Generating Data-Driven Models from Molecular-Level Kinetic Models: A Kinetic Model Speedup Strategy

Strategies to reduce the computer time to access the information in molecular-level kinetic models (MLKMs) were evaluated. A triglyceride hydroprocessing MLKM was used to generate data sets for small ranges of input parameters simulating three output parameters. The data sets were used to generate multilinear regression, polynomial regression, decision tree regression, gradient boosting regression, and artificial neural network data-driven model (DDM) representations of the MLKM. All of the DDMs were able to predict results very quickly (≪1 s). The predictive accuracy for the DDMs was compared to the polynomial regression, gradient boosting regression, and artificial neural network models, providing the best models over the entire range of the input parameters selected. However, in narrow input parameter ranges, multiple multilinear models and decision tree models also provide good accuracy, with the added benefit of easily understood parameters and faster solution times. Additionally, multilinear regression models had much lower data requirements than the decision tree regression and artificial neural network models. The major downside to all of the DDMs was shown to be the great loss in accuracy once the input parameters exceed the range of the input parameters in the data sets used to optimize the DDMs. This suggests that the extrapolation capability of the DDMs is very low, and as such, new data should be generated from the MLKM every time predictions are required outside the range of the underlying DDM data.