A Novel Hybrid Modeling Framework for Polymer Reaction Engineering with Application to Real-Time Process Monitoring

Hybrid modeling frameworks that integrate knowledge-driven
and
data-driven approaches offer a promising solution for modeling complex
physical systems, especially when data are limited and domain knowledge
is well established. In this work, a hybrid modeling strategy is proposed,
coupling Gaussian Processes with a kinetic polymerization model for
the case of the free-radical polymerization of styrene. The kinetic
model serves as the mean function of the Gaussian Process prior embedding
fundamental system knowledge into the regression process. The proposed
framework enables the model to deliver physically consistent predictions
in the absence of data while allowing fast adaptation when new data
becomes available. This is demonstrated in a real-time process monitoring
context, where data points arrive sequentially and the model updates
its predictions accordingly, while avoiding the need for extensive
training, typically required by DD models, such as neural networks.
The approach is validated using two data sets across a wide range
of reaction temperatures (85–150 °C), showcasing its performance
under varying degrees of model-data agreement. The results highlight
the framework’s ability to learn from limited data, maintain
physical plausibility through constrained predictions, and incorporate
uncertainty for improved robustness. This study, through a simple
example, lays the groundwork for broader applications of Gaussian-Process-based
hybrid models in polymerization systems, with future work aimed at
extending their scope to additional process variables and constraint
handling strategies.