Non-negative Matrix Factorization for Time-Resolved Raman Spectroscopy Data

PID (if applicable): doi:10.1007/s10910-020-01201-7

Problem Statement

Crystallization of Paracetamol in Ethanol

Object of Research and Objective

Determination of Intermediate States and their Kinetics along the Crystallization of Paracetamol in Ethanol using Raman Spectroscopy.

Procedure

Workflow Illustration

1. Data Acquisition

Time-resolved Raman Spectroscopy to follow the crystallization of Paracetamol in accoustically-levitated Ethanol droplets. Surface, temperature (22.0 +/- 1.0 °C), and relative humidity (17.5 +/- 2.5 %) of the environment are controlled by Nitrogen stream.

2. Data Extraction

Extract the spectroscopic measurement matrix $\mathbf {M} \in \mathbb {R} _{+}^{n\times m}$ ($m$ measurements of $n$ non-negative intensities) from measurement file containing spectroscopic data and metadata.

3. Data Analysis

Factorize $\mathbf {M}$ such that $\mathbf {M} =\mathbf {W} \cdot \mathbf {H}$ ($\mathbf {W} \in \mathbb {R} _{+}^{n\times r},\mathbf {H} \in \mathbb {R} _{+}^{r\times m}$ with $r$ the rank of factorization or expected number of components) using a novel non-negative matrix factorization (NMF) approach:

Pre-proceesing

Singular Value Decomposition (SVD) for primary factorization $\mathbf {M^{T}} =$ U $\mathbf {\Sigma } \mathbf {V^{T}}$.

Generate $\mathbf {U}$ from U with $\mathbf {U} [:,0]=[1,1,...,1,1]^{T}$ and $\mathbf {U} [:,1:]=$ U $[:,:r-1]$.

Ensure orthogonality among columns of $\mathbf {U}$.

Initializing

Apply PCCA+ to $\mathbf {U}$ to obtain the transformation matrix $\mathbf {A}$ to initialize $\mathbf {H}$, $\mathbf {W}$, and $\mathbf {P}$.

${\widetilde {\mathbf {H} }}=(\mathbf {UA} )^{T}$

${\widetilde {\mathbf {W} }}=\mathbf {M} (\mathbf {A^{T}U^{T}} )^{\#}$ ($\#$ the pseudoinverse of singular / non-square matrices)

${\widetilde {\mathbf {P} }}=\mathbf {A^{-1}} (\mathbf {U_{-}^{\#}U_{+}} )\mathbf {A}$ ($\mathbf {U_{+/-}}$ defined as $\mathbf {U}$ without the first / last row)

Minimizing

Objective function $\Psi ^{2}$ maintains positivity (1,2,4), column (3) and row stochastics (5), depending on $\mathbf {M}$, $\mathbf {U}$, and $\mathbf {A}$.

Minimization of $\Psi ^{2}$ with respect to $\mathbf {A}$ adjusts ${\widetilde {\mathbf {W} }}$ and ${\widetilde {\mathbf {H} }}$ numerically to the claimed structural properties.

$\Psi =\alpha ({\underset {i,j}{\min }}{\widetilde {W_{i,j}}})+\beta ({\underset {i,j}{\min }}{\widetilde {H_{i,j}}})+\gamma ({\underset {j}{\max }}|\sum \limits _{i=1}^{r}{\widetilde {H_{i,j}}}-1|)+\delta ({\underset {i,j}{\min }}{\widetilde {P_{i,j}}})+\mu ({\underset {j}{\max }}|\sum \limits _{j=1}^{r}{\widetilde {P_{i,j}}}-1|)$

Recovering

Minimization returns $\mathbf {A_{opt}}$, which allows to recover $\mathbf {H_{rec}}$, $\mathbf {W_{rec}}$, and $\mathbf {P_{rec}}$.

$\mathbf {H_{rec}} =(\mathbf {UA_{opt}} )^{T}$

$\mathbf {W_{rec}} =\mathbf {M} (\mathbf {A_{opt}^{T}{U}^{T}} )^{\#}$

$\mathbf {P_{rec}} =\mathbf {A_{opt}^{-1}} (\mathbf {U_{-}^{\#}U_{+}} )\mathbf {A_{opt}}$

4. Data Interpretation

$\mathbf {W_{rec}}$ contains the spectra of the substances involved in the crystallization process, while $\mathbf {H_{rec}}$ allows an inference on the kinetics. Interpretation of both matrices leads to the identification of intermediate states and the underlying kinetics.

Involved Disciplines

NFDI4Chem (wikidata:Q96678459)

    $\Rightarrow$ Raman Spectroscopy (wikidata:Q862228)

MaRDI (wikidata:Q108327788)

    $\Rightarrow$ Numerical Analysis (wikidata:Q11216)

    $\Rightarrow$ Mathematical Optimization (wikidata:Q141495)

    $\Rightarrow$ Linear Algebra (wikidata:Q82571)

Data Streams

NFDI4Chem $\Rightarrow$ MaRDI (.txt File containing measurement matrix $\mathbf {M}$)

MaRDI $\Rightarrow$ NFDI4Chem (.png Files containing component spectra $\mathbf {W}$ and relative concentration profiles $\mathbf {H}$)

Model

$\mathbf {M} =\mathbf {WH}$

A measurement matrix $\mathbf {M}$ is factorized in a matrix $\mathbf {W}$, containing the spectra of the substances involved, and a matrix $\mathbf {H}$, which allows an inference on the kinetics.

Discretization

• Time: Time-resolution of Raman Spectroscopy
• Space: -

Variables

Name	Unit	Symbol	dependent (measured) / independent (controlled)
Time	s	t	independent
Wavelength	cm-1	$\lambda$	independent
Intensity	-	I	dependent (measured)
Substance Matrix	-	$\mathbf {W}$	dependent (calculated)
Kinetic Matrix	-	$\mathbf {H}$	dependent (calculated)

Parameter

Name	Unit	Symbol
Temperature	°C	T
Relative Humidity	%	RH
Rank of Factorization	-	$r$
Number of Singular Values	-	$k$
Singular Value Tolerance	-	$tol$
Objective Function Parameter	-	$\alpha ,\beta ,\gamma ,\delta ,\mu$
Maximum Number of Iterations	-	$MAXITER$

Process Informationen

Process Steps

Name	Description	Input	Output	Method	Parameter	Environment	Mathematical Area
Data Acquisition	Measurement	-	.icraman	Time-resolved Raman Spectroscopy	T, RH	RXN1™	-
Data Extraction	Extract Spectroscopic Data	.icraman	.txt	-	-	iC Raman™	-
Data Analysis	Determine Component Spectra & Concentration Profiles	.txt	.png	NMF algorithm	$r,tol,k$ $MAXITER$ $\alpha ,\beta ,\gamma ,\delta ,\mu$	Matlab	Numerical Analysis, Linear Algebra, Mathematical Optimization
Data Interpretation	Determine Intermediate States & Kinetics	.png	-	-	-	-	-

Applied Methods

ID	Name	Process Step	Parameter	realised/implemented by
doi:10.1007/s10910-020-01201-7	NMF Algorithm	Data Analysis	$r,tol,k$ $MAXITER$ $\alpha ,\beta ,\gamma ,\delta ,\mu$	Matlab Script
wikidata:Q420904	SVD	Data Analysis - Pre-Processing	$k$	Matlab R2019a
wikidata:Q43219517	Pseudoinverse	Data Analysis - Initializing	$tol$	Matlab R2019a
doi:10.1016/j.laa.2004.10.026	PCCA+	Data Analysis - Initializing	$r$	Matlab Script
wikidata:Q1253278	Nelder-Mead Algorithm	Data Analysis - Minimizing	$\alpha ,\beta ,\gamma ,\delta ,\mu ,MAXITER$	Matlab R2019a
wikidata:Q43219517	Pseudoinverse	Data Analysis - Recovering	$tol$	Matlab R2019a

Software used

ID	Name	Description	Version	Programming Language	Dependencies	versioned	published	documented
-	iC Raman	Data Acquisition & Reaction Analysis	4.1	?	Windows	Yes	Yes	Yes
sw:558	Matlab	Programming and numeric Computing	R2019a	C,C++,Fortran,Java	Windows, Mac, Linux	Yes	Yes	Yes

Experimental Devices/Instruments and Computer-Hardware

ID	Name	Description	Version	Part Nr	Serial Nr	Location	Software
-	Raman RXN1	Spectrometer
-	GenunineIntel	Intel(R) Core(TM) i7-9700T CPU @ 2.00 GHz

Input Data

ID	Name	Size	Data Structure	Format Representation	Format Exchange	binary/text	proprietary	to publish	to archive
-	Spectroscopic & Meta Data	small	-	-	.icraman	binary	Yes	Yes	Yes
-	Spectroscopic Data ( $\mathbf {M}$)	small	Matlab Array	dense matrix (csv)	.txt	text	No	Yes	Yes
-	$r$	small	integer			text	No	Yes	Yes
-	$\alpha ,\beta ,\gamma ,\delta ,\mu$	small	float			text	No	Yes	Yes

Output Data

ID	Name	Size	Data Structure	Format Representation	Format Exchange	binary/text	proprietary	to publish	to archive
-	Component Spectra ($\mathbf {W}$)	small	Matlab Array	plot	.png	binary	No	Yes	Yes
-	Concentration Profile ($\mathbf {H}$)	small	Matlab Array	plot	.png	binary	No	Yes	Yes

Reproducibility

Reproducibility of the Experiments on the original Devices/Instruments/Hardware

Yes

Reproducibility of the Experiments on other Devices/Instruments/Hardware

Yes

Transferability of the Experiments to

a) other solutes, solvents, parameter

b) other chemical reactions

Legend

The following abbreviations are used in the document to indicate/resolve IDs:

doi: https://dx.doi.org/

sw: https://swmath.org/software/

wikidata: https://www.wikidata.org/wiki/