Design of experiments and regression modelling in food flavour and sensory analysis: A review

Advances in food flavor analysis and sensory evaluation techniques and applications: Traditional vs emerging.

Response surface methodology has been applied in numerous studies using polynomial models despite some of them exhibit poor correlation coefficient , which implies an incorrect optimization. In this work, a methodology for obtaining response surfaces when a small data set is available and its behavior is complex is presented. The methodology consists of four steps. First, the classic experimental design is used for obtaining a data set. Second, using the experimental results from a design of experiment, the classical kriging is employed for estimating properties at unsampled locations. Third, a response surface is obtained by training an artificial neural network using the kriging-estimated data set, and a hybrid algorithm based on differential evolution and backpropagation algorithms. The verification of the model is accomplished with the experimental data set. Fourth, the uncertainty quantification is utilized for studying the behavior of the response against uncertainties, which guarantee the robustness of the model developed. The methodology was applied to three cases, considering verification, validation, and uncertainty quantification. The results indicate that the proposed methodology is robust, and it provides more stable response surfaces than the approaches commonly used for polynomials and artificial neural networks. As a result, better optimal conditions are attained.

The following work documents the development of an artificial neural network (ANN) model of an advanced, turbocharged diesel engine, and its implementation within an overall vehicle simulation. The purpose of this research was to create a model that would help drastically reduce in-vehicle engine design cycle time, improve numerical robustness and maintain the predictive capability of its physically based counterpart. Ease of integration within an optimization framework for preliminary investigations of engine designs was also a significant need. A design of experiments (DOE) was first used to sample the engine' feasible design domain. The high-fidelity model calculated engine responses at these DOE designs; the DOE inputs and engine model responses were subsequently used to train a radial basis function ANN. A typical, vehicle-level optimization study is presented which compares results and time requirements of the ANN and high-fidelity models. Critical issues in the development of an ANN engine model are detailed and generic guidelines are presented to expedite the implementation of this surrogate modelling technique to other complex subsystems and components.

Dry reforming of methane is a potentially important process to convert the greenhouse gases carbon dioxide and methane simultaneously to syngas (CO + H2). The most serious problem with the dry reforming of methane is carbon deposition, so preparation parameters of the citric acid method were surveyed to prepare an active Co-MgO catalyst with low carbon deposition using design of experiment, artificial neural network and grid search. The preparation parameters such as Co loading, amount of citric acid, calcination temperature, and pelletization pressure were determined according to an L9 orthogonal array. After 9 data sets of the parameter activity were designed and measured in a conventional pressurized fixed bed reactor, an artificial neural network was constructed. The optimum composition was determined by a grid search and verified experimentally to be active with a small amount of carbon deposition. Design of experiment combined with an artificial neural network and grid search was useful for catalyst development.

The kinetic of catalytic upgrading of anisole as a lignin−derived bio−oil component is investigated experimentally over Pt/γAl2O3 at 573−673 K and 14 bar. According to experimental results, benzene, phenol, 2−methylphenol, 2,6−dimethylphenol, 2,4,6−trimethylphenol, and hexamethylbenzene are identified as the main products. The results indicated that the kinetically significant reaction classes are hydrogenolysis, hydrodeoxygenation (HDO), alkylation, and hydrogenation. The response surface methodology (RSM) is applied to optimize the experimental data which obtained at suggested conditions by design of experiment (DOE). Due to the complex nature of the system, artificial neural networks (ANNs) were employed as an efficient tool to model the behavior of the system. RSM and ANN methods were constructed based upon the DOE’s points and then utilized for generating extra−simulated data. Data simulated by the RSM/ANN method were used to fit power law kinetic rate expressions for the reactions. The coefficient of determination (R2 ) was obtained 0.998 and 0.973 for anisole conversion model and benzene selectivity model which represented the high accuracy of model predictions. The correlation coefficient (R) and mean square error (MSE) of ANN model equaled to 0.97 and 8.3 × 10−12 respectively means high accuracy of the developed model. The results of kinetic modeling with simulated data from the ANN and RSM models revealed that the highest reaction order during the upgrading process of anisole belongs to hydrogenolysis of anisole to phenol. Also the activation energy of hydrogenolysis reaction was lower than HDO.

The comparative analysis of predictive methodologies highlights the original contribution of this study in optimizing the prediction of Rate of Penetration (ROP) in mining drilling operations. The emphasis on employing advanced Artificial Neural Networks (ANN), fuzzy logic, and linear regression models provides new insights into enhancing predictive accuracy and operational efficiency in mining practices. This study aims to quantify the effects of three pivotal drilling parameters: compressive strength, rotational pressure, and thrust pressure on the rate of penetration, a critical performance metric in mining drilling operations. Additionally, the study seeks to develop and evaluate advanced predictive methodologies for predicting ROP. The effects of compressive strength, rotational pressure, and thrust pressure on the rate of penetration were investigated through a Design of Experiments (DOE) approach. Initially, the main effects and two-way interactions among these variables were identified using DOE. Subsequently, three predictive methodologies: linear regression, fuzzy logic, and artificial neural networks, were developed and evaluated to predict ROP based on the identified factors. The evaluation of predictive methodologies revealed that the ANN model demonstrated superior accuracy in predicting the ROP, achieving over 95 % accuracy. Additionally, the fuzzy logic model provided effective handling of nonlinearities in the data, while the linear regression model offered initial insights into the relationships between the variables. The application of advanced predictive methodologies: artificial neural networks, fuzzy logic, and linear regression to optimize the prediction of rate of penetration in mining drilling operations offers precise insights into drilling parameter interactions, enhancing operational efficiency and supporting informed decision making in mining practices.

Food flavor analysis 4.0: A cross-domain application of machine learning

Summary In this paper, apple crispness was evaluated by sensory evaluation and compared with non‐destructive measurements of portable acoustic signal to discuss the feasibility of non‐destructive evaluation for apple crispness based on portable acoustic signal. Acoustic eigenvalues from the acoustic signal were processed by time domain and Hilbert–Huang transform (HHT), followed by analysing the correlations with apple crispness that had been evaluated via sensory evaluation. Multiple linear regression (MLR) and artificial neural network (ANN) were applied to predict apple crispness. The results proved that crispness correlates significantly ( P < 0.01) with four acoustic eigenvalues, including waveform index, sound intensity, energy of low frequency and energy of high frequency. The average relative error of apple crispness predicted by ANN was 1.42 ± 1.9%, remarkably lower ( P < 0.01) that of MLR (6.79 ± 5.64%), implying that the model predicted by ANN is more accurate than that of MLR.

In this paper, optimization of the first blade of a new test rig is pursued using a hybrid model comprising the genetic algorithm, artificial neural networks and design of experiments. Blade tuning is conducted using three-dimensional geometric parameters. Taper and sweep angle play important roles in this optimization process. Compressor characteristics involving mass flow and efficiency, and stress and eigenfrequencies of the blades are the main objectives of the evaluation. Owing to the design of blade attachments and their dynamic isolation from the disk, the vibrational behavior of the one blade is tuned based on the self-excited and forced vibration phenomenon. Using a semi-analytical MATLAB code instability, the conditions are satisfied. The code uses Whitehead’s theory and force response theory to predict classical and stall flutter speeds. Forced vibrational instability is controlled using Campbell’s theory. The aerodynamics of the new blade geometry is determined using multistage computational fluid dynamics simulation. The numerical results show increasing performance near the surge line and improvement in the working interval along with a 4% increase in mass flow. From the vibrational point of view, the reduced frequency increases by at least 5% in both stall and classical regions, and force response constraints are satisfied.

Design of experiments and focused grid search for neural network parameter optimization

Editorial overview: Sensory science and consumer perception: holistic approach to sensory science to better understand consumer perception, acceptance, and behavior toward foods

In this work, a comparative study between two methods to acquire relevant information about a cosmetic formulation has been carried out. A Design of Experiments (DOE) has been applied in two stages to a capillary cosmetic cream: first, a Plackett-Burman (PB) design has been used to reduce the number of variables to be studied; second, a complete factorial design has been implemented. With the experimental data collected from the DOE, a Least Mean Square (LMS) algorithm and Artificial Neural Networks (ANN) have been utilized to obtain an equation (or model) that could explain cream viscosity. Calculations have shown that ANN are the best prediction method to fit a model to experimental data, within the interval of concentrations defined by the whole set of experiments.

This paper investigates the combined effects of parameters such as die surface finish, die stoning direction, contact pressure and draw speed on the resulting friction coefficient. The study shows the correlation between the artificial neural network (ANN) and the design of experiments (DOE). The ANN and the DOE have been implemented for minimizing the friction coefficient and for investigating the effect of the parameters in the straight pulling friction test effectively. The experiments were run in random order with at least three replications. It was found that directional stoning orthogonal to the pulling direction lowers the friction more than directional stoning parallel to the pulling direction using die material FCD55 and sheet material SPCEN.

Tilting pad journal bearings (TPJBs) are modeled with Reynold-based models or computational fluid dynamics (CFD) approach. In both cases, the estimation of the dynamic coefficients of the oil-film forces and the static characteristic, can be computationally expensive and time consuming. Artificial Intelligence (AI) is assuming a key role in engineering but is rarely applied in fluid film bearing analysis. A properly trained Deep Learning (DL) model can perform very fast predictions of TPJB behavior with accuracy comparable to more time-consuming models. In this case, the main drawback is the time required to build the training dataset. In this work, an Artificial Neural Network (ANN) is trained to predict the dynamic stiffness and damping coefficients along with the main static quantities of TPJBs, such as minimum oil-film thickness and inlet flowrate. At first, a design of experiment is performed to build an appropriate training dataset. Secondly, a Reynolds-based thermo-hydrodynamic (THD) model is used to populate the training dataset and an appropriate test dataset. Then, a feed-forward ANN is trained with Levenberg–Marquardt backpropagation and its architecture is optimized to increase accuracy. Finally, the accuracy of the ANN is tested using the test dataset and experimental data. The time and computational effort required by the ANN regression are much less than those required by the THD model. Therefore, the trained ANN is an effective and efficient tool for the characterization of TPJBs.

Pasta is a traditional wheat-based food, whose popularity, indisputable in Italy and progressive in Europe as well as in America and in the East, is essentially due to the nutritional and sensory qualities that have made it an attractive source of carbohydrates, enviable and difficult to replace in a healthy and balanced diet. The overall pasta quality is highly influenced by raw materials and production technologies. As far as the raw material, the protein content and the gluten quality of semolina are crucial factors in achieving good texture and cooking performance of the final product. Among the technological processes, the drying step is the most delicate and important The aim of the thesis was to evaluate the effect of raw material, technological processes and cooking methods on pasta quality. The work is split into three cases study. The object of first case study was to evaluate the effect of semolina from both mono-varietal and commercial mix durum wheat on qualitative characteristics of dry pasta; moreover, for each semolina was conducted a study on the effects of high and very high drying temperature on final product in order to assess if and how drying method could affect pasta. Six monovarietal semolina and two commercial mix were used to produce spaghetti than subjected to drying according to two different drying cycles: high temperature (75 ° C) and very high temperature (90 ° C) drying cycles thus obtaining a total of sixteen samples, two for each semolina sample. In the second case study, was evaluated the cooking quality of alternative cooked pasta with the purpose to understand which may be the best (reducing time and /or energy) alternatives to traditional cooking. Four different hollow shapes coming from the same production lot were used; pasta sample were than underwent to passive cooking and microwave cooking. Pasta quality was evaluated by means of sensory, physical and structural analysis by using traditional cooking as reference Finally the third study aimed to asses the influence of chickpea flour and microwave drying operation on starch digestibility and glycemic index of composite legume-wheat pasta. A substitution of 20% chickpea flour was made with respect to the reference sample (100% semolina). Six pasta samples were prepared from 100% semolina, semolina (80%) and raw chickpea flour (20%), semolina (80%) and cooked chickpea flour; each sample was than subjected to traditional drying (40°C; RH 80%) and an optimized microwave drying.. The results showed that chemical, physical, and sensory analysis were all very sensitive to changes in pasta quality. Raw material, process operations and cooking influenced the qualitative characteristics of pasta. The addition of different ingredients in pasta improved the nutritional characteristics of final product. The use of new technologies of drying brought advantages in terms of reduction of time and cost.