Quickstart

Option 0: AI Agent (zero code)

If you use an AI coding agent (Claude Code, Cursor, GitHub Copilot, Windsurf), just tell it what you want:

"Train the ridge baseline on the concentration task and show me the MAE"

The agent reads AGENTS.md and operates the full pipeline through spektran CLI commands with --json output. No manual steps needed. See the AI Agent interface for the full command reference.

Option 1: just the data (no install)

The official v0 splits are hosted on Hugging Face:

from datasets import load_dataset
ds = load_dataset("spektran/spektran-ch4-v0")

Everything below is for running the engine, regenerating data bit-for-bit, and scoring benchmark submissions.

Install

git clone https://github.com/spektran/spektran
cd spektran
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"

One clean spectrum (offline, ~10 lines)

from spektran.physics import simulate_absorbance

nu, absorbance = simulate_absorbance(
    molecule="CH4", concentration_ppm=100.0,
    temperature_K=296.0, pressure_atm=1.0, path_length_m=10.0,
    wavenumber_start_cm1=6046.0, wavenumber_end_cm1=6048.0,
)

Uses the built-in approximate CH4 demo line list. For authoritative spectra install the HITRAN extra (pip install -e ".[hitran]") and pass lines=fetch_lines("CH4", 6045.0, 6049.0) (network required on first call).

Multi-species example

from spektran.physics import demo_ch4_2nu3, demo_h2o, absorption_coefficient
import numpy as np

nu = np.linspace(6046.0, 6048.0, 2000)
alpha_ch4 = absorption_coefficient(nu, demo_ch4_2nu3(), 100e-6, 296.0, 1.0)
alpha_h2o = absorption_coefficient(nu, demo_h2o(), 0.01, 296.0, 1.0)  # 1% H2O
total = alpha_ch4 + alpha_h2o  # Beer-Lambert linear superposition

Demo line lists are also available for CO2 (demo_co2()) and CO (demo_co()). Each demo line list is centered on its own representative sensing band (CH4 ~6047 cm⁻¹, H2O ~7187 cm⁻¹, CO2 ~4978 cm⁻¹, CO ~2171 cm⁻¹), so combining two demo species over one narrow scan window mostly exercises the superposition code path rather than showing visible spectral overlap. For a window with genuine cross-species interference, fetch HITRAN lines for both species over the same range with fetch_lines().

Generate an official dataset split

Using the CLI (v0.2+):

spektran generate configs/datasets/ch4-t1-train-v0.yaml --out data

Or using the script directly:

python scripts/generate_dataset.py configs/datasets/ch4-t1-train-v0.yaml --out data

Reproducible bit-for-bit: the config pins the master seed, instrument configs, and gas-truth distributions. The four official v0 splits total ~285 MB on disk and generate in well under a minute. (configs/datasets/ch4-da-medium-v0.yaml is a standalone 10k-record demo config, not part of the benchmark splits.)

Train and score a baseline

One command (auto-generates data if missing):

spektran train --baseline ridge --json

Step by step (manual control):

pip install scikit-learn torch   # torch only needed for the CNN baseline
for s in t1-train t1-val t1-test t3-test-heldout; do
  spektran generate configs/datasets/ch4-$s-v0.yaml --out data
done
python baselines/ridge_regression/train.py   # ~20 s
spektran benchmark --task T1-concentration \
  --truth data/ch4-t1-test-v0.h5 \
  --predictions baselines/ridge_regression/predictions_t1-test.csv

The CNN baseline (baselines/cnn1d/train.py) takes ~7 minutes on CPU; full scoring commands for both models, including the T3 --t1-mae convention, are in baselines/README.md.