⚡ electrical engineering

Metriplectic Conditional Flow Matching for Dissipative Dynamics

Metriplectic Conditional Flow Matching (MCFM) is a novel framework that learns dissipative dynamics by integrating conservative-dissipative splits into both the neural vector field and a structure-preserving sampler, thereby ensuring physical consistency, stable long-horizon rollouts, and strict energy decay without relying on unstable long-rollout training.

Original authors: Ali Baheri, Lars Lindemann

Published 2026-06-12

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Ali Baheri, Lars Lindemann

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 teach a robot how to swing a pendulum. You show it thousands of short clips of the pendulum swinging and slowing down due to air resistance. Your goal is for the robot to learn the rules so it can predict the future motion of the pendulum accurately.

The problem is that standard "smart" robots (neural networks) are like over-enthusiastic students. They are great at memorizing the short clips, but when asked to predict the future for a long time, they start making up physics. They might accidentally add energy to the system, making the pendulum swing higher than it started, or they might forget how friction works entirely. In the real world, a pendulum never swings higher on its own; it always loses energy.

This paper introduces a new method called Metriplectic Conditional Flow Matching (MCFM). Think of it as giving the robot a strict set of "physical laws" built directly into its brain, rather than just hoping it learns them by accident.

Here is how it works, broken down with simple analogies:

1. The Two-Engine System (The Metriplectic Split)

The authors realized that real-world motion usually has two distinct parts:

The Conservative Engine: This is the part that keeps things moving without losing energy, like a perfect frictionless swing. It preserves the "shape" of the motion.
The Dissipative Engine: This is the part that acts like friction or air resistance. It only removes energy; it never adds it.

Standard AI models mix these two up, often letting the "friction" part accidentally add energy. MCFM forces the AI to keep these two engines separate. It builds a "Conservative Channel" that preserves energy and a "Dissipative Channel" that is strictly forbidden from adding energy—it can only take it away.

2. Learning from Short Clips (Conditional Flow Matching)

Instead of trying to simulate the whole future at once (which is hard and prone to errors), the AI learns from very short, consecutive snapshots (like looking at two frames of a video). It learns the "velocity" (how fast and in what direction the pendulum is moving) between those two frames. This keeps the learning simple and stable, avoiding the need to guess the entire future trajectory all at once.

3. The "Shrink" Step (The Sampler)

When the AI actually runs a simulation to predict the future, it uses a special step-by-step process called a Strang-prox scheme. Imagine walking across a room:

Step A: You take a perfect, energy-preserving stride (the Conservative part).
Step B: You take a "shrinking" step where you deliberately slow down, ensuring you lose a tiny bit of energy (the Dissipative part).
Step C: You take another perfect stride.

By alternating these steps, the AI guarantees that the total energy of the system goes down (or stays the same) at every single step. It physically cannot "invent" energy. If the AI tries to make the pendulum swing higher, the math of this "shrinking step" stops it.

4. The Results

The authors tested this on a damped pendulum (a swinging weight that slows down).

The Old Way (Unconstrained AI): The AI started correctly but eventually got confused. It began adding energy back into the system, causing the pendulum to swing wildly or drift away from the correct path.
The New Way (MCFM): The AI kept the pendulum swinging correctly for the entire duration. It respected the "no free energy" rule, keeping the motion smooth and realistic, just like the real physics.

The Bottom Line

This paper presents a way to build AI models that understand the difference between "swinging" and "slowing down." By baking these physical rules directly into the model's structure and its learning process, the AI produces predictions that are stable, realistic, and don't violate the basic laws of thermodynamics, all without needing complex, expensive training methods.

Technical Summary: Metriplectic Conditional Flow Matching for Dissipative Dynamics

Problem Statement

Learning dynamical models directly from data is essential for forecasting, control, and simulation. However, purely neural surrogates often violate fundamental physical laws, particularly in dissipative systems where the correct behavior involves a monotone loss of energy. Standard unconstrained neural networks tend to inject energy, drift off invariant manifolds, and destabilize over long horizons. Existing approaches face a dichotomy:

Soft penalties/Physics-informed residuals: These encourage energy decay but offer no hard guarantees and are sensitive to hyperparameters (weights and step sizes).
Structure-preserving conservative models (e.g., Hamiltonian/Lagrangian networks): These excel in closed systems but fail to address dissipation and can misrepresent frictional effects.

While recent flow-based and diffusion learners capture local velocities well from short transitions, their unconstrained parameterizations still permit non-physical energy creation during long rollouts. There is a need for a method that learns from data while strictly enforcing the conservative–dissipative split that governs real-world dynamics.

Methodology

The authors propose Metriplectic Conditional Flow Matching (MCFM), a framework that embeds the physical law of conservative-dissipative splitting directly into the model architecture and the sampling procedure.

1. Metriplectic Parameterization

The learned vector field $v_\theta(x, t)$ is constrained to a metriplectic decomposition:
$v_\theta(x, t) = J(x, t) \nabla H_\theta(x, t) - G_\theta(x, t) \nabla \Phi_\theta(x, t)$
Where:

$H_\theta$ is a learned invariant (energy-like quantity).
$\Phi_\theta$ is a learned dissipation potential (e.g., entropy or Lyapunov function).
$J$ is a fixed skew-symmetric operator ( $J^\top = -J$ ).
$G_\theta$ is a positive semidefinite metric ( $G_\theta \succeq 0$ ).

The model enforces degeneracy conditions ( $G_\theta \nabla H_\theta = 0$ and $J \nabla \Phi_\theta = 0$ ) either "hard" (by construction, restricting ranges) or "soft" (via penalties). These conditions ensure the conservative channel cannot change the dissipation potential, and the dissipative channel cannot change the energy. Consequently, the continuous-time dynamics guarantee $\frac{d}{dt}H_\theta = 0$ and $\frac{d}{dt}\Phi_\theta \leq 0$ .

2. Training: Conditional Flow Matching

Training avoids expensive long-rollout adjoint methods. Instead, it uses Conditional Flow Matching (CFM) on short, consecutive data pairs $(x_k, x_{k+1})$ collected at step $\Delta t$ .

The objective minimizes the regression loss between the vector field and the conditional target velocity $u_\tau = (x_{k+1} - x_k)/\Delta t$ .
The loss includes soft penalties for degeneracy violations and regularization terms.
If a trusted physical invariant $E_{phys}$ is known, $H_\theta$ is fixed to $E_{phys}$ , and only dissipation is learned; otherwise, both are learned from data.

3. Inference: Structure-Preserving Sampler

To ensure discrete-time energy decay, MCFM employs a Strang-prox splitting scheme:

Hamiltonian Half-Step: A symplectic update (e.g., velocity-Verlet) for the conservative part $\dot{x} = J \nabla H_\theta$ .
Metric Step: A proximal "shrink" step for the dissipative part $\dot{x} = -G_\theta \nabla \Phi_\theta$ . For common parameterizations, this has a closed-form solution that monotonically decreases the dissipation potential.
Hamiltonian Half-Step: A second symplectic update.

An optional projection step can be applied to strictly enforce $\nabla E_{phys} \cdot v \leq 0$ if a trusted energy function is available.

Key Contributions

Structure-First Parameterization: Introduction of MCFM, which encodes the conservative–dissipative split by design rather than relying on post-hoc regularizers.
Structure-Preserving Sampler: A pairing of the model with a Strang-prox integrator that guarantees discrete energy decay under mild conditions.
Theoretical Guarantees: Proofs linking the parameterization and sampler to conservation, monotonic dissipation, and stable rollouts in both continuous and discrete time.

Empirical Results

The method was evaluated on a damped pendulum (controlled dissipative system) with varying damping coefficients.

Baseline: An unconstrained neural velocity field (UF) trained with the same CFM objective and integrated via RK4.
Metrics: Energy increase fraction, max energy ratio, final energy ratio, sign-error fraction (positive energy rates), and terminal distribution fit (Sliced Wasserstein-2).

Findings:

Phase Portraits: The unconstrained baseline (UF) rapidly collapsed toward the origin and drifted from outer level sets. MCFM tracked the geometry of the true flow for the full horizon.
Energy Evolution: While both models dissipated initially, UF exhibited late-horizon energy increases (violating monotonicity). MCFM maintained a smooth, largely monotone decay with only small oscillations.
Physical Consistency: MCFM significantly reduced the fraction of energy-increase events and sign errors compared to UF.
Distributional Fit: MCFM matched the terminal distributional fit (SW-2) of the unconstrained baseline, demonstrating that physical constraints did not sacrifice data fidelity.

Significance and Claims

The paper claims that MCFM provides a "drop-in replacement" for unconstrained neural flows that produces stable, physically consistent rollouts without the need for hand-tuned penalties or expensive long-horizon training. By building the conservative–dissipative split into both the vector field and the sampler, the method ensures discrete energy decay and adherence to physical principles. The authors position this as a step toward learning dissipative dynamics that respect the fundamental laws governing real-world systems, bridging the gap between data-driven learning and geometric integration. Future work is suggested for richer parameterizations, extensions to PDEs and control, and tighter theoretical guarantees on long-horizon behavior.