On-device artificial intelligence agent based on language models for electrochemical water desalination

Electrochemical water treatment is essential for tackling global water scarcity but remains difficult to optimize due to limited expertise and computing resources at many treatment facilities. Here, we introduce an intelligent on-device platform that combines electrochemical process knowledge with large language models deployed directly on edge devices such as Raspberry Pi. This system integrates theoretical understanding with real-time optimization, eliminating the need for cloud connectivity while ensuring data privacy and accessibility. Tested against 320 published studies, it achieves a 60 % reduction in hallucination rates and maintains high predictive accuracy (R² > 0.80) for key variables such as effluent concentration and energy, even with incomplete sensor inputs. Notably, prediction accuracy for challenging parameters, such as the applied current (the driving force for electrochemical water desalination), improves from 0.03 to 0.63. By bringing intelligence to the data rather than sending it to the cloud, this approach makes advanced water-treatment intelligence feasible in resource-limited, data-imperfect, decentralized environments where physics-based models cannot be deployed.