Reproduce “Predicting Change, Not States: An Alternate Framework for Neural PDE Surrogates” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Anthony Zhou, Amir Barati Farimani
Researchers from Carnegie Mellon University introduce "derivative prediction," an alternative framework for neural PDE surrogates where models learn temporal derivatives for integration by external ODE solvers. This approach consistently reduces rollout errors and increases stability across various PDEs and architectures, offering enhanced flexibility in inference compared to direct state prediction.
Key insights
- Decoupling the learning of spatial dynamics from temporal integration results in a "smoother" loss landscape, simplifying the neural network's learning task.
- The framework provides significant inference flexibility, enabling the use of various ODE integrators and adaptive time-stepping without model retraining.
- Training models to predict derivatives allows for more effective utilization of full-resolution temporal data, avoiding the need for downsampling common in state prediction.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.