Predictions of Roasted Coffee Bean Quality Through Models

Abstract

Coffee aroma compounds are one of the roasted coffee bean quality signatures and have played a major role in establishing the popularity of coffee beverages worldwide. The formation of coffee aroma happened during roasting, where the green beans were exposed to the hot air (200- 240 ℃) for 10 to 15 minutes or until it became roasted/dark brown coloured. During coffee bean roasting, the temperature of hot air and roasting time are two important process variables that could be used to control the roasted coffee bean quality. However, previous works on coffee roasting models mainly used constant heating air that has limited application in a coffee sensory modification. Therefore, our study had focused on modelling the varying heating profile for coffee bean roasting to investigate the effect it has on roasted coffee bean qualities. Our works can be divided into 3 sub-topics; the development of 2-D single coffee bean model for bean’s temperature prediction, the coupled 2-D model with multiresponse kinetic model and neural network model for coffee aroma formation prediction, and the investigation between roasting conditions and coffee bean’s colour and sensory evaluation using the image and data analysis.First of all, we had developed a 2-D model of single coffee bean roasting. Unlike existing 3-D model of single coffee bean roasting that usually used constant heating temperature, our 2-D model could be used for varying heating profile with an affordable computational cost. The 2-D model was designed to achieve minimum discrepancies with experimental data using estimated parameters of bean’s thermal conductivity and moisture diffusivity. The model validations were carried out for different bean sizes, different heating profiles, and the cooling stage after roasting with an average root mean square error (RMSE) of 3 ℃. The 2-D model was also compared with a 3-D model which gave an absolute difference in error of less than 5 ℃. The predictive ability of the 2-D model could be used in forecasting other product qualities that are dependent on bean temperature such as the formation of aroma compounds. Secondly, we had coupled the previous 2-D model with a multiresponse kinetic model of coffee aroma formation. 2-methylpyrazine was chosen as the studied coffee aroma, which belongs to one of the major aroma group in roasted coffee called alkylpyrazine, contributes to the nutty, roasty, and earthy notes in coffee aroma. A simple first-order multiresponse kinetic model was used to describe the formation of 2-methylpyrazine during roasting. The multiresponse kinetic model has an average RMSE of 0.24 ng/mg, and poor model fit for moderate and slow heating profiles. Therefore, we had introduced a long short-term memory (LSTM) model, an artificial neural network (ANN) that was able to predict the aroma profile for multiple sets of varying heating profiles (slow, moderate, and fast). The LSTM model validations gave excellent results with an average RMSE of less than 0.1 ng/mg. Finally, we had investigated the relationship between the roasting conditions and coffee bean’s colour and sensory evaluation using the image and data analysis. Conventionally, the coffee bean’s colour and its’ sensory evaluation were used to monitor roasted coffee beans’ quality. The relationship between these two quality indicators with roasting conditions is uncertain. Therefore, we had investigated the effect of degree of roasting (light, medium, and dark roast)and multiple sets of varying heating profiles (slow, moderate, and fast) towards the coffee sensory evaluation. In addition to that, we introduced two image acquisitions; digital camera and hyperspectral imaging (HSI) to analyse the studied coffee bean powder. The data analysis used is principal component analysis (PCA) and ‘naïve’ Bayes classification which was able to separate between light and dark degree of roast. The coffee sensory attributes had shown considerable differences between light, medium, and dark roast. However, the analysed data for multiple sets of varying heating profiles showed random distribution and became dense/complex towards the dark roasting degree. In conclusions, our study had contributed to a better understanding of roasting conditions specifically on the multiple varying heating profiles. The proposed 2-D model, LSTM network, and image data analysis had proven to be powerful tools as a soft sensor to predict the roasted coffee beans’ quality during roasting. These techniques could not only be used for the in-depth study of roasted coffee bean qualities but also could be applied to other food processing.