Predicting the ethanol potential of wheat straw using near-infrared spectroscopy and chemometrics: The challenge of inherently intercorrelated response functions

- Intercorrelated response variables influence the regression models.
- Making a calibration model from NIR on straw to enzymatic released sugars is not straight forward.
- Reducing the inherent correlation makes it possible to understand the underlying dependencies in the regression models.
- Models have been validated, both through chemometric tools, as well as in detail understanding of the system.

The combination of NIR spectroscopy and chemometrics is a powerful correlation method for predicting the chemical constituents in biological matrices, such as the glucose and xylose content of straw. However, difficulties arise when it comes to predicting enzymatic glucose and xylose release potential, which is matrix dependent. Further complications are caused by xylose and glucose release potential being highly intercorrelated.This study emphasizes the importance of understanding the causal relationship between the model and the constituent of interest. It investigates the possibility of using near-infrared spectroscopy to evaluate the ethanol potential of wheat straw by analyzing more than 1000 samples from different wheat varieties and growth conditions. During the calibration model development, the prime emphasis was to investigate the correlation structure between the two major quality traits for saccharification of wheat straw: glucose and xylose release. The large sample set enabled a versatile and robust calibration model to be developed, showing that the prediction model for xylose release is based on a causal relationship with the NIR spectral data. In contrast, the prediction of glucose release was found to be highly dependent on the intercorrelation with xylose release. If this correlation is broken, the model performance breaks down. A simple method was devised for avoiding this breakdown and can be applied to any large dataset for investigating the causality or lack of causality of a prediction model.

311