Data-Driven Approach for Predictive Modeling of By-Product Formation in Methanol Synthesis

In this work, the application of a novel data-driven approach for predictive modeling of by-product formation in methanol (MeOH) synthesis is demonstrated. Due to the number of by-products present in MeOH synthesis, building reliable first-principles models for each by-product is very complex and time consuming. The total by-products of MeOH synthesis are classified into 5 main groups, namely alcohols, esters, ketones, ethers and paraffins, where alcohol and ester groups are usually predominant. In the data preparation phase, a collection of more than 900 experimental points for each individual from test campaigns conducted at a pilot plant was preprocessed using Python (TM). This cleaned dataset was used for model development in proprietary software JMP (R) via neural networks, where special care was taken to ensure a representative training vs. test set distribution, avoidance of overfitting, as well as the physical interpretability of datadriven models. The resulting predictive models have a very good generalization behavior covering the wide range of operating conditions, e.g. the alcohol predictive model has an R-2 value of 0.96, and 66.4% of experimental data points are predicted with +/- 15% accuracy. While a further refinement of the model can be possible through physical considerations, the integration of the model within the MeOH process design workflow is recommended. Moreover, the approach can be extended to other chemical processes.