Adsorption remains one of the most effective methods of decontamination of water. Adsorption processes are governed by multiple factors which contribute to the overall efficiency of the process. These include process conditions such as temperature, pH, the concentration of pollutants, and competing ions. The adsorbate properties, such as speciation, polarity, kinetic diameter, and ionic sizes, also affect adsorption performance. The adsorbent properties also play a critical role in assessing the suitability of an adsorbing material. This includes surface areas, pore volumes, chemical compositions, surface charges, etc. The complexity of the interaction between all these parameters makes it cumbersome or near impossible to predict with appreciable precision the performance of an adsorption system. Machine learning provides an opportunity for developing models for concise prediction of adsorption efficiencies for different materials. This chapter discusses the principles of various machine learning models and their application in the adsorption of pollutants from water.
TopIntroduction
Water pollution and access to clean water are global problems, especially in the 21st century. This is due to exponential growth industrialization with concomitant rise in human population. Among the widespread water contaminants are heavy metals and dyes. These are known to be toxic to both aquatic and terrestrial living organisms (Mitra et al., 2022). This explains the abundance of data and continuing research on water treatment technologies. These include; chemical oxidation, ion exchange, ozonation, Fenton's reagent, membrane filtration, electro-kinetic coagulation, irradiation, electrochemical degradation, and adsorption among others (Butler et al., 2011; Nidheesh et al., 2013; Samarghandi et al., 2020; Venkatesh et al., 2017; Dzoujo et al., 2022). All these methods have their inherent limitations. However, adsorption still remains the most popular and effective technique for removal of contaminants from water. The effectiveness of an adsorption process, measured by the percent removal (%R) of the contaminant of interest, is dependent on the process conditions such as temperature, pH, competing ions, concentration of pollutant and amount of adsorbing materials. The %R is also a function of the adsorbent properties such as surface area, pore structure, pH, functional group density, acidity and basicity among other characteristics (Zhu et al., 2019). Since all these variables may simultaneously affect the percent removal, it is time consuming and expensive to undertake experiments evaluating the role of each of these parameters for optimization. Artificial intelligence provides a powerful tool for interrogating the interplay of multiple parameters and determining the most impactful parameters and their contribution to the observed outcomes. Machine learning (ML) is presently used in almost all domains of science, medicine, engineering, agriculture, computer science and water treatment is not an exception. This is due to its simplicity, reliability and rapidity (Çelekli et al., 2013). Various computing methods have been explored to solve problems in adsorption science. This chapter highlights the principles of different machine learning models and their applications in predicting adsorption data for the removal of heavy metals and dyes from water.
Machine Learning Concepts
The Fourth Industrial Revolution (4IR or Industry 4.0) has resulted in an overabundance of data, hence the phrase “the age of data”, where all that is around us is somehow connected to some data source, and our lives are constantly being digitally captured (Cao, 2017; Sarker, 2021; Sarker et al., 2021). This data can be either (1) structured, having a well-defined structure that conforms to some standard data model such as relational databases, (2) unstructured, meaning there is no pre-defined data format for example, sensor data, word processing documents, and PDF files, or (3) semi-structured, where it has some organizational properties but in the whole it may not conform to standard data models, for example, HTML, XML, JSON documents, NoSQL databases, among others (Sarker, 2021). Therefore, to derive any intelligence from all this data, artificial intelligence (AI) expertise, and more specifically, knowledge of machine learning (ML) techniques is of paramount importance.
Machine learning is a sub-category of AI that focuses on two interrelated questions: how to design computers capable of automatic experiential improvement and the discovery of the statistical-computational-information-theoretic laws governing all learning systems such as computers, human beings, and organizations (Jordan & Mitchell, 2015). More formally, “A computer program is said to learn from experience E with respect to some class of tasks T and performance measure P, if its performance at tasks in T, as measured by P, improves with experience E” (Mitchell, 1997). ML systems can be classified along many dimensions. One classification considers three dimensions: the underlying learning strategies which look at the amount of inference the learner performs on the information provided; the type of knowledge or acquired by the learner based on rules of behavior, description of physical objects, problem-solving heuristics, and so on; and the domain of application (Jordan & Mitchell, 2015). In this study, however, we consider the more traditional approach to classifying ML models or algorithms as more appropriate. This is discussed further in the following section.