Predicting Change, Not States: An Alternate Framework for Neural PDE Surrogates
The paper explores an alternative method for neural surrogates to solve partial differential equations (PDEs). Instead of predicting the next state directly, this framework trains a neural model to predict the temporal derivative. This allows an external ODE integrator to advance the solution over time during inference, offering better accuracy and stability in finely-discretized regimes and enabling flexible time-stepping.