Modelling of spray drying of coconut milk using different artificial neural network and response surface methodology modelling approaches: A comparison

• Published in: AIP Conference Proceedings
• Main Author: 2-s2.0-85218694751
• Format: Conference paper
• Language: English
• Published: American Institute of Physics 2025
• Online Access: https://www.scopus.com/inward/record.uri?eid=2-s2.0-85218694751&doi=10.1063%2f5.0249004&partnerID=40&md5=40a433789c9baa6a47b386258f361480

Spray drying process is a complex process that involves heat transfer, and phase change in a short period of time. The application of artificial neural network in spray drying process could be an alternate modelling technique as it can interconnect different process elements and provide accurate modelling. In this study, the spray drying model of coconut milk using maltodextrin and sodium caseinate was studied and compared through artificial neural network (ANN), genetic algorithm (GA) and particle swarm optimization (PSO). The ANN model is enhanced by global search algorithm such as GA and PSO to provide better performing models. The effect of three independent variables, viz. inlet spray drying temperature (130-170°C), maltodextrin concentration (0-10 %w/w) and sodium caseinate concentration (0-10 %w/w) on the three independent variables which are moisture content (%), surface free fat (mg/100g) and outlet temperature (°C) was determined by rotation central composite design. The application of three different artificial neural network models (ANN, GA-ANN and PSO-ANN) were compared with response surface methodology (RSM) model. All models were validated using external datasets through model simulations and further compared based on performance indices like mean square error (MSE), correlation coefficients (R2), standard error of prediction (SEP) and model predictive error (MPE). All three dependent variables showed that the GA-ANN model recorded the lowest MSE, SEP and MPE followed by PSO-ANN, ANN, and RSM model. Significance results directed that moisture content from the GA-ANN model recorded the lowest MSE values (0.009) and the highest R2 value (0.932) and is supported by the lowest SEP (0.011%) and MPE (0.252%). Overall, hybrid ANN models, especially the GA-ANN is the most suitable for spray drying modelling process. Applying the combination of different techniques has been shown to increase the versatility of ANN modelling and extend the application of ANN in food simulation and control. © 2025 Author(s).