Sequential Injection/Mid-Infrared Spectroscopic Analysis of an Acetone‐Butanol‐Ethanol Fermentation: Analyte Cross‐Correlation Effects

Mid‐infrared spectroscopy together with sequential injection analysis (SIA) and partial least squares (PLS) regression analysis was used to monitor acetone‐butanol‐ethanol (ABE) fermentations under different fermentation conditions. Five analytes were simultaneously predicted (acetone, acetate, butyrate, n‐butanol, and glucose). In order to compare the overall model prediction ability, a relative average of the root mean square error of prediction (RMSEP) across all five analytes was employed. To form a PLS model devoid of any cross‐correlations between analytes, a synthetic calibration data set was created by the SIA system. As a test of their robustness, PLS models from synthetic samples and those from real fermentation samples were compared and used to predict samples from the opposite data set and from independent “acid‐crash” fermentations. The PLS model developed from the synthetic samples proved to be far more robust and accurate and used fewer factors than PLS models from the real fermentations, which were found to contain analyte cross‐correlations. The use of synthetic data enabled more accurate selection of factors and showed the importance of investigating spectral regression coefficients plots to aid and confirm appropriate factor selection. In addition, an alternative method of factor selection was proposed, using a “similarity measure” between the regression coefficient plots of factors for certain analytes and their standard spectra. Predictions using this method of factor selection over the common “minimum from an error vs. factor” plot proved to be more accurate and used far fewer factors.