(53c) The Pros and Cons of Machine Learning Versus Physical Model ? Case Study on Hydrocracking

Transformation of heavy crude oil into more valuable components: light (naphtha) and middle distillates (kerosene and diesel) is becoming increasingly important for the refining industry. Hydrocracking plays a major role in this transformation. Development of models for industrial hydrocrackers has received a great amount of attention by the scientific community over the past decades. These models can be used for several targets: chose optimal operating condition in order to optimize one target (Real Time Optimization), carry out what if analysis in order to estimate the impact of feed variations... Depending on the target and the reactor to model (Hydrotreating or Hydrocracking), several kinetic models can be used from very simple (classical lumping) to much more complex ones (single events).

These models are challenged by new techniques such as machine learning or deep learning.

This paper proposes a global methodology including design of experiments, deactivation model, Yields models and product properties models. It indicates the best models to use depending on the target (either first principles or machine learning models). And the pros and cons of Machine Learning. Some examples from hydrotreating or hydrocracking will be provided with comparison between first principles models and machine learning models when reliable.

This global methodology can be used for any kind of process.