BOOK OF ABSTRACTS 53 IHSS&IWA26 / BRNO / CZECHIA / 23–28 August 2026 Tuesday, 25 August 2026 / Hall B+C Analysis and Characterization sources and solid and liquid products following the procedure outlined in ISO Method 19822, omitting separation of the hydrophobic fraction. The volume of the FF was reduced to 15-20 ml by rotary evaporation, and the concentrated FF was freeze dried to a powder. After determining the ash contents, solutions of known FF concentration were prepared with pH adjusted to pH 3 for each sample. These solutions were then used to prepare stock solutions with optical density of 1±0.05 at 240 nm. Integration times were adjusted to obtain a fluorescence intensity of 20,000 ± 500 raw signal counts for the CCD detector. In addition, EEMs were collected from several samples of verified adulterants including lignosulfonates, corn steep liquor and molasses, using identical procedures and parameters. Once EEMs were collected, the emission step size was adjusted to 5 nm, after which the data were corrected for inner filter effects and Rayleigh scattering. Finally, the data were normalized using the builtin Raman Scattering Unit adjust feature, based on the integration time and CCD settings used to collect individual EEMs. Using the Eigenvector Solo (Eigenvector Inc., USA) toolbox, the normalized EEM data from the FF and adulterant samples were first subjected to discrimination for Pass/Fail analysis for the calibration data (n=92 Pass and n=15 Fail) and validation data (n=106 Pass and n=10 Fail) analyses. No sample replicates were shared between the calibration and validation data. The Support Vector Machine Discrimination Analysis preprocessing included a Full Rank extended mixture model and mean centering with Partial Least Squares spectral compression using 2 Latent variables. The resulting confusion tables yielded 100% accuracy (Matthew’s Correlation =1) for calibration and prediction for both the Most Probable and Strict (threshold=0.5 prediction probability) evaluation modes. Very similar results were achieved with Partial Least Squared and Artificial Neural Network algorithms. Samples with verified Passing scores were then evaluated with respect to FF concentration by linear regression using the Local Weighted Regression (LWR) algorithm based on 20 similar points. The calibration and validation data included n=147 and n=31 files and no sample replicates were shared between the two sets. Here preprocessing included autoscaling of the x- and y-data and a General Least Squares Weighting alpha value of 0.02; additionally, the x-block data was compressed using 7 latent variables and 4 principal components for the LWR. The RMSEC and RMSEP were reported over a range of 0 to >560 mg/L as 0.32 and 0.16 mg/L, respectively. The R2 for calibration and validation were both >0.9999. Hence the full range LOD and LOQ for the prediction data were estimated to be 0.528 and 1.62 mg/L for FF. Thus, we conclude that the proposed discrimination-regression method can be used to both accurately authenticate FF samples and quantify the combined hydrophobic and hydrophilic (FF)
RkJQdWJsaXNoZXIy NDA4Mjc=