Multi-output machine learning models for kinetic data evaluation : A Fischer-Tropsch synthesis case study

Predicting the impact of input process variables on chemical processes is key to assess their performance of the latter. Models explaining this impact through a mechanistic approach are rarely readily available, complex in nature and/or require long development time. With increased automation in industries and the availability of high-throughput experimental data, data-driven machine learning models are gaining popularity due to their simplicity and reduced computational effort. In this work, multi-output Fischer-Tropsch Synthesis data generated via mechanistic Single-Event MicroKinetic (SEMK) model is analyzed with different machine learning models (ML) such as Lasso, K Nearest Neighbors (KNN), Support Vector Regression (SVR), and Artificial Neural Network (ANN) regression. Temperature and pressure are identified as the dominant input variables. Among the considered ML models, ANN emerged as the superior performing one with respect to benchmark SEMK results. In addition, the validity of neural network predictions is verified using the so-called Shap-value interpretation technique. The relative impact of input variables obtained using Shap values, on conversion follow the order of temperature (1x) > pressure (0.22x) > space-time (0.1x) > syngas ratio (0.03x). Temperature (1x) and pressure (0.26x) remain the dominant input variables for light olefin selectivity, but that of space-time (0.03x) and syngas ratio (0.03x) becomes comparable. This work provides a reference method for the identification of suitable ML models for multi-output prediction in chemical processes.

Automated summary

The impact of input variables on chemical processes can be predicted using machine learning models. In this study, the dominant input variables for Fischer-Tropsch Synthesis were identified as temperature and pressure, with Artificial Neural Network (ANN) model performing the best among others.