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

Area of Science:

• Environmental Science
• Chemical Engineering
• Artificial Intelligence

Background:

• Electrochemical water treatment is crucial for global water scarcity but faces optimization challenges due to limited resources and expertise.
• Current methods often require cloud connectivity, raising concerns about data privacy and accessibility.

Purpose of the Study:

• To develop an intelligent on-device platform for optimizing electrochemical water treatment.
• To integrate large language models (LLMs) with electrochemical process knowledge directly on edge devices.

Main Methods:

• Deployment of LLMs on edge devices (e.g., Raspberry Pi) for real-time data analysis and prediction.
• Integration of theoretical knowledge with on-device processing, eliminating the need for cloud connectivity.
• Validation against 320 published studies to assess predictive accuracy and performance.

Main Results:

• Achieved a 60% reduction in hallucination rates compared to traditional methods.
• Maintained high predictive accuracy (R² > 0.80) for effluent concentration and energy consumption.
• Significantly improved prediction accuracy for applied current from 0.03 to 0.63, crucial for desalination.

Conclusions:

• The on-device platform makes advanced water treatment intelligence feasible in resource-limited, decentralized environments.
• This approach ensures data privacy and accessibility by processing data locally.
• Enables effective electrochemical water treatment optimization even with incomplete sensor data and limited computational resources.