Hybrid modeling in chemical processes

At RISE, we combine physical models for reaction kinetics with data-driven methods. Modern machine learning can be used to complement physical models in cases where the underlying mechanisms are only partially known. The result is a kind of hybrid model that can be used for forecasting, optimisation and experiment planning.

Data is costly and time consuming to produce. This is also true for the development of mechanistic models that describe chemical processes in detail. In contrast to methods from classical machine learning, hybrid models can be trained on relatively small amounts of data, and do not require a complete physical description of the process. In addition, hybrid models, like mechanistic models, can predict the outcome even outside the measurement range, which is not possible with purely data-driven methods.

Simpler processes can be described by a complete physical model dependent on a number of parameters. In turn, these parameters can be modelled using neural networks and other proven AI technology. In more complex cases, where a complete model cannot be formulated, more sophisticated hybrid models are needed.

Simulation illustrating a hybrid model of three reactions involving five substances of concentrations A,B,C,D and E. Solid lines are predictions according to the hybrid model and dotted lines are ground truth simulations using a white-box model. The training data is plotted as circles.

In our prototype lab we have set up a reaction that produces hydroxymethylfurfural (HMF) from fructose, using an acid catalyst. HMF is a precursor to the production of recyclable alternatives to plastic. The figure shows experimental data of HMF yield as well as the yield prediction of a hybrid model.

Illustration of a simple parametric model of a reaction network of three substances A->B->C. The reaction rates for each substance depends non-linearly on temperature. White-box and gray-box (hybrid) model yields are shown, as well as reaction rate coefficients and training loss.

Contact person

David Eklund

Senior Forskare

+46 10 228 42 16