Machine Learning Performance Prediction of an Ion Deposited Membrane Atmospheric Water Harvesting Unit

Water scarcity is certainly one of the most relevant and worrying issues of the 21st century, affecting billions of people and limiting their access to safe water. Atmospheric water harvesting technologies have been under research and development for decades now to relieve — and eventually overcome — water scarcity in affected areas. Such technologies have been drawing increasing attention in the last years, due to the possibility of employing renewable energy sources to power the main physical processes involved. This study analyzes the performance of a novel membrane based harvesting unit under various operating conditions and using multiple structural designs. The lab-scale prototype (featuring a single membrane of surface area 160 cm2, active contact area 123.5 cm2) yield was as high as 354 ml/day of water, with an average of 155 ml/day, corresponding to water harvesting rates of 22.13 kg/m2/day and 9.69 kg/m2/day, respectively. When designing a commercial scale plant, as well as during its operation, it is important to have an easy tool to predict the production of water from the system: in this sense, machine learning models represent a much more convenient solution than complex and computationally expensive physics-based models. This study compares the performance and stability of models created using four machine learning techniques. The techniques applied and discussed are Support Vector Regression, Gradient Boosting Regression, Multilayer Perceptron, and Generalized Regression Neural Networks, all of which are particularly efficient in describing non-linear relationships even when the data set and the number of features per sample are of moderate dimensions. The first three methods achieved noteworthy results, thus proving suitable for employment in the regulation of an atmospheric water harvesting plant featuring the analyzed technology.