Deep neural network for generalizing and forecasting on-demand drying kinetics of droplet solutions

Abstract

Recent trend in consumer preferences toward healthy and on-demand-to-make items has mandated food manufacturers to seek more efficient production processes in order to make their businesses sustainable. For powder producers in particular, the business models are drifting from ‘ready-made’ toward on-the-go items, thus requiring an urgent attention. Here, we present a long short-term memory (LSTM) model for forecasting drying kinetics histories of lactose, low-fat and high fat milk droplet solutions; covering a wide gas-and-material spectrum, utilizing only their initial conditions. Within the examined range, results show that the forecast lactose temperature and mass drying curves have accuracies within 0.9987 and 0.9841 respectively; that improve in the order of training-material-blend/training-data-rows as thus: lactose/3612, lactose-fat/4515, lactose-protein/5317 and lactose-fat-protein/6220. This indicates the impact of data size on the model accuracy and generalization of the trained LSTM network. For low/high fat milk (20–30% total solid), accuracy margins are within 0.0179 and 0.0655 in the range of 1–8 combined test samples. Beyond 8 samples up to 20 combined scenarios, accuracies are capped at 0.9749 for temperature and 0.9000 for mass profiles. Since only the initial conditions are required by the developed model to provide forecasting, this removes cost and time barriers inherent in traditional approaches during new product launch. Future application of deep learning models would integrate the presented LSTM network to consider actual characteristics of material mixing such as colour, texture and taste.