no. 16 · Tuesday 17 November 2026
UNIDE: Cross-PDE Learning for Noise-Robust Discovery of Governing Equations
Abstract
Governing partial differential equations (PDEs) provide compact and interpretable descriptions of physical systems. Classical methods, such as sparse and symbolic regression, fit a new equation to each observed field: what is learned on one system is not carried to the next, and observation noise corrupts the derivative features they rely on. This paper introduces UNIDE (Unified Identification of Differential Equations), a data-driven approach that learns across PDE families to identify governing equations from new noisy trajectories. It integrates three components: a self-supervised denoiser trained without clean targets, an encoder shared across PDE families that predicts which candidate terms are active, and weak-form least squares with residual-based pruning that estimates their coefficients. We demonstrate UNIDE on synthetic one-dimensional trajectories from six PDE families with known term structures, including the Burgers and Kuramoto–Sivashinsky equations. We train one model per noise level and systematically compare UNIDE with SINDy, WSINDy, and PySR. The results show that UNIDE recovers the exact set of governing terms more often than all three at every noise level from $1\%$ to $80\%$. At $80\%$ noise, where the noise standard deviation is $0.8$ times that of the field, it recovers the exact terms for $92\%$ of held-out trajectories, whereas the best baseline recovers about $1\%$. UNIDE is also efficient: because its neural components are reused rather than refit, it identifies each new trajectory in $0.45$ seconds, $2.7$ times faster than WSINDy and $51$ times faster than PySR. These results demonstrate the potential of learning across PDE families for efficient equation identification under substantial observation noise.
- comments
- 30 pages, 9 figures
- fields
- Numerical Analysis