Application of machine learning and knowledge-based systems to support decision-making on fate and behaviour of engineered nanoparticles in aqueous environment

Abstract

In recent decades, modern science has transformed due to the recognition and usage of engineered nanoparticles (ENP), which are substances that have a peripheral dimension in the geometric range of 1 to 100 nm. These materials have evolved into multifunctional materials that are used for engineering and innovations in a wide range of fields, including agriculture, medicine, food, industry, biomedical, and energy. Apart from their unique functionality and numerous benefits, the surge in production of multi-enabled nano-products and their emission in ecology systems, specifically to aquatic systems has raised serious environmental concerns about the potential deleterious effects of ENPs on the aquatic biota. So far, to address the existing environmental safety concerns, a volume of experimental data has been generated using natural freshwater and like systems for characterisation of the ENP colloidal stability. However, this data is knowledge-poor, heterogeneous, highly multifaceted (not easily discernible data variables relationship), and uncertain (due to the multiplicity of data); thus, highly challenging to be utilised to support the decision-making. Therefore, this work describes the application of data modelling techniques for the development of computer-based intelligent systems to support decision-making in dealing with the fate and behaviour of ENPs. At the same time, the study aims to provide a coherent understanding of the mechanisms and interactions that underpin these ENP transformation behaviours in aqueous environments. The field of data modelling has gained prominence across various fields including numerous environmental domains with the advancement of artificial intelligence (AI) research, digital computers, and big data. The modelling techniques of interest in this study included machine learning (ML) (i.e., adaptive neuro-fuzzy inference system (ANFIS), artificial neural network (ANN), support vector regression (SVR), random forest regression (RFR), k-nearest neighbour (KNN), extreme gradient boosting (XGBoost) and multi-linear regression (MLR)) and knowledge-based system (KBS) (i.e., fuzzy logic, semi-quantitative analysis). The results showed that ML was quite useful for modelling heterogeneity and non-linear data. It also revealed that diverse ENP transformation processes are influenced by variant parameters and that significant variables reported experimentally are not fundamentally good predictive variables. The RFR algorithms had the highest performance with the coefficient of determination (R2) and Nash-Sutcliffe efficiency (NSE) greater than 0.80 and 0.70, respectively for predicting the dynamic aggregation of ENPs. Furthermore, to predict the dissolution of nZnO the models that performed the best were the RFR and XGBoost algorithms with R2 values of 0.85 and 0.92, respectively. Overall, ML techniques of RFR, XGBoost, SVR, and ANN models yielded satisfactory to a very good level of accuracy in predicting both the aggregation and dissolution of ENPs. However, MLR showed poor performance, for both processes an indication of no underlying linear relationship between the model inputs and output. In addition, to ML algorithms demonstrating high prediction accuracy, and meta-analysis aiding to quantitatively evaluate highly heterogeneous data from multiple literature sources: to account for the scarcity of quantitative data the domain knowledge was encoded using rules and scores to develop intelligent KBS. These included computer-based semi-quantitative analysis integrated with decision tree classifiers (SQADTC) and fuzzy decision-making systems (FDMS). SQADTC used several weights/scores allocated to different factors (inputs) or linguistic variables and their sub-level (intermediate outputs). The functionality of SQADTC was illustrated using worked case studies of silver (nAg), nZnO, and nTiO2. The results demonstrate that our proposed model can be highly effective and valuable for preliminary screening of the exposure of ENPs. SQA application is relatively cost-effective and easy to use since no software or computational tools are required. In addition, non-experts can easily understand the hierarchical nature, Boolean logic, and visual representations of DTCs; which is highly valuable given that testing each variation of ENPs is tedious and associated with high cost. Furthermore, the FDMS constituted 321 (three hundred and twenty-one) if-then conditional statements in the fuzzy inference system. Modelling results using FDMS in the case studies of nTiO2 and nZnO demonstrated that the representation of qualitative knowledge by fuzzy sets and its application as very successful in handling partial truth information. FDMS provides flexibility to reduce bias and integrate the uncertainty that arises with the modelling of expert intuitions or perceptions. FDMS was able to replicate human-like reasoning using natural language in complex scenarios with no sharp boundaries, which makes the model ideal in various real-world scenarios. Overall, this thesis work offers the application of ML and KBS as a basis to maximise and leverage accessible data (structured input-output data pairs and unstructured expert knowledge) to support ENP monitoring, initial screening, and exposure assessment. The developed decision support system could aid to reduce the costs associated with experimental testing and support the establishment of robust frameworks for nano-safety. This is necessary to balance the advancements in nanotechnology and long-term environmental protection. Additionally, the efficiency of developed models can be extended to other ENPs and readily scaled when new and more information becomes accessible without having to reconstruct the frameworks of these models.