Using Fourier Neural Operator (FNO) for Spatial Compression​

A Brief Synopsis

Similar to my GPLaSDI project at Lawrence Livermore National Laboratory, I explored the application of Fourier Neural Operator (FNO) for spatial compression. Instead of performing operator learning, the FNO used compresses the data by taking fourier transform and then learns the evolution of the dominant modes through system id methods. Pre-eliminary results on the standard 1D Burgers equations showed a speed up of about 4x on the training process while achieving a similar accuracy. The results to this combination of FNO + LaSDI work can be found here. The corresponding implementation of this work can be found here.