🤖 machine learning

Learned Lagrangian Models of PDEs via Euler-Lagrange Residual Minimization

This paper introduces a novel method that directly learns continuous Lagrangian models for partial differential equations by minimizing Euler-Lagrange residuals via a mesh-free optimization-based integrator, enabling stable, long-range predictions with linear scalability and generalization to arbitrary boundary conditions without retraining.

Original authors: Lyra Zhornyak, Eric Forgoston, M. Ani Hsieh

Published 2026-05-11

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Lyra Zhornyak, Eric Forgoston, M. Ani Hsieh

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict the future path of a complex system, like a double pendulum swinging wildly or waves crashing across the ocean. In the world of physics and machine learning, we usually try to teach a computer the "rules" of how these things move. Once the computer learns the rules, we ask it to simulate what happens next.

The problem with most current methods is that they are like a clumsy accountant. They might get the math right for the very next second, but as they keep adding up the numbers over hours or days, tiny rounding errors pile up. Eventually, the computer's prediction breaks the laws of physics: energy might magically appear out of nowhere, or waves might stop moving. The simulation becomes physically impossible.

This paper introduces a new method called ELM (Euler–Lagrange Minimization) that acts more like a master architect than a clumsy accountant. Here is how it works, using simple analogies:

1. The "Learned Blueprint" (The Lagrangian)

Instead of teaching the computer to guess the next step in a sequence (like predicting tomorrow's weather based on today's), the authors teach the computer a blueprint called a "Lagrangian." Think of this as a master recipe that describes the energy and rules of the entire system at once.

The Innovation: Usually, these blueprints are written by human experts. Here, the computer learns the blueprint directly from data. It figures out the underlying physics on its own.

2. The "Local Patchwork" (The Integrator)

Once the computer has the blueprint, it needs to simulate the movement. Old methods try to solve the whole puzzle at once, which is slow and prone to errors.

The ELM Approach: Imagine you are trying to smooth out a wrinkled sheet of fabric. Instead of trying to fix the whole sheet at once, you take small, overlapping patches. You smooth out one patch, then move to the next, overlapping slightly with the previous one.
The Magic: ELM does this with "space-time patches." It looks at a small chunk of time and space, checks if the movement fits the learned blueprint, and makes tiny adjustments to make it fit perfectly. It does this locally, patch by patch, rather than trying to solve the whole universe at once.

3. "Near-Symplectic" (The Energy Guardian)

In physics, "symplectic" is a fancy word for "energy-conserving." A symplectic integrator is like a bank account that never loses or gains money by mistake; it only moves money around.

The Problem: Most AI models are like a bank account with a leaky faucet. Over time, energy leaks out (or floods in), and the simulation becomes garbage.
The ELM Solution: Because ELM is built on the "patchwork" method that strictly follows the blueprint's rules, it acts like a near-perfect vault. Even if the blueprint isn't 100% perfect, the method ensures that energy is conserved over long periods. The simulation doesn't drift into nonsense; it stays physically realistic for a very long time.

4. Why This Matters (The "Plug-and-Play" Physics)

The most exciting part of this paper is that the computer learns a continuous blueprint, not a grid-based one.

The Analogy: Imagine learning how water flows in a specific, square swimming pool. Most AI models learn the rules only for that square pool. If you put them in a round pool, they break.
ELM's Superpower: Because ELM learns the continuous rules (like the actual physics of water), you can take the same trained model and put it in a square pool, a round pool, or a pool with a weird wall in the middle. It doesn't need to be retrained.
Real Results: The authors tested this on:
- A chaotic double pendulum (keeping energy stable for 100,000 seconds).
- 1D and 2D waves (showing how waves bounce, pass through, and interfere with each other correctly, even in new environments the model never saw before).

The Bottom Line

This paper presents a way to teach computers the deep "rules of the game" (the Lagrangian) and then gives them a special, stable way to play the game (the ELM integrator) that prevents them from breaking the laws of physics over time. It allows AI to simulate complex physical systems for long periods without drifting into nonsense, and it lets those simulations work in new, different environments without needing to be retrained.

Technical Summary: Learned Lagrangian Models of PDEs via Euler–Lagrange Residual Minimization

Problem Statement

Forecasting the long-term dynamics of systems governed by partial differential equations (PDEs) presents a unique challenge for machine learning. While modern physics-informed and structure-preserving methods have improved the ability to learn model structures (such as Hamiltonians or Lagrangians) from data, they often treat the numerical integrator as a secondary implementation detail. This oversight is critical because forecasting demands long-term stability and bounded error, whereas solving for a local solution prioritizes immediate accuracy.

In systems that conserve energy (e.g., waves, vibrations), small per-step integration errors in standard methods compound over time, violating conservation laws and causing trajectories to diverge from the learned physical regime. Existing learned-physics methods often fuse the dynamics with their integrator or rely on fixed spatial discretizations. This coupling prevents the separation of model error (phase error) from integration error (conservation error) and limits generalization to arbitrary boundary conditions or spatially varying dynamics without retraining. Furthermore, prior work on learned variational integrators has been largely restricted to finite-dimensional ordinary differential equations (ODEs) or specific 1D PDEs with rigid grid structures.

Methodology: Euler–Lagrange Minimization (ELM)

The authors propose Euler–Lagrange Minimization (ELM), a variational integrator designed specifically for learned continuous Lagrangian models of PDEs. Unlike traditional approaches that discretize the action functional symbolically (which is impossible for learned black-box Lagrangians), ELM employs an optimization-based framework constructed at three scales:

Local Residual: The method defines a local, pointwise squared Euler–Lagrange residual, $R^2$ , which measures the deviation of a candidate path from the learned dynamics. For a Lagrangian density $L$ , the residual is:
$R = \frac{\partial L}{\partial q} - \frac{d}{dt}\frac{\partial L}{\partial q_t} - \frac{d}{dx}\frac{\partial L}{\partial q_x}$
The goal is to minimize $R^2$ rather than the action directly.
Patch Error: The domain is divided into local space-time patches. Within each patch, a Hermite interpolation $\hat{q}$ is constructed using a finite set of basis functions (e.g., monomials) fitted to known node states (values and derivatives). The patch error $J$ is the integral of the squared residual over the patch, approximated via quadrature:
$J \approx \sum_{i=1}^{n_\omega} \omega_i R^2[\hat{q} | (t_i, x_i)]$
Global Integrator: To advance the solution by one time step, the method seeks the node values at the new time step that minimize the patch error given the known states. This is solved via a damped Newton iteration swept across overlapping patches using Jacobi iteration. This approach avoids a coupled global solve, allowing the method to scale linearly with domain size.

Near-Symplecticity:
When the residual $R$ is driven to zero, the integrator becomes symplectic, inheriting the long-term energy conservation properties of classical variational integrators. In practice, the authors oversample the quadrature points (using more points than strictly required for symplecticity) to ensure the stability of the patch solver and robustness against imperfections in the learned Lagrangian. Consequently, ELM is described as near-symplectic, maintaining energy conservation orders of magnitude better than non-symplectic methods while avoiding the strict constraints of exact symplecticity.

Key Contributions

First Learned Continuous Lagrangian Integrator for PDEs: ELM is the first method to directly use a learned continuous Lagrangian for forecasting PDE dynamics, extending variational integration beyond finite-dimensional ODEs.
Decoupling of Errors: By relying on optimization rather than fixed time-stepping, ELM decouples model error (phase error) from integration error (conservation error). This allows the learned Lagrangian to represent the physics while the integrator ensures stability.
Mesh-Free and Boundary-Agnostic: The method is mesh-free and does not require pre-discretization of space. It handles arbitrary boundary conditions (Dirichlet, Neumann, driven, absorbing) and spatially varying dynamics purely through constraints on the optimization variables, without retraining the model.
Scalability: The use of Jacobi iteration for the global solve ensures the method scales linearly with domain size, bypassing the computational bottlenecks of global coupled solves inherent in implicit schemes for high-dimensional PDEs.

Experimental Results

The authors validate ELM on three systems of increasing complexity:

Chaotic Double Pendulum (ODE): ELM achieves energy drift comparable to state-of-the-art symplectic methods (specifically the 4th-order Gauss-Legendre Runge-Kutta) over 100,000 seconds, significantly outperforming non-symplectic integrators (RK4, Dormand-Prince) which exhibit steady energy accumulation.
1D Wave Equation: ELM outperforms Lagrangian Neural Networks (LNN) in both accuracy and energy conservation over 10,000 seconds. Unlike LNN, which suffers from fixed discretization errors that limit accuracy improvements with smaller time steps, ELM's accuracy scales with resolution. ELM maintains discernible wavefronts and symmetry where LNN devolves.
2D Wave Equation: ELM provides the first forward simulation of a 2D PDE using a learned continuous Lagrangian density. Over 1,000 seconds, ELM tracks the analytical ground truth with minor phase drift, whereas Fourier Neural Operators (FNO) diverge within 20 seconds and PDE-Refiner exhibits non-physical global oscillations. ELM achieves this with 40,000 times fewer parameters than PDE-Refiner.

Generalization to New Domains:
A critical finding is the model's ability to generalize without retraining. When a model trained on homogeneous media is applied to a domain with a sharp impedance interface (1D) or a double-slit barrier (2D), ELM correctly predicts reflection, transmission, and interference patterns. This is achieved by simply varying the Lagrangian spatially or constraining degrees of freedom, demonstrating that the learned model acts as a "physics primitive."

Significance and Claims

The paper argues that the integrator deserves the same level of design care as the model itself to achieve flexible physics-learning systems. The primary significance of ELM lies in its ability to separate the learning of physical laws (the Lagrangian density) from the simulation of those laws (the integrator).

The authors claim that ELM offers a starting point for learned models that can be deployed, constrained, and combined at inference time rather than requiring retraining for every new scenario or boundary condition. By avoiding the "baking in" of discretization, ELM allows for physically correct handling of complex boundary effects and spatial variations. While the method is currently more expensive per time step than standard integrators for ODEs, the authors posit that the trade-off in stability and generalization is essential for long-horizon forecasting of complex, conservative systems.

Limitations Acknowledged:
The paper modestly notes that ELM is currently more computationally expensive per step than standard integrators, particularly for ODEs where symplectic methods are already available. The Jacobi sweep converges slowly for strongly coupled systems in higher dimensions. Additionally, the current implementation accepts only Lagrangian densities with up to first derivatives, excluding higher-order field theories like beam mechanics. The authors suggest extending the framework to dissipative systems and adaptive mesh refinement as natural future directions.