Abstract
Emerging artificial intelligence for science (AI-for-Science) algorithms, such as the Fourier neural operator (FNO), enabled fast and efficient scientific simulation. However, extensive data transfers and intensive high-precision computing are necessary for network training, which challenges conventional digital computing platforms. Here, we demonstrated the potential of a heterogeneous computing-in-memristor (CIM) system to accelerate the FNO for scientific modeling tasks. Our system contains eight four-kilobit memristor chips with embedded floating-point computing workflows and a heterogeneous training scheme, representing a heterogeneous CIM platform that leverages precision-limited analog devices to accelerate floating-point neural network training. We demonstrate the capabilities of this system by solving the one-dimensional Burgers' equation and modeling the three-dimensional thermal conduction phenomenon. An expected nearly 116 times to 21 times increase in computational energy efficiency was achieved, with solution precision comparable to those of digital processors. Our results extend in-memristor computing applicability beyond edge neural networks and facilitate construction of future AI-for-Science computing platforms.