Machine Learning-Based Experimental Design in Chemical Industry: Taking C4 Olefin Preparation from Ethanol as an Example

Abstract:
With the acceleration of global carbon-neutralizing process, green chemical industry has attracted more and more attention. Finding appropriate parameters for experiments is important for green chemical reaction. This work focuses on using experiment parameters to predict chemical reaction yield of products based on machine learning methods, which takes C4 olefin preparation from ethanol as an example. To explore how experiment parameters influence the reaction yield of products, figures are plotted to show the correlations between parameters and reaction index. Then the conversion models and selectivity models are constructed based on polynomial regression, decision tree, random forest, Light GBM and neural networks. Different evaluations of the models all indicate that LightGBM performs best to predict the conversion of ethanol, while the selectivity of C4 olefin prefers neural network. Comprehensively we propose a hybrid method to predict yield of C4 olefin which significantly exceeded all single regression models. Furthermore, the sensitivity of optimized conversion/selectivity/yield models are shown in heatmaps that imply the robustness of our models.