(452a) Using Artificial Neural Networks to Estimate Xylose Conversion and Furfural Yields in Autocatalytic, Organic/Aqueous Solvent Systems | AIChE

(452a) Using Artificial Neural Networks to Estimate Xylose Conversion and Furfural Yields in Autocatalytic, Organic/Aqueous Solvent Systems

Authors 

Wettstein, S. - Presenter, Montana State University
Stratton, S., Tulane University
Umhey, C. E., Washington State University
Job, A. L., Colorado School of Mines
Hoo, K. A., Gonzaga University
Artificial neural networks are becoming more commonly used in order to estimate outputs based on important inputs. In this work, an ANN was built to predict xylose conversion and furfural yield in autocatalytic organic/aqueous reactions that explained over 90% of variance for each output. This ANN allows for methodical solvent selection versus the traditional “guess and check” method, which reduces solvent waste and is better for the environment.

Platform chemicals such as furfural can be derived from biomass and present a valuable source as intermediates for fuels and chemical commodities. The dehydration reactions that form furfural can be catalyzed in a hydrothermal system as the result of hydronium ions from high temperature water and reaction byproducts, such as formic acid. In this work, we created a feed-forward artificial neural network (ANN) that estimates the xylose conversion and furfural yield from the reaction severity, which is a function of temperature and time, as well as the Hansen Solubility Parameter (HSP) for polarity for reaction systems consisting of organic/aqueous solvent mixtures. Using both severity and polarity as independent input variables, the percent variance explained by the ANN model was 95.5% for xylose conversion and 92.1% for furfural yield. In the case of xylose conversion, only one predicted point was outside of the 95% prediction interval and for furfural yield, only two points were outside. Importantly, the 10 experimental points that were not used to train the ANN were all estimated within the 95% interval, which indicates the ANN model captured the nonlinear correlations between the input and output variables for the entire experimental range of polarity and severity.