Introduction
The automated discovery of governing partial differential equations (PDEs) from experimental data is a cornerstone of modern data-driven physics. While methods like Sparse Identification of Nonlinear Dynamics (SINDy) have proven effective for clean or slightly noisy data, they often fail when faced with high-magnitude outliers—sparse but significant errors common in real-world sensor measurements. This paper introduces Double Low-rank Sparse Regression (DLrSR), a framework designed to robustly identify governing equations even when data is heavily corrupted.
The DLrSR Framework
The core innovation of DLrSR lies in exploiting the inherent low-rank structure of both the physical state data and its partial derivatives. The framework operates in three distinct stages to isolate the underlying physics from noise and outliers.
- Data Cleaning (Robust PCA): The noisy observation matrix $U$ is decomposed into a clean, low-rank matrix $Z$ and a sparse outlier matrix $E_1$.
- Library Construction: A library of candidate terms (e.g., $u^2, u_x, u_{xx}$) is built using the cleaned data $Z$.
- Robust Discovery (LrSTR): A new algorithm, Low-rank Sequential Threshold Ridge regression (LrSTR), solves for PDE coefficients. It simultaneously identifies and removes outliers ($E_2$) that may persist in the time-derivative data ($U_t$), ensuring the final regression fits the physics rather than the noise.
Low-Rank Structure and Robustness
A key insight of this work is that physical processes are not random; they exhibit dominant patterns that result in low-rank data matrices. By extending this observation to the derivatives of the data, DLrSR provides a dual-layer filter against corruption.
The authors demonstrate that outliers are far more disruptive than standard Gaussian noise. A single erroneous sensor reading can significantly bias standard sparse regression. DLrSR mitigates this by using the Augmented Lagrangian Method (ALM) to balance nuclear norms (promoting low-rankness) with $\ell_1$ and $\ell_0$ norms (promoting sparsity in both outliers and the final equation).
Experimental Results
The framework was evaluated on seven canonical models, including standard PDEs (Burgers, KdV, NLS, KS) and parametric PDEs with time-varying or spatially dependent coefficients. To simulate extreme conditions, the researchers added 90% sparse outliers to the datasets.
- Burgers' Equation: DLrSR achieved a discovery error of only 3.084% under heavy corruption, whereas standard Sequential Threshold Ridge (STR) regression failed completely.
- Parametric Discovery: For the spatially dependent Kuramoto-Sivashinsky equation, DLrSR was the only method capable of correctly identifying the governing terms and coefficients, maintaining an error rate of approximately 13.9%.
- Clean Data: The method remains competitive even with perfect data, proving that the low-rank assumption does not compromise accuracy in ideal scenarios.
Conclusion
DLrSR represents a significant step forward in making data-driven PDE discovery practical for real-world applications where data quality is often compromised. While the method is computationally more intensive than simpler regression techniques due to repeated Singular Value Decompositions (SVD), its robustness to extreme outliers is a critical advantage. Future work may involve integrating this low-rank approach with deep learning architectures, such as Physics-Informed Neural Networks (PINNs), to handle more complex geometries and higher-dimensional systems.
Summary updated