2026-08-14 –, Room 4
We present MLThermoProperties.jl, a Julia package that provides a variety of state-of-the-art thermodynamic models that combine modern machine learning methods with physical knowledge. These hybrid models obey hard physical constraints while being more accurate and applicable to a wider scope of substances than established models. MLThermoProperties.jl is built upon the Clapeyron.jl package, leveraging its rich thermodynamic solver ecosystem. The MLThermoProperties.jl models significantly improve molecular property prediction in various applications in science and engineering, e.g., chemical process engineering. Exemplary applications will be demonstrated in the talk by coupling MLThermoProperties.jl with Julia's rich ecosystem for scientific modelling and simulation.
Knowledge of thermodynamic properties of fluids is crucial in many fields of engineering and science, e.g., for developing new chemical and biotechnological processes, optimizing heat pumps, or designing carbon capture and storage technologies. However, experimental data on thermodynamic properties are notoriously scarce due to the high cost and effort of measurements, making reliable prediction models indispensable. In recent years, machine learning (ML) has emerged as a particularly promising approach to thermodynamic modeling [1].
In our research group, many state-of-the-art thermodynamic ML models are developed under the umbrella of MLPROP, a collection of open-source ML models for molecular property prediction. These models, despite being neural networks at their core, are thermodynamically consistent, i.e., they obey hard physical constraints. This consistency is either achieved by an appropriate architecture or by exploiting existing thermodynamic models to form new hybrid models. Based only on the SMILES code of a substance – a textual representation of the molecular structure – the MLPROP models predict multiple thermodynamic properties, outperforming established models in both accuracy and scope [2]. Among the covered properties are phase equilibria of pure substances and mixtures [2], which are central to chemical process design, as well as transport properties such as diffusion coefficients [3] that govern molecular mass transfer. The models utilize different ML architectures, including the chemical language model ChemBERTa [4] and molecular graph neural networks.
In Julia, a rich ecosystem for thermodynamics and chemical engineering already exists. Clapeyron.jl [5], a mature and comprehensive thermodynamic package, is a central part of this ecosystem. It combines a large number of thermodynamic models (including equations of state (EoS) and Gibbs excess energy models) with advanced and efficient solvers. Leveraging Julia’s excellent extensibility, several packages extend Clapeyron.jl, e.g., EntropyScaling.jl for modeling transport properties, or Langmuir.jl for modeling adsorption.
This talk introduces MLThermoProperties.jl, which provides Julia implementations of the thermodynamic models from MLPROP, based on Clapeyron.jl – enabling the prediction of thermodynamic properties for any substance in Julia. MLThermoProperties.jl not only provides implementations of existing models, but also serves as a central anchor point for the development of new models. Due to the excellent extensibility of Julia packages, MLThermoProperties.jl substantially broadens the
capabilities for modeling chemical processes, based on state-of-the-art molecular property prediction. The individual models are presented and their integration into the existing thermodynamic ecosystem is explained. Additionally, illustrative applications of these models in chemical process simulations are showcased using packages from the wider scientific-modeling ecosystem in Julia, particularly from the SciML organization, e.g., ModelingToolkit.jl and DifferentialEquations.jl.
References
[1] H. Hasse, S. Schmitt, and F. Jirasek: Artificial Intelligence in Thermodynamics: Hybrid Modeling of Thermophysical Properties of Fluids, Current Opinion in Chemical Engineering 51 (2026) 101236, DOI: https://doi.org/10.1016/j.coche.2026.101236.
[2] T. Specht, M. Nagda, S. Fellenz, S. Mandt, H. Hasse, and F. Jirasek: HANNA: Hard-Constraint Neural Network for Consistent Activity Coefficient Prediction, Chemical Science (2024), DOI: https://doi.org/10.1039/D4SC05115G.
[3] J. Wagner, Z. Romero, K. Münnemann, S. Schmitt, T. Specht, H. Hasse, and F. Jirasek: Hybrid Machine Learning for Enhanced Prediction of Diffusion Coefficients in Liquids, to be published (2026).
[4] W. Ahmad, E. Simon, S. Chithrananda, G. Grand, and B. Ramsundar: ChemBERTa-2: Towards Chemical Foundation Models, arXiv, DOI: https://doi.org/10.48550/arXiv.2209.01712.
[5] P. J. Walker, H.-W. Yew, and A. Riedemann: Clapeyron.jl: An Extensible, Open-Source Fluid Thermodynamics Toolkit, Industrial & Engineering Chemistry Research 61 (2022) 7130–7153, DOI: https://doi.org/10.1021/acs.iecr.2c00326.
I'm a postdoctoral researcher at Laboratory of Engineering Thermodynamics (LTD) of the University of Kaiserslautern (RPTU), Germany. My work focuses on hybrid thermodynamic models that combine physical knowledge with machine learning.
See here for more.