NEXT: Physics-Informed Neuro-Spectral Exponential Time Differencing Architectures
NEXT targets the instability that appears when neuro-spectral PDE models meet stiff equations.
The paper proposes combining NeuSA-style spectral solution representations with high-order exponential time differencing. Its key move is to integrate the stiff linear part exactly with matrix exponentials, while a neural network handles the remaining nonlinear term. In benchmark stiff PDEs, the authors report NEXT stays stable and accurate where NeuSA diverges numerically. They also show it can handle inverse problems with sparse data, including unknown parameters or boundary conditions. ArXiv · AI/CL/LG's note
The paper proposes combining NeuSA-style spectral solution representations with high-order exponential time differencing. Its key move is to integrate the stiff linear part exactly with matrix exponentials, while a neural network handles the remaining nonlinear term. In benchmark stiff PDEs, the authors report NEXT stays stable and accurate where NeuSA diverges numerically. They also show it can handle inverse problems with sparse data, including unknown parameters or boundary conditions. ArXiv · AI/CL/LG's note
score 4