Fingerprinting and thermodynamic modeling of poorly specified mixtures with NMR spectroscopy and machine learning
Final Report Abstract
Mixtures of partially unknown composition (poorly specified mixtures) are ubiquitous in many fields of science and engineering, including chemistry, biotechnology, and environmental science. Such mixtures constitute a significant challenge for conceptual process design and optimization since classical thermodynamic models cannot be applied in this situation as they require complete information on the speciation. In this project, a novel framework for addressing this challenge was developed, which consists of three steps combining NMR spectroscopy, machine-learning (ML) algorithms, and thermodynamic models. In the first step, called "NMR fingerprinting", information on the structural groups as the building blocks of components are derived from a set of standard NMR spectra of the unknown sample in an automated way. For this purpose, a support vector classification (SVC) algorithm from ML was developed and trained on the NMR spectra of almost 3000 pure components. In the second step, the structural groups identified and quantified by NMR fingerprinting are used to define pseudo-components in the unknown mixture. For this purpose, k-medians clustering from ML was employed, which was carried out based on self-diffusion coefficients measured by pulsed-field gradient NMR spectroscopy. In the third and final step, the obtained mixture composition in terms of pseudocomponents is used as input for thermodynamic group-contribution methods for modeling the properties of the unknown mixture. The newly developed framework was validated and tested using unknown pure components and poorly specified mixtures inspired by fermentation broths. The automatically obtained characterizations were systematically compared to the ground truth, and excellent results were obtained. Furthermore, the obtained characterizations were used to simulate two thermal separation processes with poorly specified feeds: liquidliquid extraction and single-stage batch distillation. Again, the results were compared to those obtained when using the full speciation of the feed, and an excellent agreement was found. The developed framework was made available to the scientific community, among others, in an interactive website and will be extended and tailored to specific situations in follow-up projects.
Publications
-
Fingerprinting and thermodynamic modeling of poorly specified mixtures with NMR spectroscopy and machine learning, Symposium on Thermophysical Properties, online, 20.-25.06.2021 (Talk)
F. Jirasek, T. Specht & H. Hasse
-
NEAT 2.0 – Thermodynamic modeling of poorly specified mixtures with NMR spectroscopy and machine learning, Thermodynamik-Kolloquium 2021, online, 27.-29.09.2021 (Talk)
T. Specht, K. Münnemann, F. Jirasek & H. Hasse
-
Quantitative fingerprinting and thermodynamic modeling of poorly specified mixtures with NMR spectroscopy and machine learning, European Symposium of Applied Thermodynamics, online, 05.-09.07.2021 (Talk)
T. Specht, K. Münnemann, F. Jirasek & H. Hasse
-
The NEAT toolbox for thermodynamic modeling of poorly specified mixtures, Thermodynamik-Kolloquium, Chemnitz, 26.- 28.09.2022 (Poster)
T. Specht, K. Münnemann, H. Hasse & F. Jirasek
-
Automated NMR fingerprinting and thermodynamic modeling of poorly specified mixtures, Thermodynamik-Kolloquium, Hannover, 25.-27.09.2023 (Poster)
J. Arweiler, T. Specht, K. Münnemann, H. Hasse & F. Jirasek
-
Automated nuclear magnetic resonance fingerprinting of mixtures. Magnetic Resonance in Chemistry, 62(4), 286-297.
Specht, Thomas; Arweiler, Justus; Stüber, Johannes; Münnemann, Kerstin; Hasse, Hans & Jirasek, Fabian
-
Modeling Unknown Mixtures, Mathematical Methods in Process Engineering (MMiPE), Kaiserslautern, 05.-06.10.2023.
F. Jirasek
-
Predictive Thermodynamic Modeling of Poorly Specified Mixtures and Applications in Conceptual Fluid Separation Process Design. Industrial & Engineering Chemistry Research, 62(27), 10657-10667.
Specht, Thomas; Hasse, Hans & Jirasek, Fabian
-
Rational method for defining and quantifying pseudo-components based on NMR spectroscopy. Physical Chemistry Chemical Physics, 25(15), 10288-10300.
Specht, Thomas; Münnemann, Kerstin; Hasse, Hans & Jirasek, Fabian
-
SMART NMR fingerprinting of mixtures using benchtop spectrometers, Quantitative NMR Methods for Reaction and Process Monitoring (NMRPM), Kaiserslautern, 24.-26.05.2023 (Poster)
J. Arweiler, T. Specht, K. Münnemann, H. Hasse & F. Jirasek
-
Thermodynamic modeling of unknown mixtures with NMR spectroscopy and machine learning, Quantitative NMR Methods for Reaction and Process Monitoring (NMRPM), Kaiserslautern, 24.-26.05.2023 (Talk)
F. Jirasek
