Single-shot prediction of parametric partial differential equations

Khalid Rafiq, Wenjing Liao, Aditya G. NairarXiv 2025 · DOI · arXiv

Abstract

We present Flexi-VAE, a novel data-driven framework designed to predict the future states of complex physical systems governed by parametric partial differential equations (PDEs) in a single "shot." Unlike traditional iterative solvers or sequence-based deep learning models, Flexi-VAE evolves system dynamics within a low-dimensional latent space, bypassing slow time-stepping and mitigating error accumulation. Our results on 1D viscous Burgers’ and 2D advection–diffusion equations demonstrate up to 90x speedups on GPU and superior stability in extrapolation tasks compared to standard autoencoder-LSTM architectures.

Introduction

Simulating physical phenomena—such as fluid flow, heat transfer, or wave propagation—requires solving Partial Differential Equations (PDEs). Traditional numerical solvers, like Finite Element Methods, are computationally expensive because they must calculate the system's state step-by-tiny-step over time.

In recent years, researchers have turned to "Reduced-Order Models" (ROMs) using Autoencoders to compress these high-dimensional states into a small "latent space." However, most existing models still rely on iterative sequence models (like LSTMs) to move forward in time. This creates two major challenges: error accumulation (the "drift" problem) and computational bottlenecks for long-term forecasts. Flexi-VAE addresses these by predicting the state at any future time $\tau$ directly from the current state in one go, even when underlying physical parameters change.

Architecture and Methodology

Flexi-VAE combines a Variational Autoencoder (VAE) with a parametric neural propagator. An encoder compresses the physical state into a latent vector $z(t, \zeta)$. A propagator then takes this vector, along with the desired time offset ($\tau$) and physical parameters ($\zeta$), and maps it to a new latent vector representing the future state. Finally, a decoder turns that latent vector back into a full-scale physical simulation.

Overview of the Flexi-VAE architecture for single-shot prediction.

The authors explored two propagator strategies: the Direct Concatenation Propagator (DCP), which appends the time and parameters to the latent vector, and the Positional Encoding Propagator (PEP), which uses sinusoidal embeddings similar to those in Transformers.

Geometric Intuition and Stability

A key innovation of this work is the use of Jacobian spectral analysis to study the "geometry" of the latent space. The researchers found that the propagator does not just move the state forward in time; it steers the system toward regions of the latent manifold where the decoder is less sensitive to noise. These "stable regions" make the predictions significantly more robust to perturbations than those generated by standard encoders.

Geometric interpretation of the latent space evolution and decoding process.

Key Results and Discussion

The model was validated on the 1D viscous Burgers’ equation and the 2D advection–diffusion equation, testing both interpolation and extrapolation (unseen Reynolds numbers and time horizons).

  • Massive Speedups: Flexi-VAE achieved a 50x speedup on CPU and a 90x speedup on GPU for large temporal shifts. Because it is "single-shot," the inference time is independent of the length of the forecast.
  • Superior Generalization: While LSTMs began to drift and lose accuracy at long horizons due to compounding errors, Flexi-VAE remained stable. Surprisingly, the simpler DCP method outperformed the more complex PEP in generalization, fostering a "disentangled" latent space where dimensions naturally aligned with physical features like peak position.
  • Theoretical Grounding: The paper provides a formal proof showing that if PDE solutions lie on a low-dimensional manifold, a neural network can approximate the evolutionary operator with an error that scales based on the intrinsic dimension of the system, rather than the high-dimensional grid size.

Conclusion

Flexi-VAE offers a robust, high-speed alternative to traditional PDE solvers and iterative deep learning models. While currently purely data-driven, future iterations could incorporate "physics-informed" constraints to explicitly enforce conservation laws. Additionally, scaling the framework to high-dimensional, chaotic systems like 3D turbulent flows remains a promising direction for future research.

Summary updated

BibTeX
@misc{rafiq2025single,
  author = {Khalid Rafiq and Wenjing Liao and Aditya G. Nair},
  title = {{Single-shot prediction of parametric partial differential equations}},
  year = {2025},
  eprint = {2505.09063},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2505.09063},
}
Figure