TL;DR
This paper introduces a novel framework that synergistically combines data-driven methods with data assimilation to simultaneously identify governing physical processes and estimate their associated model parameters. This approach overcomes limitations of existing data-driven techniques by handling nonlinear model parameters, making it applicable to a wider range of complex real-world problems.
What Problem Does This Paper Address?
Accurately modeling complex physical systems often requires understanding the underlying physical processes and their governing equations. While data-driven methods have shown promise in discovering these equations, they typically assume model parameters are known or have simple linear dependencies. This assumption severely limits their applicability to real-world scenarios where parameters are often uncertain and exhibit nonlinear relationships. Traditional data assimilation methods, on the other hand, require a known model structure, creating a Catch-22 when the governing processes themselves are unclear. This paper addresses this critical gap by developing a unified framework that can both identify the dominant physical processes and simultaneously estimate the uncertain, potentially nonlinear, parameters within their governing equations.
Methods
The core of the proposed method lies in a two-stage process:
- Data-Driven Discovery: Spatiotemporal measurement data is split into training and testing sets. A data-driven approach, using sparse regression techniques, is employed on the training data to learn a candidate library of physical processes and their empirical models. This stage aims to identify which processes are active and select appropriate model forms.
- Data Assimilation for Parameter Estimation: A prediction error is calculated using the learned equation on the testing data. This prediction error, along with the training data, is then used within a data-assimilation framework (specifically, a gradient-based method) to iteratively update and estimate the uncertain model parameters. This process continues until convergence, effectively refining the model parameters based on how well the learned equation predicts unseen data.
The paper also details the use of finite difference schemes for calculating necessary derivatives from the spatiotemporal data and a polynomial smoothing technique to preprocess noisy data before derivative calculation.
Datasets
The study utilizes synthetically generated spatiotemporal measurement data derived from a benchmark problem of contaminant solute transport in subsurface formation, simulated using the MT3DMS software. Three distinct scenarios were designed:
- Scenario 1: Advection (ADV) and Dispersion (DIS).
- Scenario 2: ADV, DIS, and Freundlich sorption (F-SORP).
- Scenario 3: ADV, DIS, and Langmuir sorption (L-SORP).
These scenarios were further tested with varying levels of synthetic data noise (1%, 5%, and 10%) to evaluate the robustness of the proposed method.
Key Insights and Ideas
- Bridging the Gap: The most significant insight is the successful integration of data-driven discovery with data assimilation. This combination allows for tackling problems where both the governing equations and their parameters are unknown, a common scenario in complex physical systems.
- Handling Nonlinearities: The framework's ability to handle nonlinear model parameters is crucial. Existing data-driven methods often struggle with this, but this approach effectively incorporates them into the learning process.
- Robustness to Noise: The method demonstrates a degree of robustness to measurement noise, especially when combined with data preprocessing techniques like polynomial smoothing. While accuracy decreases with higher noise levels, the core identification and estimation capabilities remain functional.
- Parsimonious Models: The approach naturally leads to parsimonious models by discarding terms with near-zero coefficients, simplifying the discovered governing equations.
Key Results
The proposed method successfully identified the dominant physical processes and estimated the model parameters for all three solute transport scenarios.
- Process Identification: The method accurately distinguished between scenarios with different combinations of advection, dispersion, and sorption processes. For instance, it correctly identified that sorption was present in scenarios 2 and 3, and distinguished between the Freundlich and Langmuir sorption models.
- Parameter Estimation: The estimated parameters (e.g., Freundlich exponent $\alpha$, Langmuir constant $K_l$) were close to their true values, with small standard deviations, particularly in clean data scenarios. For example, in scenario 2, the estimated $\alpha$ had a mean of 0.700 with a standard deviation of 0.0031.
- Impact of Noise: With clean data, the learned equations showed high accuracy. As noise levels increased to 1%, 5%, and 10%, the accuracy of the learned equations and parameter estimates generally decreased, but the method still provided reasonable results, especially after data smoothing. For instance, in Table 1, the learned equations for ADV and DIS show small deviations in coefficients across different noise levels.
Limitations and Future Work
- Computational Cost: While the method is effective, the iterative nature of data assimilation, especially with multiple initial samples, can be computationally intensive.
- Data Requirements: The method relies on spatiotemporal measurement data, and the accuracy of derivative calculations is sensitive to data quality and density, particularly near boundaries.
- Future Work: The authors suggest exploring other data-driven methods (e.g., LASSO) to potentially improve performance and developing more sophisticated numerical schemes for learned PDEs to further refine coefficients and parameters. They also note that the method is amenable to problems involving coupled equations and constitutive relationships, hinting at broader applications. The challenge of data noise is acknowledged, suggesting further research is needed in this area.
Summary updated