Machine learning-driven discovery of optimal designs for water electrolysis devices

Area of Science:

• Electrochemical Engineering
• Materials Science
• Artificial Intelligence

Background:

• Green hydrogen production via water electrolysis is crucial for decarbonization but faces challenges with energy-intensive designs.
• Current electrolyzer designs suffer from gas bubble entrapment and inefficient manual optimization processes.

Purpose of the Study:

• To introduce a machine learning (ML) strategy for the autonomous design of high-efficiency electrolyzer flow channels.
• To identify novel channel geometries that enhance bubble removal and improve overall electrolyzer performance.

Main Methods:

• Utilized a mixture-of-experts framework within an ML strategy to autonomously design electrolyzer flow channels.
• Performed parametric analysis to establish structure-performance relationships for array-type channel geometries.
• Employed data-driven screening as an efficient alternative to traditional computational fluid dynamics simulations.

Main Results:

• Identified an array-type channel geometry that significantly enhances bubble removal efficiency.
• Demonstrated a ~23% improvement in current density at 2 V in a prototype electrolyzer with the AI-optimized channel compared to conventional designs.
• Validated the consistent performance enhancement of the optimized design in scaled-up devices.

Conclusions:

• The ML-driven autonomous design strategy effectively identifies high-performance electrolyzer flow channel structures.
• The optimized array-type channel design offers a scalable pathway for next-generation electrochemical systems, improving green hydrogen production efficiency.
• This approach decodes complex relationships between topological features and multiphase transport for accelerated innovation in electrochemical devices.