RADAR

Raman spectral Analysis using Deep learning for Artifact Removal.

RADAR splits Raman spectra into four additive components with convolutional neural networks:

spectrum = baseline + cosmic rays + noise + peaks

• Baseline: smooth background (fluorescence, etc.)
• Cosmic rays: narrow spikes
• Noise: random noise
• Peaks: the Raman bands, i.e. the corrected, denoised spectrum

The package contains:

• Two pre-trained networks, standard and ensemble, each trained on 240 million synthetic spectra.
• Training and evaluation tools: a synthetic data generator for training and testing, training and fine-tuning routines, and evaluation metrics (RMSE, MAE, max error, R²).
• File I/O for the common Raman formats: SPC, JCAMP-DX, Renishaw WDF, Horiba LabSpec, TriVista, text exports, Excel, Parquet, SQLite, HDF5, MATLAB, JSON and NumPy.
• Interfaces: a command-line tool.

Installation

cd RADAR
pip install .                 # core
pip install ".[formats]"      # + Excel / Parquet / Feather
pip install ".[renishaw]"     # + Renishaw .wdf
pip install -e ".[dev]"       # everything, for development

Requires Python ≥ 3.9 and TensorFlow ≥ 2.16 (Keras 3). Check the installation with radar-raman info.

Quick start

import radar_raman
comps = radar_raman.decompose("spectra.spc")   # file (any supported format) or NumPy array
comps.baseline, comps.cosmic_rays, comps.noise, comps.peaks
comps.plot(0)
comps.save("spectra_components.h5")            # .csv .xlsx .parquet .db .mat .json .spc .jdx ...

The API in eight functions

radar_raman.decompose(spectra, model="ensemble", output=None)  # array / file -> Components
radar_raman.read(path)                                   # any format -> SpectralData
radar_raman.write(obj, path)                             # spectra, components or data sets
radar_raman.formats(show=True)                           # supported formats
radar_raman.load_model("ensemble")                       # pre-trained or saved model
radar_raman.train(data=None, model="ensemble", init=None, save_to="my.keras", epochs=..., ...)
radar_raman.generate(1000)                               # synthetic spectra with ground truth
radar_raman.evaluate(model, test_set)                    # RMSE, MAE, max error, R²

Saved models record the RADAR, Keras and TensorFlow versions they were made with: radar_raman.model_info("my.keras", show=True) or radar-raman info my.keras.

Supported file formats

Instrument text exports from Renishaw WiRE, Horiba LabSpec, Ocean Insight, WITec and RRUFF are recognised automatically. Raman maps keep their grid shape (comps.as_map("peaks"), comps.band_image(990, 1010)).

Command line

radar-raman decompose sample.wdf -o sample_components.h5
radar-raman decompose data/*.spc -o "results/{stem}.csv" -m standard
radar-raman convert measurement.spc measurement.jdx
radar-raman generate -n 1000 -o test.h5
radar-raman train -o my_model.keras --epochs 12 --steps 200000
radar-raman evaluate test.h5 -m my_model.keras --statistics
radar-raman formats

Documentation

To build the HTML documentation: pip install ".[docs]" && mkdocs serve.

Examples

Tests

pip install -e ".[test,formats,renishaw]"
pytest                    # full suite
pytest -m "not slow"      # skip the training tests

The suite checks that the networks reproduce the outputs of the original implementation (to within 1e-5), and that every writable format round-trips. Optional tests on real instrument files run when RADAR_SAMPLE_DATA points to the public test data of the SPC, JCAMP-DX and rosettasciio projects (see tests/test_formats.py).

How to cite

If you use RADAR, please cite the software:

Sjöberg, J., & Oancea, B. (2026). RADAR: Raman spectral Analysis using Deep learning for Artifact Removal (Python package) (Version 0.1.0) [Computer software]. Zenodo. https://doi.org/10.5281/zenodo.23760379

radar_raman.citation() prints the references, and CITATION.cff has them in machine-readable form (GitHub's "Cite this repository" button uses it).

Licence and credits

BSD 3-Clause, see LICENSE. The networks, the synthetic data generator and the pre-trained weights come from the original implementation by Joel Sjöberg (repository). The Python package is by Bogdan Oancea. Both copyright notices must be kept in redistributions. See AUTHORS.md.

Optional readers: renishawWiRE (MIT) and rosettasciio (GPL-3.0) are separate packages that RADAR uses only if you install them.