FTIR.fun FTIR.fun Spectrum evidence workspace
Three-Axis Engine · v3

One spectrum. Three independent verdicts.

FTIR.fun does not rank by a single similarity score. Every uploaded spectrum is cross-examined along three independent axes — library similarity, characteristic-peak reasoning, and knowledge-graph validation — and only when they agree do we surface a high-confidence call. When they disagree, we say so out loud.

Axis 1 · Library Similarity Axis 2 · Peak Reasoning Axis 3 · KG Validation Explainable
Try a sample now See full service portfolio
// 01 — Engine

Three independent axes, one verdict

Each axis is computed by a different module, on different data, with different failure modes. Disagreement is surfaced — not hidden — so reviewers can see why a hit is trusted or why a sample needs a second look.

// AXIS 01

Spectral Library Similarity

Hash-based pre-screen narrows 130k+ reference spectra to a workable Top-N, then curve-level cosine and derivative scoring rank the candidates.

  • Peak20 / Peak50 indexing for sub-second recall at scale
  • Top-1 vs Top-2 gap surfaced — not just the winner
  • Robust to baseline drift and spectrum rescaling
top1=PMMA-std-021
cosine=0.972 · deriv=0.948
gap(top1-top2)=+0.071
// AXIS 02

Characteristic-Peak Reasoning

Auto-extracted peaks are mapped to functional-group rules, with shift tolerance and shape constraints — not a one-shot list compare.

  • Tolerant matching: ±5–15 cm⁻¹, FWHM-aware
  • Group-level rules linked to the underlying chemistry
  • Outputs an assignment table, not just a score
C=O ester @ 1729±4 ✓
O-CH₃ @ 1148±6 ✓
α-CH₃ @ 2950±8 ✓
// AXIS 03

Knowledge-Graph Validation

21M+ curated relations link materials → peak families → functional groups. The candidate must be supported by the graph, with no hard conflicts.

  • Material → group → peak triples, sourced from literature & curated cases
  • Catches plausible-looking but chemically impossible hits
  • Surfaces conflicts as red flags, not silent rejections
material=PMMA · family=acrylic_ester
supports={C=O, O-CH₃, α-CH₃}
conflicts=∅
// 02 — Feedback Loop

The engine sharpens every time you contribute

Three axes are only as honest as the data behind them. Every shared standard, every confirmed verdict, every corrected assignment flows back into the indexes, the rules, and the graph.

Why it gets better month over month

When users upload a known standard, it doesn't just sit in their workspace — it expands the library coverage on Axis 1 and seeds new training cases for the peak rules on Axis 2.

When users confirm or correct a verdict, that signal lands in the graph on Axis 3 — strengthening the relations the engine got right and weakening the ones it got wrong.

Concretely, contributions feed three pipelines: (a) library expansion, (b) peak-rule calibration, (c) graph-fact reweighting. Reviewer-confirmed cases carry more weight than passive uploads, by design.

130k+ Standard spectra
21M+ KG relation facts
600k+ KG entities
// 03 — Worked Examples

Two verdicts, two contracts with the user

High match returns a specific material. Low match returns a direction with concrete next-step verification, never a wild guess. Both come with a peak-by-peak interpretation and a literature trail.

Case A · tier=definitive claim_kind = material

Transparent acrylic pellet → identified as PMMA

Sample: transparent thermoplastic pellet, ATR pickup, no obvious filler. Uploaded as an unknown; operator did not pre-label.

AXIS 01 Top-1 PMMA-std-021, cosine 0.972 · Top-2 gap +0.071 · above the definitive_l1_similarity_min = 0.90 threshold
AXIS 02 Five diagnostic peaks of the methacrylate ester family all matched within ±5 cm⁻¹ tolerance ✓
AXIS 03 Material PMMA → family acrylic_ester; agreement_count = 3, no hard_forbidden conflict ✓
Verdict Poly(methyl methacrylate) — PMMA. All three axes agree at the material level; the case satisfies the definitive tier rule (L1 ≥ 0.90, ≥ 1 supporting layer, agreement_count ≥ 1, L3 has at least one citation-bound peak match).

Detailed explanation

The Top-1 spectrum-level similarity (0.972) clears the definitive threshold and the +0.071 gap to Top-2 indicates a clean separation rather than a tie. Both Axis 02 (peak rules) and Axis 03 (knowledge graph) independently identify the methacrylate-ester family, with the graph's PMMA → acrylic_ester edge fully covering the observed peak set. No competing family has a hard_forbidden veto. The verdict is therefore released as material-level (claim_kind = material) without recourse to alternatives.

Characteristic-peak interpretation

  • 1729 cm⁻¹ ester C=O stretch — acrylate carbonyl, position diagnostic of methacrylate (cf. aliphatic ester ~1738, aromatic ester ~1715)
  • 1240 cm⁻¹ C(=O)–O asymmetric stretch of the ester linkage
  • 1148 cm⁻¹ O–CH₃ stretch — methyl ester signature, distinguishes methacrylate from acrylate
  • 988 cm⁻¹ C–C–O methacrylate skeletal mode
  • 2950 / 2995 cm⁻¹ α-CH₃ asymmetric / symmetric stretch — quaternary carbon environment of the methacrylate backbone

References

  1. Coates, J. Interpretation of Infrared Spectra, A Practical Approach. In Encyclopedia of Analytical Chemistry; Meyers, R.A., Ed.; Wiley, 2006.
  2. Lin-Vien, D.; Colthup, N.B.; Fateley, W.G.; Grasselli, J.G. The Handbook of Infrared and Raman Characteristic Frequencies of Organic Molecules; Academic Press, 1991 (esters, Ch. 9).
  3. Socrates, G. Infrared and Raman Characteristic Group Frequencies, 3rd ed.; Wiley, 2001, pp. 99–116.
  4. Willis, H.A.; Cudby, M.E.A. Spectrochim. Acta A, 1968, 24, 1701 (PMMA backbone vibrations).
Case B · tier=directional claim_kind = family

Translucent film fragment → polyester family, but PET / PBT cannot be separated by IR alone

Sample: thin translucent film fragment from a packaging return. ATR pickup, no overt filler, slight surface haze. Uploaded as an unknown.

AXIS 01 Top-1 PET cosine 0.62, Top-2 PBT 0.59; gap +0.03 — below definitive_l1_similarity_min = 0.90, no material-level winner
AXIS 02 Terephthalate ester peak set fully present (1714 / 1240 / 1095 / 1410 / 870 / 723); aliphatic-spacer fingerprint region inconclusive ✓ !
AXIS 03 Three axes converge on family polyester / aromatic_terephthalate_ester; alternatives kept = {PET, PBT}; no hard_forbidden conflict ✓
Verdict Aromatic polyester (terephthalate ester family); most likely PET, PBT cannot be ruled out by FTIR. Case satisfies the directional tier rule (≥ 2 supporting layers, agreement at family level). Two alternatives kept; further verification required before any material-level claim.

Detailed explanation

The terephthalate carbonyl, ring and ester C–O–C bands are unambiguous, so the engine commits to the polyester / aromatic terephthalate ester family. The reason it stops there: PET and PBT differ only in the aliphatic glycol spacer (–OCH₂CH₂O– vs –O(CH₂)₄O–), which affects only the weak CH₂ rocking/wagging modes between roughly 970 and 1000 cm⁻¹. Under ATR with no orientation control these contrasts are below the engine's confidence floor — a definitive material call would be over-claiming.

Characteristic-peak interpretation

  • 1714 cm⁻¹ aromatic ester C=O stretch — terephthalate carbonyl (lower than aliphatic ester ~1738 due to ring conjugation)
  • 1410 cm⁻¹ aromatic C=C ring stretch of para-disubstituted benzene
  • 1240 cm⁻¹ C(=O)–O asymmetric stretch of the aromatic ester
  • 1095 cm⁻¹ C–O–C symmetric stretch of the glycol portion (PET) / butanediol portion (PBT)
  • 870 cm⁻¹ in-plane C–H bending of the para-disubstituted benzene ring
  • 723 cm⁻¹ out-of-plane C–H bending of the para-disubstituted ring — a fingerprint of terephthalate-based polyesters
  • ~975 cm⁻¹ weak CH₂ rocking — in principle the PET/PBT discriminator, but here too weak / overlapped to be decisive !

Recommended next steps

  • Py-GC/MS — pyrolysis fragments cleanly separate ethylene glycol (PET) from 1,4-butanediol (PBT) units. The single most decisive test.
  • DSC — Tm separates PET (~250 °C) from PBT (~225 °C); a 5 mg cut and one heating ramp are usually enough.
  • Density / solubility — PET ≈ 1.38 g·cm⁻³, PBT ≈ 1.31 g·cm⁻³; HFIP / o-chlorophenol solubility differ. Use only as supplementary evidence.
  • Re-sample — if film is too thin or oriented, request a melt-pressed flake to remove orientation artefacts before any IR re-run.

References

  1. Cole, K.C.; Ben Daly, H.; Sanschagrin, B.; Nguyen, K.T.; Ajji, A. A new approach using FTIR to characterise PET and PBT. Polymer, 1999, 40, 3505–3513.
  2. Socrates, G. Infrared and Raman Characteristic Group Frequencies, 3rd ed.; Wiley, 2001, pp. 99–123 (terephthalate esters).
  3. Pretsch, E.; Bühlmann, P.; Badertscher, M. Structure Determination of Organic Compounds, 4th ed.; Springer, 2009.
  4. ASTM E1252-98(2021). Standard Practice for General Techniques for Obtaining Infrared Spectra for Qualitative Analysis.
  5. Tsuge, S.; Ohtani, H.; Watanabe, C. Pyrolysis-GC/MS Data Book of Synthetic Polymers; Elsevier, 2011 (PET vs PBT pyrograms).

Run your sample through all three axes.

Upload a spectrum and read the AI report — every claim is traceable to one of the three axes.

Start a search Talk to an expert
Submit Request Form