A neural operator framework for data-driven discovery of stability and receptivity in physical systems

This paper introduces an equation-free, data-driven framework using neural operators and automatic differentiation to automatically identify stability properties and optimal forcing responses in complex physical systems directly from observation data, bypassing the need for known governing equations.

The Gist

Imagine you are trying to understand how a complex machine works—like a car engine, a weather system, or even the human brain. Usually, to predict what will happen if you push a button or turn a knob, you need the blueprint (the mathematical equations) of that machine. You need to know exactly how the gears, pistons, and wires interact.

But what if you don't have the blueprint? What if the machine is too complex, or the equations are too messy to write down?

This paper introduces a clever new way to figure out how these machines behave without needing the blueprint. Instead, it uses a "digital twin" built entirely from watching the machine run.

Here is the breakdown of their method using simple analogies:

1. The Problem: The "Black Box"

Scientists often study systems by looking at how they react to small nudges.

• Stability: If I tap a pendulum, does it swing back to rest, or does it start spinning wildly?
• Receptivity: If I blow on a candle, which direction makes the flame flicker the most?

Traditionally, to answer this, you need the "equations of motion" (the blueprint). But for things like turbulent water, climate change, or brain activity, those equations are either unknown or too hard to solve.

2. The Solution: The "Digital Twin" (Neural Operator)

The authors built a Neural Network (a type of AI) that acts like a video game simulator.

• They feed the AI thousands of videos of the system moving (e.g., water flowing, air swirling).
• The AI learns to predict the next frame of the video based on the current one.
• Eventually, this AI becomes a "Digital Twin." It doesn't know the physics equations, but it knows exactly how the system behaves because it has "memorized" the patterns.

3. The Magic Trick: The "Microscope" (Automatic Differentiation)

This is the most creative part. Usually, an AI just predicts the future. But the authors realized they could use the AI as a mathematical microscope.

Imagine you are driving a car. You know the car is stable because if you turn the wheel slightly, the car turns slightly.

• The AI has learned the "rules" of the car so well that it can tell you: "If I nudge the steering wheel by 0.001 degrees, exactly how much will the car move?"
• The authors use a mathematical tool called Automatic Differentiation to ask the AI this question instantly.
• This allows them to extract the "Jacobian"—which is just a fancy word for the "local rulebook" of how the system reacts to tiny nudges at any specific moment.

4. The Results: Finding the "Weak Spots"

Once they have this "local rulebook" from the AI, they can run standard tests to find two critical things:

• The "Tipping Points" (Stability Analysis): They can see if a tiny nudge will die out (the system is stable) or grow into a disaster (the system is unstable). They found this works even when the system is behaving chaotically, where traditional math fails.
• The "Knobs to Turn" (Receptivity Analysis): They can identify exactly where and how to push the system to get the biggest reaction.
  • Analogy: Imagine a giant drum. You want to make the loudest sound possible. Where should you hit it? This method tells you the exact spot and the exact rhythm to hit the drum to get the maximum volume, even if you don't know the physics of the drum skin.

5. Why This Matters

The authors tested this on four different "machines":

1. Lorenz System: A simple weather model (like a chaotic butterfly).
2. Ginzburg-Landau: A model for how waves grow in fluids.
3. Channel Flow: Water flowing through a pipe.
4. Cylinder Flow: Air swirling around a pole (like a flag flapping in the wind).

In all cases, the AI learned the behavior from data alone and successfully predicted the "weak spots" and "instabilities" just as well as (and sometimes better than) traditional methods that required complex physics equations.

The Big Picture

Think of this framework as a universal translator.

• Old Way: You need to speak "Physics" (equations) to understand the system.
• New Way: You just need to speak "Data" (observations). The AI translates the data into a map of the system's behavior, showing us where it is stable, where it is fragile, and how to control it.

This is a huge step forward for fields like climate science (predicting storms), neuroscience (understanding brain seizures), and engineering (designing better airplanes), where the "blueprints" are often missing or too complicated to use.

Technical Summary

1. Problem Statement

Understanding how complex physical systems respond to perturbations—specifically determining their stability (whether they remain stable or transition to instability) and receptivity (how they respond to external forcing)—is a fundamental challenge in science and engineering.

• Limitations of Traditional Methods: Classical approaches like Linear Stability Analysis (LSA) and Resolvent Analysis rely on known governing equations. They require linearizing these equations around a base state to compute the Jacobian operator. This limits their application to systems where equations are unknown, poorly modeled, or highly nonlinear.
• Limitations of Existing Data-Driven Methods: Purely data-driven techniques like Dynamic Mode Decomposition (DMD) assume the system evolution can be approximated by a global linear operator. This assumption fails in strongly nonlinear regimes, leading to inaccurate eigenmode extraction and an inability to capture transient growth or optimal forcing structures in complex systems.
• The Gap: There is a need for a framework that can extract stability properties and input-output structures (eigenmodes and resolvent modes) directly from observational data without requiring explicit governing equations, while remaining valid in strongly nonlinear regimes.

2. Methodology

The authors propose a data-driven framework that utilizes a Neural Network (NN) as a dynamics emulator to approximate the system's local linear operator (Jacobian) via automatic differentiation. The workflow consists of four main stages:

A. Neural Network Emulator Training

• Architecture: The system state $q_n$ at time $t$ is mapped to the next state $q_{n+1}$ using a neural network $f_\theta$:
  $$q_{n+1} = f_\theta(q_n)$$
• Training Strategy: The network is trained using a rollout strategy (autoregressive training). Instead of minimizing only one-step prediction errors, the loss function minimizes the error over multiple time steps ($t$). This forces the network to learn long-term dependencies and correct for compounding errors, crucial for high-dimensional nonlinear systems.
• Differentiability: The trained NN acts as a differentiable surrogate for the classical numerical solver.

B. Jacobian Extraction via Automatic Differentiation

• Linearization: The discrete-time NN map is linearized around a base state (equilibrium point or steady state) $q_b$.
• Computation: The local Jacobian matrix $\mathbf{N}$ of the NN is computed using automatic differentiation:
  $$\mathbf{N} = \left. \frac{\partial f_\theta}{\partial q} \right|_{q_b}$$
• Relationship to Continuous Operator: The NN-based Jacobian $\mathbf{N}$ relates to the continuous-time linear operator $\mathbf{A}$ (the true Jacobian of the governing equations) via the exponential map:
  $$\mathbf{N} = \exp(\mathbf{A}\Delta t)$$
  This relationship allows the recovery of the continuous system's spectral properties from the discrete NN.

C. NN-Based Linear Stability Analysis

• Eigenvalue Decomposition: The eigenvalues ($\Lambda_N$) and eigenvectors ($V_N$) of the NN Jacobian $\mathbf{N}$ are computed.
• Mapping: The continuous eigenvalues $\sigma$ (growth rate $\lambda$ and frequency $\omega$) are recovered via:
  $$\Lambda_A = \frac{1}{\Delta t} \ln(\Lambda_N)$$
• Outcome: This identifies dominant unstable modes and stability boundaries without solving the governing equations.

D. NN-Based Resolvent Analysis

• Dimensionality Reduction: To avoid the ill-posedness of directly taking the logarithm of the Jacobian (due to the multivalued nature of the complex logarithm), the authors project the system onto a reduced subspace spanned by the top $r$ eigenvectors identified in the stability analysis.
• Weighted SVD: The resolvent operator $\mathbf{H}(\omega) = (-i\omega \mathbf{I} - \mathbf{A})^{-1}$ is approximated in this reduced basis. A weighted Singular Value Decomposition (SVD) is performed to identify:
  • Optimal Forcing Modes: The input patterns that yield the largest response.
  • Response Modes: The resulting system response.
  • Gain Distribution: The amplification factor across frequencies.

3. Key Contributions

1. Equation-Free Modal Analysis: The framework enables both linear stability and resolvent analysis purely from data, bypassing the need for governing equations or adjoint solvers.
2. Nonlinear Capability: Unlike DMD, which assumes global linearity, the NN emulator captures local nonlinear dynamics. The method successfully extracts accurate stability and receptivity characteristics even in strongly nonlinear regimes.
3. Bridging Discrete and Continuous: The paper rigorously establishes the mathematical link between the discrete NN Jacobian and the continuous system operator, allowing for the recovery of physical growth rates and frequencies.
4. Hybrid Approach: It combines the flexibility of deep learning (universal function approximation) with the interpretability of classical operator theory (eigenmodes and resolvent modes).

4. Results and Validation

The framework was validated on four systems of increasing complexity:

• Example 1: Lorenz-63 System (Chaotic ODEs)

  • The NN accurately predicted trajectories for unseen initial conditions.
  • The NN-derived Jacobian matched the analytical Jacobian element-wise at equilibrium points and arbitrary trajectory points, validating the automatic differentiation approach.

• Example 2: Complex Ginzburg-Landau (CGL) Equation

  • Linear vs. Nonlinear: Tested on both linear and nonlinear datasets.
  • Stability: The NN correctly identified the leading unstable eigenvalue and mode shapes in the nonlinear regime, whereas DMD failed completely (identifying spurious unstable modes).
  • Resolvent: The NN accurately predicted the gain curves and forcing/response modes near the dominant frequency, though higher-order modes showed some noise.

• Example 3: 2D Transitional Channel Flow (Navier-Stokes, Re=2000)

  • Stability: Successfully identified the dominant A-branch (wall) and P-branch (center) modes in both weakly and strongly nonlinear datasets.
  • Resolvent: Accurately predicted the peak gain frequency ($\omega = 0.31$) and the spatial structure of the most sensitive regions for perturbation amplification.
  • Comparison: DMD results were scattered and inaccurate in the nonlinear case, while the NN results remained robust.

• Example 4: 2D Cylinder Flow (Re=100)

  • Stability: Correctly identified the single unstable eigenvalue associated with the Hopf bifurcation and the corresponding spatial eigenmode (vortex shedding structure).
  • Resolvent: Despite the lack of a true steady-state solution (the flow is inherently unsteady), the analysis around the pseudo-equilibrium provided physical insights, identifying the "wavemaker" region downstream of the cylinder as the most sensitive to forcing.

5. Significance and Impact

• Broad Applicability: This methodology provides a general tool for analyzing complex, high-dimensional datasets in fields where governing equations are unknown or intractable, such as climate science, neuroscience, epidemiology, and fluid engineering.
• Overcoming Nonlinearity: It addresses a critical bottleneck in dynamical systems theory by enabling modal analysis in regimes where linearization assumptions (required by traditional methods) and global linearity assumptions (required by DMD) break down.
• Interpretability: By recovering physically meaningful structures (eigenmodes, resolvent gains), the method bridges the gap between "black-box" deep learning and interpretable physics-based modeling.
• Future Directions: The authors suggest that this approach can be enhanced by integrating autoencoders for dimensionality reduction and adapting to non-uniformly sampled data, potentially revolutionizing how stability and control strategies are designed for complex real-world systems.