(346k) Machine Learning Empowered Continuum Lumping Kinetics Model

Catalytic or thermal cracking of complex hydrocarbon mixtures is petroleum refining industry to treat oil. Hydrocarbon mixtures, such as petroleum feedstocks and coal liquid usually are the feedstock for these processes. The reaction involves an indefinitely large number of species, the analysis of which is costly and near impossible. As a result, continuum lumping kinetics, a ‘white box’ method, is used to describe the kinetic behaviour of lump of chemically-similar species. For example, in hydrocracking of paraffin, the disappearance rate of total compounds having similar carbon number is measured, rather than describing the kinetics of each single compound. The lumping methodology effectively simplifies the reaction system of a complex reactive mixture, and therefore, can analyse the product distribution, which is of great significance to optimise the operation process. However, to describe the highly complex mixture kinetics, the lumping method needs simplified assumptions and can lead to the loss of model accuracy.

The recent development of artificial intelligence (AI) has created huge opportunities to use machine learning as a robust predictive tool to effectively solve such highly complex chemical process. The machine learning algorithms, typically Deep Neural Network (DNN) and Artificial Neural Network (ANN) can effectively resolve the relationship between the kinetic data and the product distribution. However, DNN and ANN are just a data-driven ‘black box’ approach, where the description of the reaction process is not clear and with poor interpretability. The relevance of the results obtained through DNN and ANN depends heavily on the quantity and quality of data samples. This implies that the model could be easily over or under fitted, resulting in poor prediction and generalisation ability. To fill the research gaps and overcome the shortcomings in both machine learning and classic mechanistic models, this study proposed a novel solution to hybridise the data-driven framework with mechanism-driven model to create a new ‘grey box’ that can largely improve the interpretability and traceability of the machine learning. The new hybrid model effectively embeds the physical-meaningful continuum lumping kinetic method into data-driven framework, which enabled better kinetic data correlation guided by a clear reflection of the process principles.