spektran/spektran-multigas-v0

SPEKTRAN Multi-Gas v0 — Gas Mixture Benchmark Data

AI Agent Ready — load this dataset and run the full ML pipeline via natural language.

Multi-gas mixture regression benchmarks from SPEKTRAN. Real-world gas sensing always involves mixtures — these datasets challenge models to disentangle overlapping absorption features from multiple species.

Why Multi-Gas?

Single-species benchmarks are a simplification. In practice, laser-based gas analyzers must quantify target gases in the presence of spectral interference from other species. These datasets provide the first open benchmark for multi-species spectral regression:

• CH4 + CO2 + H2O: Combustion exhaust — the three dominant carbon/water species
• CO + CO2: Incomplete combustion detection — CO is the key safety indicator

Configs

Config	Target	Interferents	Application	Train	Test
ch4_co2_h2o	CH4	CO2 + H2O	Combustion exhaust	5,000	2,000
co_co2	CO	CO2	Incomplete combustion	5,000	2,000

Quick Start

from datasets import load_dataset

# Triple mixture: CH4 + CO2 + H2O
ds = load_dataset("spektran/spektran-multigas-v0", "ch4_co2_h2o")

# Dual mixture: CO + CO2
ds_co = load_dataset("spektran/spektran-multigas-v0", "co_co2")

Each record contains the target molecule's concentration_ppm plus interferent concentrations as interferent_1_concentration_ppm, interferent_2_concentration_ppm, etc.

Dataset Details

Config	Band (cm⁻¹)	Target range	Interferent ranges
ch4_co2_h2o	6047	CH4: 1–1,000 ppm	CO2: 100–50,000 ppm; H2O: 100–20,000 ppm
co_co2	2171	CO: 1–5,000 ppm	CO2: 100–50,000 ppm

• Technique: TDLAS direct absorption (DA)
• Instruments: 3 tiers (easy/medium/hard)
• Physics: Multi-species Beer-Lambert superposition with HITRAN demo lines

Citation

@software{spektran,
  title   = {SPEKTRAN: Simulation Engine and ML Benchmark for Optical Gas Sensing},
  url     = {https://github.com/spektran/spektran},
  doi     = {10.5281/zenodo.21790394},
  version = {0.5.1},
  license = {Apache-2.0}
}

Links

• GitHub · Docs · PyPI
• CH4 Dataset · CO2 Dataset · Industrial Dataset