UltraIR

An absorption band near 1700 cm⁻¹ tells a chemist there is a carbonyl in the sample, but turning a whole infrared spectrum into a structure, a mixture composition, or a sample identity still depends on expert peak assignment, rule tables, and a reference spectrum close enough to match — which fails precisely where IR is most valuable, on unfamiliar compounds and complex real-world samples. Machine-learning readers have eroded the problem task by task, but each is fitted to one objective on one dataset, needs a large labeled spectral collection, and transfers poorly to another analytical goal or spectrometer.

UltraIR attacks that shortage from the simulation side. Calculation produces IR spectra far faster than a laboratory can measure them, so a shared spectral encoder is pretrained once on approximately 60 million simulated spectra, then adapted to each downstream objective with the limited experimental labels that actually exist. Pretraining combines three complementary objectives: wavelet-domain spectral reconstruction, which forces the encoder to retain both the coarse spectral envelope and fine absorption line shapes; molecular fingerprint similarity alignment, a soft-Tanimoto contrastive loss that organizes the latent space by graded structural similarity instead of hard positive and negative pairs; and multi-label functional-group prediction, which anchors the representation to interpretable chemical motifs.

The model comes from the Information Hub at the Hong Kong University of Science and Technology (Guangzhou), with collaborators at Jilin University, the University of Auckland, Shanghai AI Laboratory, Hunan University and HKUST. It is the infrared counterpart of UltraNMR from an overlapping group and shares its simulation-to-real thesis; where SpecMol reads spectra as text tokens inside a language model, UltraIR is a dedicated spectral encoder built around the shape of an absorption curve.

#Key Features

• Simulation-to-real transfer: Chemical breadth comes from simulated spectra, task supervision from the few thousand experimental spectra a lab can realistically label. A matched ablation without pretraining fell below conventional baselines on medicinal-herb constituent quantification.
• Derivative-aware spectral input: A Savitzky–Golay filter and two learnable derivative convolutions add first- and second-derivative channels alongside the raw absorbance, gated by an input-dependent weighting so the model can emphasize line shape or intensity as the task requires.
• One encoder, many output heads: Functional-group prediction, structure elucidation, property regression, mixture-component detection and quantification, and sample classification all initialize from the same pretrained encoder.
• Cross-instrument generalization: Trained on soil spectra from one laboratory and evaluated zero-shot on a subset acquired with a different FTIR instrument, it was the only method to hold a positive R² across every evaluated property.
• Biological and botanical sensing: It classifies fast-growing bacterial isolates from green snow to genus, and traces Jinyinhua and Shanyinhua honeysuckle to their province of origin while quantifying their chemical constituents.

#Technical Details

UltraIR is a hybrid convolutional–Transformer encoder with more than 100 million parameters. Spectra are cropped to 400–4000 cm⁻¹, resampled onto a 3,600-point grid and min–max normalized; the derivative module feeds four residual spectral blocks with squeeze-and-excitation recalibration and multi-scale fusion, whose output is tokenized by overlapping patch embedding and passed to a Transformer returning a [CLS] representation. The pretraining corpus draws roughly 1.5 million published simulated spectra from IRtoMol and QM9S, about 7.5 million newly generated molecular-dynamics spectra, and about 51 million from a machine-learning spectral predictor, with downstream-dataset molecules removed to prevent molecule-level overlap. Around 120,000 experimental spectra support adaptation, in which encoder and freshly initialized head are optimized jointly; structure elucidation is cast as formula-conditioned autoregressive SMILES generation with beam search and formula-consistency reranking. Across NIST, SDBS and USPTO the model leads on exact match ratio for functional-group prediction and on top-1, top-5 and top-10 structure recovery, and reaches R² of 0.956, 0.986 and 0.996 for targeted fractional contribution estimation.

#Applications

The model suits labs that already run FTIR and want more out of it: proposing structures for unknowns, screening functional groups, and quantifying a target component in a mixture from a single spectrum. Its accuracy at small labeled-data fractions matters most where reference measurements are expensive — herbal medicine authentication, microbial identification from whole-cell spectra, and soil chemistry surveys.

#Impact

UltraIR carries the simulation-to-real recipe already demonstrated for NMR into vibrational spectroscopy, and the soil case study is its clearer contribution: a representation surviving a change of instrument, laboratory and sample preparation addresses the transfer failure that keeps most spectral models confined to the bench that trained them. The limitations are stated plainly by the authors. Every objective still requires supervised adaptation and its own output head, so what is on offer is transferability after fine-tuning rather than universal inference, and a domain gap persists because calculated spectra do not reproduce instrument response, resolution, scattering or matrix effects. The work is an arXiv preprint awaiting peer review, and the code and checkpoint repositories are declared MIT but carry neither code nor weights yet.