Simulation Data and Classification Results for Flow Matching-Based Infrared Gas Identification

Description

This dataset supports the study “Flow Matching-Based Spectral-Shape Transfer for Infrared Spectral Gas Identification.” It contains 3,500 infrared absorption spectra simulated using HAPI and HITRAN spectroscopic parameters for ten gases: H₂O, CO₂, O₃, N₂O, CO, CH₄, NO, SO₂, NO₂, and NH₃. The spectra cover 400–4000 cm⁻¹ at 2 cm⁻¹ intervals and include single gases and binary and ternary mixtures. The archive also provides the scripts, configuration, source-sample indices, and ground-truth labels needed to reconstruct a separate synthetic test set of 13,200 spectra. This set comprises 45 binary and 120 ternary gas combinations, with 80 spectra per combination. The mixtures are synthesized from single-component reference spectra obtained from the NIST Chemistry WebBook (SRD 69); they are not independently measured mixtures. The NO₂ label uses a nitrogen dioxide/dinitrogen tetroxide reference record. The spectra/ folder contains the simulation arrays and sample metadata. The recipe/ folder contains the NIST record and download links, original contributor credits, reconstruction scripts, configuration, and software versions. The test/ground_truth.csv file gives the components present in each test sample. Original NIST spectra are not redistributed.

Steps to reproduce

Extract the ZIP archive. To use the 3,500 simulated spectra, open the NPZ file in the spectra/ folder with NumPy and consult the accompanying metadata CSV. No NIST download is required for this part. To reconstruct the synthetic mixture test set, obtain the ten JCAMP-DX files listed in recipe/references.csv. Create a folder named raw/ in the extracted archive and save each file using its gas label, for example raw/H2O.jdx and raw/CO2.jdx. Install NumPy, pandas, and SciPy using the versions listed in recipe/versions.txt. From the extracted archive root, run: python recipe/generate.py raw generated. Use an output folder that does not already exist. The script reconstructs 3,600 binary and 9,600 ternary spectra and checks the SHA-256 hashes of the spectral array, label array, and wavenumber grid against the archived test set. Use the combination and sample identifiers (combo_id, sample_id) to match samples to test/ground_truth.csv.

Institutions

• Southeast University

  Jiangsu, Nanjing

Categories

Licence