The Global Minimum: Quantifying the Requirements for Extending AI-Powered Water Supply Optimisation Algorithms to New Applications and "Data-Poor" Env

AI is used to automate and optimise water sector operations in HICs (including optimising energy usage, detecting leaks and improving water quality in water distribution systems) with unparallelled levels of performance. However, whilst vast datasets are relied upon to train AI in HICs, many low-and-middle-income countries (LMICs) lack sufficient data to produce even basic coverage estimates for water supply and sanitation infrastructure.

The need to improve data availability in many LMICs is well documented. However, the requirements for bridging the gap between the rapid AI development in HICs and the insufficient enabling environments observed in LMICs have not been quantified. Without these estimates, there is no way to reliably propose appropriate actions and predict the associated costs necessary for LMICs to prepare themselves to embrace the significant benefits enabled by AI systems.

This is also a problem in HICs, because without an understanding of the relationship between AI performance and dataset characteristics it is difficult to determine whether existing systems have been overdesigned. These knowledge gaps also increase the difficulty of predicting AI performance such that systems can be relied upon to inform safety critical applications such as detecting drinking water contamination events.

Knowledge gaps will be investigated by developing AI algorithms which meet suitable performance benchmarks for applied case studies. Initial plans for water sector case studies involve developing anomaly detection algorithms for water distribution systems or algorithms to predict freshwater quality trends. Algorithms will be trained on large datasets which can be systematically reduced in size and quality to assist with modelling the effects of sparse data on algorithm performance.

The intention of this study is to improve the understanding of the relationship between dataset characteristics and the performance of AI algorithms to reduce system overspecification in HICs and to inform roadmaps for streamlining AI adoption in LMICs.