Bruno Perret

Determination of Major Compounds of Alcoholic Fermentation by Middle-Infrared Spectroscopy: Study of Temperature Effects and Calibration Methods

The potential of Fourier transform middle-infrared spectroscopy has been demonstrated for the quantitative analysis of substrates (glucose and fructose) and metabolites (glycerol and ethanol) involved in alcoholic fermentation. Temperature variations between samples and water background reference caused changes in absorbance, and therefore the prediction of concentrations with partial least-squares (PLS) regressions was affected. The same temperatures for the calibration, validation, and prediction sets gave the smallest standard error of prediction (SEP): SEPglucose = 3.9 g L−1; SEPfructose = 4.3 g L−1; SEPglycerol = 0.5 g L−1; SEPethanol = 1.3 g L−1. In order to take different working temperatures (18, 25, and 35 °C) into account, an artificial neural network was used to create a nonlinear multivariate model. Compared to PLS regression, this method provided better results, especially for glycerol and ethanol, where SEP decreased by 0.3 g L−1 and 0.4 g L−1, respectively.

A hybrid recurrent neural network model for yeast production monitoring and control in a wine base medium

Abstract A dynamic model based on a recurrent neural network was established to follow the growth of yeast in a wine-base medium. It leads to the estimation and prediction of the yeast concentration in batch cultures, based on the on-line measurement of the volume of CO 2 released and the initial yeast concentration. The mean error of the predicted value of the final yeast concentration is lower than 5%. A hybrid model combining this model with a measurement model (based on linear correlations reflecting the reaction scheme) also leads to the estimation and prediction of the sugar and ethanol concentrations in the culture medium with respective mean errors of 1.6 and 1 g l −1 . Moreover, this model was used in an open-loop control strategy in order to achieve a final concentration of yeast by setting the culture temperature. Adjusting culture temperature during growth was necessary for only 4% of the cultures, in order to remain within the range of measurement error (3×10 6 cells ml −1 ) of yeast concentration. The performance of the model and of the control algorithm used could be assessed by controlling six successive cultures.

Characterizing acidification kinetics by measuring pH and electrical conductivity in batch thermophilic lactic fermentations

Abstract On-line measurements of pH and electrical conductivity data have made it possible to access time and rate feature points of thermophilic lactic acid fermentations. Ten feature points characterize curves of acidification and conductivity changes using the main points of inflection observed. The presence or absence of urease also changes the observed kinetics and corresponding feature points. These phenomena and the time patterns of the biomass and products permit an understanding of the meanings of the feature points. These points showed the excellent reproducibility of fermentations conducted in standard conditions, with coefficients of variation lower than 5.1% for nine of them. They were also used to compare the effects of the type of starter, temperature and culture medium. The temperature affects the urease activity and acidification optima. The presence of fat (32 g/l) in the medium does not change any feature point, while the presence of sucrose (90 g/l) results in a decrease in the acidification rate and a longer fermentation time.

Yeast concentration estimation and prediction with static and dynamic neural network models in batch cultures

The second fermentation is one of the most important steps in Champagne production. For this purpose, yeasts are grown on a wine based medium to adapt their metabolism to ethanol. Several models built with various static and dynamic neural network configurations were investigated. The main objective was to achieve real-time estimation and prediction of yeast concentration during growth. The model selected, based on recurrent neural networks, was first order with respect to the yeast concentration and to the volume of CO2 released. Temperature and pH were included as model parameters as well. Yeast concentration during growth could thus be estimated with an error lower than 3% (±1.7×106 yeasts/ml). From the measurement of initial yeast population and temperature, it was possible to predict the final yeast concentration (after 21 hours of growth) from the beginning of the growth, with about ±3×106 yeasts/ml accuracy. So a predictive control strategy of this process could be investigated.

A Decision Support System Coupling Fuzzy Logic and Probabilistic Graphical Approaches for the Agri-Food Industry: Prediction of Grape Berry Maturity.

Agri-food is one of the most important sectors of the industry and a major contributor to the global warming potential in Europe. Sustainability issues pose a huge challenge for this sector. In this context, a big issue is to be able to predict the multiscale dynamics of those systems using computing science. A robust predictive mathematical tool is implemented for this sector and applied to the wine industry being easily able to be generalized to other applications. Grape berry maturation relies on complex and coupled physicochemical and biochemical reactions which are climate dependent. Moreover one experiment represents one year and the climate variability could not be covered exclusively by the experiments. Consequently, harvest mostly relies on expert predictions. A big challenge for the wine industry is nevertheless to be able to anticipate the reactions for sustainability purposes. We propose to implement a decision support system so called FGRAPEDBN able to (1) capitalize the heterogeneous fragmented knowledge available including data and expertise and (2) predict the sugar (resp. the acidity) concentrations with a relevant RMSE of 7 g/l (resp. 0.44 g/l and 0.11 g/kg). FGRAPEDBN is based on a coupling between a probabilistic graphical approach and a fuzzy expert system.

On-line ethanol estimation and prediction: application to the production of low alcohol wines

Abstract An on-line method of estimating ethanol concentration during production of semi-sparkling wine is described. During the first stage, in an open reactor, a linear correlation between the volume of CO 2 released and the concentration of ethanol produced allowed estimation of ethanol up to 18 g/litre (mean error = 0·3 g/litre). During the second stage, in a closed reactor, the concentration of ethanol reached 24 g/litre and could be estimated from the pressure increase (mean error = 0·29 g/litre) or predicted by means of a simple model.

A probabilistic graphical model for describing the grape berry maturity

Dynamic Bayesian networks for coupling heterogeneous data and expertise knowledge.The modeling of grape berry maturity over the time tainted with uncertainty.Prediction of sugar, acidity and anthocyanin concentrations over the maturity. Grape berry maturation depends on complex and coupled physiological and biochemical reactions which are climate dependant. Moreover one experiment represents one year and the climate variability could not be covered exclusively by the experiments. Consequently, harvest mostly relies on expert prediction. A big challenge for the wine industry is nevertheless to be able to anticipate the reactions for sustainability purposes. We propose to implement a robust mathematical model able (1) to capitalize the heterogeneous fragmented available knowledge including data and expertise by means of probabilistic graphical approaches; and (2) to predict sugar, acidity and anthocyanin concentrations over the maturity.

Camembert-type cheese ripening dynamics are changed by the properties of wrapping films

Four gas-permeable wrapping films exhibiting different degrees of water permeability (ranging from 1.6 to 500 g/m(2) per d) were tested to study their effect on soft-mold (Camembert-type) cheese-ripening dynamics compared with unwrapped cheeses. Twenty-three-day trials were performed in 2 laboratory-size (18L) respiratory-ripening cells under controlled temperature (6 ± 0.5°C), relative humidity (75 ± 2%), and carbon dioxide content (0.5 to 1%). The films allowed for a high degree of respiratory activity; no limitation in gas permeability was observed. The wide range of water permeability of the films led to considerable differences in cheese water loss (from 0.5 to 12% on d 23, compared with 15% for unwrapped cheeses), which appeared to be a key factor in controlling cheese-ripening progress. A new relationship between 2 important cheese-ripening descriptors (increase of the cheese core pH and increase of the cheeses creamy underrind thickness) was shown in relation to the water permeability of the wrapping film. High water losses (more than 10 to 12% on d 23) also were observed for unwrapped cheeses, leading to Camembert cheeses that were too dry and poorly ripened. On the other hand, low water losses (from 0.5 to 1% on d 23) led to over-ripening in the cheese underrind, which became runny as a result. Finally, water losses from around 3 to 6% on d 23 led to good ripening dynamics and the best cheese quality. This level of water loss appeared to be ideal in terms of cheese-wrapping film design.

Comparison of Partial Least Squares Algorithm and Artificial Neural Networks for the Prediction of the Concentration of Molecules Involved in Alcoholic Fermentation with Mid Infra-Red Spectroscopy

The potential of the Fourier Transform Infra-Red (FT-IR) spectroscopy was demonstrated for the quantitative analysis of biological molecules. But, the complexity of spectra obtained from fermentation media required the use of multivariate methods for the quantitative analysis. For the analysis of compounds, the classic laboratory technics such as flow injection analysis, gas chromatography, high performance liquid chromatography are tedious and need samples preparations (centrifugation, filtration, precipitation, dilution, etc.). But, for several years, FT-IR spectroscopy is well known to realize fast (some seconds for a spectrum with a high signal-to-noise ratio) and accurate analysis of many biological [1] and fermentation media [2]. For the quantitative analysis of this complex multi-component media, many chemometrical technics [3] are used such as multilinear regression (MLR), principal component regression (PCR), partial least squares (PLS). Among all this multivariate technics, PLS algorithm give the best results [4]. Recently, the artificial neural networks (ANN) had a great expansion. For instance, Bhandare et al. [5] showed that ANN could improve the prediction of glucose in whole blood when the relationship between spectral data and concentrations of this compound became nonlinear. The nonlinear phenomenons can be of several orders such as the response of the detector, the multiple scattering of the IR light in the sample, the temperature and the composition of the sample, etc... The goal of this study is to compare the prediction of the concentration of compounds involved in alcoholic fermentation with PLS algorithm and ANN.