Comparison of machine learning methodologies for predicting kinetics of hydrothermal carbonization of selective biomass

We have examined performance of various machine learning (ML) methods (artificial neural network, random forest, support vector-machine regression, and K nearest neighbors) in predicting the kinetics of hydrothermal carbonization (HTC) of cellulose, poplar, and wheat straw performed under two different conditions: first, isothermal conditions at 200, 230, and 260 degrees C, and second, with a linear temperature ramp of 2 degrees C/min from 160 to 260 degrees C. The focus of this study was to determine the predictability of the ML methods when the biomass type is not known or there is a mixture of biomass types, which is often the case in commercial operations. In addition, we have examined the performance of ML methods in interpolating kinetics results when experimental data is available for only a handful of time-points, as well as their performance in extrapolating the kinetics when the experimental data from only a few initial time-points is available. While these are stringent tests, the ML models were found to perform reasonably well in most cases with an averaged mean squared error (MSE) and R-2 values of 0.25 +/- 0.06 and 0.76 +/- 0.05, respectively. The ML models showed deviation from experimental data under the conditions when the reaction kinetics were fast. Overall, it is concluded that ML methods are appropriate for the purpose of interpolating and extrapolating the kinetics of the HTC process.