Machine Learning and Optimal Experimental Design for Thermodynamic Property Modeling


Start: 2025-03-01
End: 2028-02-29
Principal Investigators: Roland Herzog, Markus Richter
Staff: Viktor Martinek, Ophelia Frotscher
Funded by: DFG within the Priority Program funding scheme
Part of: Machine Learning in Chemical Engineering (SPP 2331)

Project Description

For many tasks in chemical and energy engineering, the accurate knowledge of thermodynamic properties (e.g., pressure and temperature with density and speed of sound) and the phase behavior of the involved fluids play a key role. Such properties are required for the fundamental understanding of chemical-physical behavior and for the development of predictive models. Thermodynamic properties are also the basis for the design of safe, sustainable and energy efficient processes and machinery. However, the quality of property predictions depends largely on the availability and accuracy of experimental or simulated data, as well as on the modeling techniques used. Thermodynamic property measurements are typically the most accurate data source and are often carried out on a dense grid of measurement points, which delivers a seemingly comprehensive data set. With the goal to develop an accurate thermodynamic property model, this approach is time-consuming, while it is unclear whether all data contribute substantially to the model development. As a result, although fundamental knowledge can be generated, the required time and financial expenditure make the generation of reliable models rather limited. It is thus highly desirable to significantly reduce the model development time by limiting the amount of acquired data, whether experimental or simulated, to the required extent. Additionally, models developed should exhibit concise functional forms to allow for efficient evaluation and facilitate their use in process simulations. Therefore, the major goal of the proposed project is to tackle the aforementioned issues using a synergistic combination of (1) interpretable machine learning in the form of symbolic regression, to find better functional forms to model thermodynamic properties, (2) optimal experimental design, to find the most appropriate data points, and (3) the actual data acquisition. During the first phase of the priority program, a workflow was applied, where, starting from initial thermodynamic property data, machine learning-based equation-of-state modeling is used to create a first functional form. This form is used to identify the next most informative measurements, which can then be used as input for further equation-of-state modeling. In the second project phase this workflow is applied to investigate mixtures. To this end, liquid-liquid equilibria are investigated, and application-specific models are created for use in process designs. In addition, symbolic regression will be used to develop physics-based fundamental equations of state for pure substances and particularly mixtures.

Associated Software

  • Viktor Martinek
    Fast symbolic regression benchmarking, 2025
    doi:10.5281/zenodo.15469873
    https://github.com/viktmar/FastSRB/
    bibtex 
    @SOFTWARE{Martinek:2025:1,
  AUTHOR = {Martinek, Viktor},
  URL = {https://github.com/viktmar/FastSRB/},
  DATE = {2025},
  DOI = {10.5281/zenodo.15469873},
  TITLE = {Fast symbolic regression benchmarking},
}
  • Viktor Martinek
    Thermodynamics-informed symbolic regression (TiSR). A tool for the thermodynamic equation of state development, 2023
    doi:10.5281/zenodo.8317547
    https://github.com/scoop-group/TiSR/
    bibtex 
    @SOFTWARE{Martinek:2023:1,
  AUTHOR = {Martinek, Viktor},
  URL = {https://github.com/scoop-group/TiSR/},
  DATE = {2023},
  DOI = {10.5281/zenodo.8317547},
  TITLE = {Thermodynamics-informed symbolic regression (TiSR). A tool for the thermodynamic equation of state development},
}

Agency logo