Carl-Zeiss-Foudation-Juniorprofessorship Machine Learning in Process Engineering
In the couse of "digitalization", process engineering is facing novel opportunities and challenges. One of the key challenges of "digitalized" process engineering will be to integrate and establish Machine Learning (ML) methods in a beneficial way. Application fields for this purpose are manifold, e.g., for predicting fluid properties or process simulations.
To date, ML methods are successul in particular in fields in which huge amounts of data are available. Process engineering significantly differs from classical application fields of ML methods in that respect. As an example, an algorithm for text and image recognition can be trained using a basically unlimited number of data points, whereas is process engineering, the generation of each data point is usually very expensive, e.g., the experimental measurement of a fluid property of a component or mixture. On the other hand, process engineering can - in addtion to available data - rely on rich physical knowledge and experience. Consequently, the following questions arise:
- How can ML methods be applied to incomplete and heterogenous data sets as they are common in process engineering?
- How can available physical knowledge be integrated in a-priori purely data-driven ML methods in a suitable way to obtain hybrid approaches?
- How reliable are predictions of (hybrid) ML methods? Can they be deployed for safety-related issues?
- If new data points are required: which experiments should be carried out to obtain the biggest improvement of a given method?
Carl-Zeiss-Foundation Junior Professor of Machine Learning in Process Engineering
JP Dr.-Ing. Fabian Jirasek
- Prediction of fluid properties wit Machine Learning (ML)
- Development of hybrid approaches: integration of physical knowledge in ML methods
- Interpretability of ML methods in process engineering
- Design of Experiment and Active Learning
- Automation of chemical plants
Prediction of Fluid Properties with Matrix Completion Methods
As the experimental determination of fluid properties is elaborate and expensive, prediction methods are indispensable in process engineering. The properties of binary mixtures can naturally be represented in matrix form, where rows and columns constitute the mixture components; the resulting matrix is only partially observed in basically all cases. The prediction of properties of not yet studied mixtures can be considered as matrix completion problem. This problem is well known in Machine Learning and several approaches to predict the unobserved entries based only on the observed entries exist. At the LTD, these approaches are applied to predict different fluid properties, such as activity coefficients, Henry's law constants, and diffusion coefficients.
Development of Hybrid Fluid Property Models: Machine Learning meets Physical Knowledge
Machine Learning methods are a-priori purely data-driven. Hence, they ignore the explicit knowledge of physical relationsships, which is available in many cases in process engineering. At the LTD, we explore how Machine Learning methods can be augmented with explicit phyisical knowledge to develop novel hybrid approaches. As an example, we consider the integration of pure component descriptors and physical models in matrix completion methods, yielding hybrid predictive fluid property models.
Elucidation of Structural Groups in Mixtures with NMR Spectroscopy and Support Vector Classification
NMR spectroscopy is a powerful tool for structure elucidation of components and quantitative analysis of mixtures. The evaluation of NMR spectra is, however, oftentimes elaborate and requires great personal experience. At the LTD, we study how Support Vector Classification can be used to facilitate or even automate the evaluation of NMR spectra. We thereby develop methods to identify and quantify relevant structural groups in complex mixtures based on simple one-dimensional NMR experiments.
Process Design and Simulation with Poorly Specified Mixtures
Mixtures of which the composition is only partially known, i.e., poorly specified mixtures, are common in process engineering, e.g., fermentation broths in biotechnology. Such mixtures are especially challenging for process engineers as their properties cannot be described by classical thermodynamic models. At the LTD, we develop methods to characterize poorly specified mixtures by "NMR Fingerprinting", which we then apply for process design and optimization.