Towards domain-adapted large language models for water and wastewater management: methods, datasets and benchmarking
Boyan Xu, Guanlan Wu, Zihao Li, Guangming Xu, Huabin Zeng, Rui Tong, How Yong Ng
Abstract
Large language models (LLMs) have shown significant promise for water and wastewater management. However, current foundation models are not yet reliable. This Perspective outlines a pathway for customizing foundation models into WaterGPTs (specialized LLMs for water and wastewater management). We present key methodologies for adapting foundation models into WaterGPTs, including prompt engineering, knowledge and tool augmentation, and fine-tuning, and they are illustrated through representative examples. Then, we highlight the importance of diverse and ethically sourced datasets to customize foundation models, and we propose strategies for efficiently extracting high-quality information to customize foundation models. Further, we advocate for the development of a secure, informative, and dynamic evaluation benchmark that will guide the creation of more reliable WaterGPT. To illustrate practical LLM deployment in water sectors, we envision a specialized WaterGPT in wastewater treatment plants, which could integrate specific biological/chemical knowledge and advanced tools to manage intricate processes of activated sludge. Collectively, we aim to lower barriers for non-AI water-domain-specific experts and bridge the gap between experimental and computational research in water and wastewater management.