spektran/spektran-baselines-v0

SPEKTRAN Baseline Checkpoints

Pre-trained baseline model weights for the SPEKTRAN gas-sensing benchmark. These allow inference on SPEKTRAN data without retraining.

Models

Checkpoint	Task	Input	Output	Val MAE
ridge-t1-da/	T1 Concentration	DA raw scan (2000 pts)	ppm	2.80 ppm
cnn1d-t1-da/	T1 Concentration	DA raw scan (2000 pts)	ppm	16.79 ppm
ridge-t4-wms/	T4 WMS Concentration	WMS 2f signal	ppm	12.46 ppm
ridge-t9-temperature/	T9 Temperature	DA raw scan (2000 pts)	K	7.03 K

Quick Start

Ridge (scikit-learn)

import numpy as np

data = np.load("ridge-t1-da/weights.npz")
# Normalize input
X_scaled = (raw_scan - data["scaler_mean"]) / data["scaler_scale"]
# Predict
concentration_ppm = X_scaled @ data["coef"] + data["intercept"][0]

CNN (PyTorch)

import torch
import torch.nn as nn
import numpy as np

norm = np.load("cnn1d-t1-da/normalization.npz")
state = torch.load("cnn1d-t1-da/model.pt", weights_only=True)

model = nn.Sequential(
    nn.Conv1d(1, 16, 15, stride=2, padding=7), nn.ReLU(),
    nn.Conv1d(16, 32, 9, stride=2, padding=4), nn.ReLU(),
    nn.Conv1d(32, 64, 5, stride=2, padding=2), nn.ReLU(),
    nn.AdaptiveAvgPool1d(8), nn.Flatten(),
    nn.Linear(64 * 8, 64), nn.ReLU(), nn.Linear(64, 1),
)
model.load_state_dict(state)

# Normalize and predict
x = torch.tensor((raw_scan - norm["input_mean"]) / norm["input_std"],
                  dtype=torch.float32).unsqueeze(0).unsqueeze(0)
with torch.no_grad():
    z = model(x).item()
concentration_ppm = float(np.expm1(z * norm["target_std"][0] + norm["target_mean"][0]))

Training

All models were trained on SPEKTRAN v0.5.0 official splits:

pip install spektran[dev] scikit-learn torch
python scripts/train_and_export_baselines.py --out checkpoints/

See baselines/README.md for full reproduction instructions and benchmark results.

Citation

@software{spektran2026,
  title = {SPEKTRAN: Synthetic ML Training Data for Gas Sensing},
  url = {https://github.com/spektran/spektran},
  doi = {10.5281/zenodo.21790394},
  version = {0.5.0},
  year = {2026},
}