When Water Data Is Scarce: Teaching AI to Listen to Science

Many of the water decisions that matter most—tracking contamination in seawater, understanding how rocks and soils release chemicals, or predicting a river’s response after a storm—happen with fewer measurements than anyone would like. This episode looks at a machine-learning idea built for that reality: instead of asking AI to learn everything from scratch, start with the scientific rule we already trust, then train the AI to learn what that rule misses. We unpack Knowledge-based Residual Learning, or KRL, through everyday analogies and water-relevant examples, including radioactive chemical measurements in seawater near Fukushima and chemical weathering in soils. The promise is not “AI replaces science,” but something more useful: AI can become a careful assistant when field data are expensive, patchy, and hard-won.

Citation: Guanjie Zheng, Chang Liu, Hua Wei, Porter Jenkins, Chacha Chen, Tao Wen, and Zhenhui Li. “Knowledge-based Residual Learning.” Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence (IJCAI-21), 2021, pp. 1653–1659.

Disclosure: this Waterlines episode package is written for production with AI-generated voices.