Matrix and Tensor Completion Methods for Predicting Mixture Properties

Reliable mixture data are essential for the modeling, design, and optimization of processes in industry and nature. However, since experimental data are available only for a small fraction of the relevant systems, predictive methods are needed to enable reliable predictions even for mixtures that have not yet been studied. In this project, matrix and tensor completion methods are used to systematically fill these gaps. The basic idea is to organize data for binary mixtures into matrices or tensors, whose rows and columns correspond to the substances and conditions involved, and whose entries correspond to the properties of interest. Missing entries can thus be predicted in a data-driven manner, with prior physical knowledge from established physical models being specifically integrated into the machine-learning models.

One area of focus is the prediction of diffusion coefficients at infinite dilution in binary liquid mixtures. These transport properties are relevant, among other things, for describing mass transfer, reaction, and separation processes; however, experimental data on them are severely lacking. Tensor completion methods allow for the inclusion of temperature dependence, enabling the prediction of diffusion coefficients not only at individual temperatures but also over an extended temperature range. Furthermore, matrix completion methods are applied to electrolyte solutions by predicting parameters of established activity coefficient models for previously unstudied cation-anion combinations. This significantly expands the applicability of physically interpretable electrolyte models. A current application context is phosphorus recovery from wastewater, where the activities of ionic species in aqueous solution play a decisive role in solubility, precipitation, and process control.

Active Learning for Efficient Model Improvement with Targeted New Measurements

The performance of data-driven prediction methods depends largely on the quality and distribution of the available training data. Since new measurements are time- and ressource-intensive, it is not sufficient to simply expand the experimental data at random. Instead, we develop and apply active learning strategies to specifically identify those material systems whose measurement is expected to yield the greatest information gain for the model. This ensures that experimental resources are used efficiently and that model accuracy can be improved with as few additional measurement points as possible.

In the context of matrix and tensor completion, active learning strategies are used in particular to plan new diffusion measurements. Selected solute-solvent systems are experimentally investigated using pulsed-field-gradient NMR spectroscopy and subsequently used for further model refinement. This approach is currently being extended to electrolyte systems. The resulting integration of data-driven modeling, prior physical knowledge, and adaptive experimental planning enables the targeted expansion of chemical property databases. In the long term, this approach aims to contribute to making predictive models for thermodynamic and transport-related material properties robust and data-efficient.

Teaching Assistance

• ChemCar-Wettbewerb I (WS 2022/2023)
• Wärmeübertragung (WS 2023/24)
• Grundlagenlabor für das Schwerpunktfach I (SS 2024)
• Elektrolytthermodynamik (WS 2024/25)
• Maschinelles Lernen in der Verfahrenstechnik (SS 2025)
• Thermodynamik der Transportprozesse (WS 2025/26)

Appointments

Inquire by E-Mail

Publications

• Z. Romero, K. Münnemann, H. Hasse, F. Jirasek: Improvement of Diffusion Coefficient Prediction by Active Learning, J. Phys. Chem. B 2025, 129, 36, 9219–9228. [doi]
• Z. Romero, K. Münnemann, H. Hasse, F. Jirasek: Prediction of Diffusion Coefficients in Mixtures with Tensor Completion, Phys. Chem. Chem. Phys. 2026. [doi] [xiv]

• Z. Romero, M. Kohns, F. Jirasek: Predicting Activities in Aqueous Electrolyte Solutions with Hybrid Machine Learning, arXiv, 2607.19114. [xiv]

   

Presentations

• S. Mross, Z. Romero, K. Münnemann, F. Jirasek, H. Hasse: Measurement and Prediction of Self-Diffusion Coefficients with NMR Spectroscopy and Active Learning Strategies, Thermodynamik-Kolloquium, Hannover, 25.09.-27.09.2023.
• Z. Romero, K. Münnemann, F. Jirasek, H. Hasse: Improvement of Diffusion Coefficient Prediction by Active Learning, 33^rd European Symposium on Applied Thermodynamics, Edinburgh, UK, 09.06.-12.06.2024.
• Z. Romero, K. Münnemann, F. Jirasek, H. Hasse: Improvement of Diffusion Coefficient Prediction by Active Learning, 20^th International Conference on Diffusion in Solids and Liquids (DSL), Barcelona, Spain, 24.06.-28.06.2024.
• Z. Romero, M. Kohns, H. Hasse, F. Jirasek: A Hybrid Model for Predicting Activities in Aqueous Electrolyte Solutions, 23^rd European Conference on Thermophysical Properties (ECTP), Gouvieux, France, 21.-24.06.2026.
• Z. Romero, N. Hayer, H. Hasse, F. Jirasek: Probabilistic Matrix Completion for Chemical Engineering Applications, StanCon 2026, Uppsala, Sweden, 17.-21.08.2026.
• Z. Romero, M. Kohns, H. Hasse, F. Jirasek: Advanced Matrix and Tensor Completion for Predicting Thermophysical Properties of Mixtures, Thermodynamik-Kolloquium, Dortmund, 30.09.-02.10.2026.

Posters

• Z. Romero, K. Münnemann, F. Jirasek, H. Hasse: Prediction of Diffusion Coefficients in Mixtures with Tensor Completion: European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases ECML PKDD, Vilnius, Lithuania, 9. September 2024 (Poster-Präsentation).
• Z. Romero, K. Münnemann, F. Jirasek, H. Hasse: Prediction of Diffusion Coefficients in Mixtures with Tensor Completion: Thermodynamik-Kolloquium, Stuttgart, 25. September 2024 (Poster-Präsentation).
• Z. Romero, I. Bezverkhyy, J.-M. Simon, H. Hasse, F. Jirasek, Prediction of Gas Adsorption Isotherms with Machine Learning, Thermodynamik-Kolloquium, Dortmund, 30.09.-02.10.2026.