Memristive floating-point Fourier neural operator network for efficient scientific modeling.

Area of Science:

• Artificial Intelligence for Science (AI-for-Science)
• Computational Science
• Materials Science

Background:

• AI-for-Science algorithms like Fourier neural operators (FNOs) offer efficient scientific simulation capabilities.
• Conventional digital computing faces challenges with the extensive data and high-precision computing demands of FNO training.

Purpose of the Study:

• To demonstrate the potential of a heterogeneous computing-in-memristor (CIM) system for accelerating AI-for-Science tasks.
• To leverage precision-limited analog devices within a CIM platform for efficient neural network training.

Main Methods:

• Developed a heterogeneous CIM system comprising eight 4-kilobit memristor chips with embedded floating-point computing.
• Implemented a heterogeneous training scheme to accelerate floating-point neural network training.
• Applied the system to solve the 1D Burgers' equation and model 3D thermal conduction.

Main Results:

• Achieved a significant increase in computational energy efficiency, ranging from 116x to 21x.
• Maintained solution precision comparable to conventional digital processors.
• Successfully demonstrated the system's capability in solving complex scientific modeling problems.

Conclusions:

• The heterogeneous CIM system effectively accelerates AI-for-Science simulations, particularly FNO-based tasks.
• This approach extends the applicability of in-memristor computing beyond edge AI.
• The findings facilitate the development of future AI-for-Science computing platforms.