McrNmf

McrNmf is a Python package for Multivariate Curve Resolution (MCR) using a range of Nonnegative Matrix Factorization (NMF) algorithms. MCR is widely used in chemometrics, for instance, to decompose mixture spectra into their pure component spectra and associated concentration profiles.

Key Features

• Multiple solvers – unified interface for three NMF variants: classic Alternating Least Squares (FroALS), Fast Projected Gradient (FroFPGM), and Minimum-Volume formulation (MinVol).
• Built-in constraints – supports closure, normalization, equality, and per-component unimodality constraints.
• Robust initialisation – includes Successive Nonnegative Projection (SNPA) algorithm for generating reliable starting estimates of spectra and concentrations
• Lightweight implementation – written almost entirely in NumPy, with Numba used only to speed up unimodal regression.

For more details, see the documentation.

Installation

With pip:

pip install mcrnmf

Note for macOS Apple Silicon (M4) users: When running the McrNmf package installed via pip, you may encounter runtime numerical warnings (e.g., divide/overflow). To avoid these, please install via conda-forge.

With conda:

conda install -c conda-forge mcrnmf

mcrnmf requires:

• Python ≥ 3.11.9, < 3.14
• NumPy ≥ 1.24, < 2.3
• Numba ≥ 0.61.2

Quick Start

from mcrnmf.datasets import load_rxn_spectra
from mcrnmf import SNPA, MinVol

# load a Raman spectroscopy data (wv × time)
X, wv, t = load_rxn_spectra()

# specify the number of components
rank = 4

# generate initial guess for W and H using SNPA
snpa = SNPA(rank=rank, iter_max=1000)
snpa.fit(X)
Wi = snpa.W  # intial guess for W (pure spectra)
Hi = snpa.H  # intial guess for H (conc. profiles)


mvol = MinVol(rank=rank,
              constraint_kind=1,
              unimodality={"H": True},
              iter_max=2000,
              tol=1e-4)
# fit the Minimum Volume NMF model
mvol.fit(X=X, Wi=Wi, Hi=Hi)
# access the decomposed factors
W = mvol.W
H = mvol.H
# access relative reconstruction error at each iter.
rel_recon_err = mvol.rel_reconstruction_error_ls

More worked examples can be found in the Usage section of the documentation.

Citation

If you use McrNmf in your work, please consider citing the following preprint:

Vasudevan SA, Kohlbrenner R, Ritschel N, Vent-Schmidt T. Beyond Classical MCR-ALS—Minimum Volume NMF and Strategic Sampling for Robust Reaction Monitoring. ChemRxiv. 2025; doi:10.26434/chemrxiv-2025-qs5g4

License

MIT – see LICENSE.