PriorIDENT: Prior-Informed PDE Identification from Noisy Data

The paper proposes PriorIDENT, a prior-informed weak-form sparse-regression framework that integrates compact physics priors into dictionary construction to robustly identify governing partial differential equations from noisy spatiotemporal data while outperforming existing baselines in accuracy and stability.

The Gist

Imagine you are a detective trying to figure out the rules of a game just by watching the players move around. You have a video of the game, but the camera is shaky, the lighting is bad, and there's static on the screen (this is your noisy data). Your goal is to write down the exact laws of physics that govern the game (the Partial Differential Equations or PDEs).

This paper, titled PriorIDENT, introduces a new, smarter way for detectives to solve this mystery.

Here is the breakdown of the problem and their solution, using simple analogies:

The Problem: The "Guessing Game" Trap

Traditionally, scientists tried to find these laws by throwing a giant net of every possible mathematical formula into a computer and seeing which ones fit the data.

• The Issue: Because the data is noisy (shaky camera), the computer gets confused. It starts picking formulas that look like they fit the noise but are actually nonsense. It's like trying to find a specific needle in a haystack, but the haystack is on fire, and you keep grabbing random pieces of straw because they look like needles for a split second.
• The Result: The computer finds "laws" that break the laws of physics (e.g., creating energy out of nothing).

The Solution: PriorIDENT (The "Smart Detective")

The authors propose a method called PriorIDENT. Instead of guessing blindly, they give the detective a set of clues (Priors) before they even start looking.

Think of it like this:

• Old Way: "Here is a bag of 10,000 random words. Write a sentence that describes this picture." (The computer might write "The cat ate the moon" because it fits the pixels, even though it's nonsense).
• New Way (PriorIDENT): "Here is a bag of words, but we know for a fact this is a story about a cat. So, we only give you words related to cats, fur, and mice. Now, write the sentence."

The paper uses three specific types of "clues" (Priors) depending on the type of system:

1. The "Energy Saver" Clue (Hamiltonian):

   • Analogy: Imagine a pendulum swinging. It never gains energy on its own; it just swaps back and forth between height and speed.
   • The Trick: The computer is told, "You are only allowed to pick formulas that never create or destroy energy." This stops the computer from inventing magic physics.

2. The "Traffic Flow" Clue (Conservation Law):

   • Analogy: Imagine cars on a highway. If 10 cars enter a tunnel, 10 cars must exit. Cars don't just vanish or appear out of thin air.
   • The Trick: The computer is told, "You are only allowed to pick formulas that look like traffic flow (things moving in and out)." This ensures the math respects the rule that matter is conserved.

3. The "Rolling Downhill" Clue (Energy Dissipation):

   • Analogy: Imagine a ball rolling down a hill. It slows down due to friction and eventually stops at the bottom. It never rolls up the hill on its own.
   • The Trick: The computer is told, "You are only allowed to pick formulas that look like things slowing down or settling." This prevents the computer from inventing systems that get more chaotic over time when they should be calming down.

The Secret Weapon: The "Smooth Lens" (Weak Form)

Even with the right clues, looking at a shaky video is hard. If you try to measure how fast a car is moving by looking at two blurry frames, the math gets messy.

The paper uses a technique called Weak Formulation.

• Analogy: Instead of trying to measure the speed of a single car at a single instant (which is noisy), imagine you are looking at the average flow of traffic over a whole minute.
• How it works: They use a "smooth lens" (a mathematical test function) to blur out the tiny, jittery errors in the data. This makes the signal clear and the math stable, even if the original data is very noisy.

The Result: A Cleaner, Smarter Detective

The paper tested this method on famous physics problems:

• The Three-Body Problem: Predicting how three planets orbit each other.
• Shallow Water: Predicting how waves move in the ocean.
• Diffusion: Predicting how a drop of ink spreads in water.

The Outcome:
When the data was very noisy (like a bad video), the old methods failed and picked the wrong rules. PriorIDENT, however, kept finding the correct laws. It didn't just guess; it used the "clues" to filter out the nonsense and the "smooth lens" to ignore the static.

In a Nutshell

PriorIDENT is a new tool for discovering the laws of nature from messy data. It works by:

1. Restricting the search: Only looking for math that makes physical sense (like saving energy or conserving matter).
2. Smoothing the noise: Ignoring the tiny jitters in the data to see the big picture.

It's the difference between a detective guessing randomly in a dark room and a detective who knows the rules of the game and has a flashlight.

Technical Summary

Here is a detailed technical summary of the paper "PriorIDENT: Prior-Informed PDE Identification from Noisy Data".

1. Problem Statement

The discovery of governing Partial Differential Equations (PDEs) from spatiotemporal data is a critical task in scientific machine learning. However, existing data-driven methods face two fundamental challenges:

• Noise Amplification: Numerical differentiation of noisy data amplifies measurement errors, leading to unstable derivative estimates.
• Model Ambiguity & Overfitting: The use of large, over-complete feature dictionaries (containing all possible polynomial and derivative terms) creates redundancy. This often leads to the selection of non-physical terms that merely fit the noise rather than the underlying physics, resulting in models that violate fundamental conservation laws or thermodynamic principles.

Current approaches, such as sparse regression (e.g., SINDy) or neural PDEs, often treat the dictionary construction as a generic process without embedding specific physical constraints, making them fragile under high noise levels.

2. Methodology: PriorIDENT

The authors propose PriorIDENT, a unified framework that integrates physical priors directly into the dictionary construction stage and utilizes weak-form formulations to handle noise. The method operates on the principle that candidate terms should be physically admissible by construction.

A. Core Framework

The identification pipeline follows a five-step sparse regression process (based on Subspace Pursuit), but with a critical modification in Step 1:

1. Prior-Constrained Dictionary Construction: Instead of generating a generic library of all possible terms, the dictionary is built by applying structure-preserving operators to a parameterized basis of latent physical objects (e.g., Energy, Flux, Hamiltonian).
2. Sparse Regression: The Subspace Pursuit (SP) algorithm identifies the sparse coefficient vector.
3. Trimming: Marginal features with negligible contributions are removed to prevent over-selection.
4. Model Selection: The optimal sparsity level is determined using the Reduction in Residual (RR) criterion.
5. Reconstruction: Final coefficients are estimated via least squares.

B. Three Specific Priors

The paper instantiates this framework for three distinct classes of physical systems:

1. Hamiltonian Prior (Skew-Gradient Structure):

   • Concept: Systems conserving energy and symplectic volume (e.g., oscillators, celestial mechanics).
   • Implementation: The dictionary is constructed as $D_{Ham} = \{ J\nabla \phi_\alpha \}$, where $J$ is the symplectic matrix and $\phi_\alpha$ are basis functions for the Hamiltonian $H$.
   • Effect: Enforces $\dot{z} = J\nabla H$, ensuring energy conservation and reducing the search space from independent equations to a single scalar energy functional.

2. Conservation-Law Prior (Flux-Form Representation):

   • Concept: Systems governed by continuity equations (e.g., fluid dynamics, shallow water).
   • Implementation: The dictionary is constructed as divergences of admissible fluxes: $D_{Flux} = \{ \nabla \cdot \phi_\alpha \}$. Coefficients are tied across spatial directions to enforce isotropy or specific flux symmetries.
   • Effect: Ensures the identified model strictly satisfies local mass/momentum conservation ($\partial_t u + \nabla \cdot F = 0$).

3. Energy-Dissipation Prior (Gradient-Flow Structure):

   • Concept: Systems evolving to minimize an energy functional (e.g., diffusion, phase separation).
   • Implementation: The dictionary is constructed via variational derivatives: $D_{GF} = \{ -\frac{\delta}{\delta u} \int \phi_\alpha dx \}$.
   • Effect: Restricts candidates to thermodynamically consistent operators, ensuring monotonic energy decay ($dE/dt \leq 0$).

C. Weak-Form Formulation

To mitigate noise sensitivity, the method employs a weak formulation. Instead of differentiating noisy data directly, derivatives are transferred to smooth test functions ($\psi$) via integration by parts:
$$\int \int u \mathcal{F}^*[\psi] \, dx \, dt = 0$$
This shifts the differentiation burden from the noisy state $u$ to the smooth test function $\psi$, significantly enhancing numerical stability.

3. Key Contributions

1. Novel Dictionary Design: A method to embed physical priors directly into the dictionary construction, ensuring all candidate features are physically admissible before regression begins.
2. Unified Framework: Integration of three distinct structural priors (Hamiltonian, Conservation, Gradient-Flow) with a weak-form sparse regression pipeline enhanced by trimming and residual-based model selection.
3. Robustness to Noise: Demonstrated ability to recover correct PDE structures and coefficients even under substantial noise (up to 50% or 100% noise-to-signal ratio), outperforming both strong-form and weak-form baselines without priors.

4. Experimental Results

The method was evaluated on canonical systems across the three prior categories:

• Hamiltonian Systems:
  • Harmonic Oscillator & Three-Body Problem: The method achieved a True Positive Rate (TPR) of 1.0 across all noise levels (up to 50%), whereas baselines failed significantly at high noise. It successfully reduced the 18 coupled equations of the three-body problem to a single scalar energy functional.
• Conservation Laws:
  • 1D Inviscid Burgers & 2D Shallow Water Equations: The method correctly identified flux terms (e.g., $(u^2)_x$) and preserved mass/momentum conservation. In the 2D Shallow Water case, the total relative error remained below 3% even at 50% noise, while baselines suffered from spurious term selection.
• Energy Dissipation:
  • Diffusion & Allen-Cahn Equations: The method correctly identified gradient flow structures (e.g., $u_{xx}$ and $u^3-u$) and maintained thermodynamic consistency. At extreme noise levels (100%), the prior-informed method maintained a median TPR of 1, while unconstrained baselines collapsed.

Performance Metrics:

• TPR (True Positive Rate): Consistently higher than baselines, often reaching 1.0 where baselines dropped below 0.5 under noise.
• Coefficient Recovery: Stable recovery of physical constants (e.g., diffusivity $\nu$, gravity $g$).
• Dynamics Preservation: Identified models reproduced spatiotemporal trajectories (e.g., orbital paths, wave propagation, phase separation) that remained qualitatively correct even when quantitative errors existed due to noise.

5. Significance

PriorIDENT addresses the "black box" nature of data-driven PDE discovery by enforcing physical interpretability at the algorithmic level. By constraining the hypothesis space to physically valid operators, the method:

• Reduces Degrees of Freedom: Drastically shrinking the search space, which improves conditioning and reduces the risk of overfitting.
• Ensures Physical Faithfulness: Guarantees that discovered models respect fundamental laws (conservation, energy dissipation, symplecticity) by construction, not just by post-hoc validation.
• Provides a Unified Route: Offers a robust, single framework capable of handling diverse physical regimes (conservative, dissipative, and mixed) from noisy data, bridging the gap between scientific modeling and data-driven learning.

In conclusion, the paper demonstrates that combining compact structural priors with weak formulations is a superior strategy for identifying governing equations from noisy data, yielding models that are both mathematically robust and physically meaningful.