is this?"
- Small-molecule structure elucidation where ranked, evidence-backed candidates are the deliverable, not a single forced answer
Do not use for: verifying a known expected structure against a batch
spectrum (that is paramus-nmr-identity-verification — known to pass/flag; this
skill is unknown to ranked candidates), or crystal-structure work from XRD data
(paramus-crystal-structure-analysis — different evidence chain). No two skills
share a trigger: this skill's trigger is the unknown-spectrum question.
Domain Judgment
- Ranking is not confirmation: the top candidate is a hypothesis with evidence, and the chemist confirms — the skill never reports a ranked list as an identification.
- Evidence must be traceable to individual chemical shifts: a bare score without per-shift agreement and unexplained signals is not a deliverable.
- Confidence is conditional on correct referencing/processing and on the true structure being inside the searchable space — with formula-based retrieval only, say so in every report.
- Abstain is a valid outcome: large residuals everywhere, inseparable analogues, or missing metadata are abstain cases, not "pick the least bad" cases.
Failure modes
- No elemental formula available -> stop and ask for it (MS or elemental analysis); formula-based retrieval cannot start without it.
- True structure outside the formula-searchable space -> all candidates retain large residuals; abstain and say the retrieval space is the suspected limit — never force a ranking.
match_spectrarejects NMR input -> per-shift residual fallback from this procedure, flagged as manual verification; never silently drop the comparison.- Missing metadata (referencing, solvent, processing) -> abstain; a mis-referenced spectrum ranks the wrong candidates confidently.
- Candidate tools unavailable -> report which step is blocked and stop; a ranking without the prediction step is a guess and is labelled as such.
Example
Asked: "1H and 13C spectra of an unknown, formula C8H8O2 from MS — what is this?"
The skill ingests the JCAMP export with read_jcamp, obtains the measured shift
list with pick_peaks, retrieves structure candidates for the formula with
search_chemical_by_formula, predicts shifts per candidate with
brain_hpc_orca_nmr, and ranks them against the measurement with match_spectra.
The report is a ranked candidate list with per-candidate match value, residual
error, per-shift agreement, and unexplained signals — labelled a decision aid,
with the chemist confirming. Source for the retrieval-plus-ranking pattern and
its limits: the public 1H foundation-model preprint (Di Iorio et al., ChemRxiv
2026, doi:10.26434/chemrxiv.15007823/v1) — a 7.8M-parameter 1D 1H-spectrum
encoder trained on ~3.2M spectra with embeddings, denoising, peak maps, and
retrieval heads, explicitly not an elucidation solver; its weights and code are
Apache-2.0 (verified 2026-10-17), so a future shift-embedding retrieval tool
is legally clear to build. The same source frames the rules this skill applies:
identification is retrieval + ranking, not solving; "ranking is not confirmation";
abstain when residuals stay large, analogues cannot be separated, or metadata are
missing.
Input
measured spectrum (JCAMP-DX export), elemental formula (e.g. from MS),
Output
ranked candidate structures with per-candidate evidence (match value,
Orchestrates · 6 tools
The skill decides at runtime which to call.
read_jcamp pick_peaks search_chemical_by_formula brain_hpc_orca_nmr brain_hpc_pyscf_nmr match_spectra
Frequently Asked Questions
What does Nmr Unknown Identification do?
Identify an unknown small molecule from its 1H/13C NMR spectrum - retrieve candidate structures from the elemental formula, predict their shifts, rank them against the measurement, and report per-candidate evidence (match value, residual error, per-shift agreement, unexplained signals) as a decision aid; the chemist confirms. Use when a user has NMR data of an unknown and asks "what is this?". Multi-tool WORKFLOW with candidate-recall, prediction-depth, and abstain judgment.
What do I provide and what do I get back?
You provide: measured spectrum (JCAMP-DX export), elemental formula (e.g. from MS),. You get: ranked candidate structures with per-candidate evidence (match value,.
Which tools does it orchestrate?
6 catalog tools, chosen at runtime: read_jcamp, pick_peaks, search_chemical_by_formula, brain_hpc_orca_nmr, brain_hpc_pyscf_nmr, match_spectra.