Metriplectic Conditional Flow Matching (MCFM)

• The paper presents a metriplectic decomposition that splits dynamics into conservative (energy-preserving) and dissipative components to avoid spurious energy injection.
• MCFM employs a CFM-style local training on short transitions, using learned potentials and degeneracy conditions to enforce physical plausibility.
• A structure-preserving sampler based on Strang splitting ensures stable rollouts by balancing Hamiltonian and dissipative flows throughout integration.

Searching arXiv for the MCFM paper and closely related flow-matching / metriplectic papers to ground the article. arXiv શોધ પરિણામો (MCFM and related structural papers):

1. "Metriplectic Conditional Flow Matching for Dissipative Dynamics" — (Baheri et al., 23 Sep 2025) Published: 2025-09-23 Relevance: Direct source on MCFM.
2. "Improving Flow Matching by Aligning Flow Divergence" — (Huang et al., 31 Jan 2026) Published: 2026-01-31 Relevance: CFM error analysis and divergence alignment, explicitly noted as complementary to MCFM.
3. "Solving physics-constrained inverse problems with conditional flow matching" — (Dasgupta et al., 14 Mar 2026) Published: 2026-03-14 Relevance: Conditional FM in inverse problems; finite-data degeneracy analysis relevant to MCFM.
4. "UPMRI: Unsupervised Parallel MRI Reconstruction via Projected Conditional Flow Matching" — (Luo et al., 19 Dec 2025) Published: 2025-12-19 Relevance: Projected CFM with measurement-constrained geometry; comparison point for structured CFM.
5. "Multimarginal flow matching with optimal transport potentials" — (Kansal et al., 3 Jun 2026) Published: 2026-06-03 Relevance: Multimarginal FM via potential-guided targets; useful structural comparison.
6. "Metriplectic relaxation to equilibria" — (Bressan et al., 11 Jun 2025) Published: 2025-06-11 Relevance: General metriplectic theory underlying conservative–dissipative decompositions. Metriplectic Conditional Flow Matching (MCFM) is a structured variant of conditional flow matching for dissipative dynamical systems in which the learned continuous-time vector field is parameterized as a conservative–dissipative split and sampled with a matched structure-preserving integrator. In the formulation introduced in "Metriplectic Conditional Flow Matching for Dissipative Dynamics" (Baheri et al., 23 Sep 2025), MCFM retains the short-transition, simulation-free training style of conditional flow matching while imposing metriplectic structure directly in both the vector field and the inference scheme. Its stated purpose is to learn dissipative dynamics without violating first principles, in particular to avoid the energy injection, instability, and qualitatively incorrect long-horizon behavior that can arise in equally expressive unconstrained neural flows trained only by local velocity regression (Baheri et al., 23 Sep 2025).

1. Conceptual setting and metriplectic background

MCFM is formulated for finite-dimensional dissipative dynamical systems whose evolution can be decomposed into a conservative part and a dissipative part. The paper’s canonical example is the damped pendulum,

$$\dot q = \frac{p}{M}, \qquad \dot p = -MgL\sin q - \gamma p,$$

used to illustrate the distinction between transport along energy level sets and monotone dissipation toward lower-energy regions (Baheri et al., 23 Sep 2025).

The metriplectic background used by MCFM is consistent with the broader theory developed in "Metriplectic relaxation to equilibria" (Bressan et al., 11 Jun 2025). In that framework, metriplectic dynamical systems consist of a special combination of a Hamiltonian and a generalized entropy-gradient flow, such that the Hamiltonian is conserved and entropy is dissipated or produced depending on sign convention. The strong form is

$$\partial_t u = J(u)\frac{\delta \mathcal{H}(u)}{\delta u} - K(u)\frac{\delta \mathcal{S}(u)}{\delta u},$$

with compatibility conditions ensuring conservation of $\mathcal H$ and monotonicity of $\mathcal S$ (Bressan et al., 11 Jun 2025). In the sign convention of that paper, $\mathcal S$ is nonincreasing: $$\frac{d\mathcal{H}(u)}{dt}=0, \qquad \frac{d\mathcal{S}(u)}{dt}\le 0.$$

Within this lineage, MCFM can be read as a finite-dimensional, learning-oriented instantiation of the same conservative–dissipative principle. Its novelty is not the metriplectic idea by itself, but the coupling of that idea to conditional flow matching on short transitions and to a structure-preserving sampler tailored to the learned split (Baheri et al., 23 Sep 2025).

2. Metriplectic parameterization of the learned vector field

The central modeling assumption of MCFM is that the learned velocity field admits the decomposition

$$v_\theta(x,t) = J(x,t)\nabla H_\theta(x,t) - G_\theta(x,t)\nabla \Phi_\theta(x,t). \tag{1}$$

Here $H_\theta$ is a learned energy-like invariant, $\Phi_\theta$ is a learned dissipation potential, $J(x,t)$ is skew-symmetric, and $$\partial_t u = J(u)\frac{\delta \mathcal{H}(u)}{\delta u} - K(u)\frac{\delta \mathcal{S}(u)}{\delta u},$$0 is positive semidefinite: $$\partial_t u = J(u)\frac{\delta \mathcal{H}(u)}{\delta u} - K(u)\frac{\delta \mathcal{S}(u)}{\delta u},$$1 The paper further imposes the metriplectic degeneracy conditions

$$\partial_t u = J(u)\frac{\delta \mathcal{H}(u)}{\delta u} - K(u)\frac{\delta \mathcal{S}(u)}{\delta u},$$2

Under these conditions, the conservative channel $$\partial_t u = J(u)\frac{\delta \mathcal{H}(u)}{\delta u} - K(u)\frac{\delta \mathcal{S}(u)}{\delta u},$$3 is tangent to level sets of $$\partial_t u = J(u)\frac{\delta \mathcal{H}(u)}{\delta u} - K(u)\frac{\delta \mathcal{S}(u)}{\delta u},$$4, while the dissipative channel $$\partial_t u = J(u)\frac{\delta \mathcal{H}(u)}{\delta u} - K(u)\frac{\delta \mathcal{S}(u)}{\delta u},$$5 decreases $$\partial_t u = J(u)\frac{\delta \mathcal{H}(u)}{\delta u} - K(u)\frac{\delta \mathcal{S}(u)}{\delta u},$$6 monotonically (Baheri et al., 23 Sep 2025).

The continuous-time conservation and dissipation laws follow directly: $$\partial_t u = J(u)\frac{\delta \mathcal{H}(u)}{\delta u} - K(u)\frac{\delta \mathcal{S}(u)}{\delta u},$$7 and

$$\partial_t u = J(u)\frac{\delta \mathcal{H}(u)}{\delta u} - K(u)\frac{\delta \mathcal{S}(u)}{\delta u},$$8

The paper also writes this using the $$\partial_t u = J(u)\frac{\delta \mathcal{H}(u)}{\delta u} - K(u)\frac{\delta \mathcal{S}(u)}{\delta u},$$9-norm,

$\mathcal H$0

so that

$\mathcal H$1

A corollary states that if $\mathcal H$2 on its range for some $\mathcal H$3, then $\mathcal H$4 whenever $\mathcal H$5 (Baheri et al., 23 Sep 2025).

When a trusted physical energy is available, MCFM allows fixing

$\mathcal H$6

For mechanical systems with $\mathcal H$7, mass matrix $\mathcal H$8, and potential $\mathcal H$9, the trusted energy is

$\mathcal S$0

with gradient

$\mathcal S$1

If no trusted invariant is available, both $\mathcal S$2 and $\mathcal S$3 are learned (Baheri et al., 23 Sep 2025).

A broader metriplectic implication, made explicit in the relaxation theory, is that preservation of an invariant is not by itself sufficient; the nullspace of the dissipative operator matters. "Metriplectic relaxation to equilibria" emphasizes the desirability of specific or minimal degeneracy,

$\mathcal S$4

because overly degenerate dissipators can preserve invariants yet fail to relax to the intended constrained entropy extrema (Bressan et al., 11 Jun 2025). This suggests that MCFM’s hard-degeneracy constructions are not merely aesthetic constraints but are tied to the structure of admissible equilibria.

3. Conditional flow matching objective and learnable components

MCFM trains on short consecutive samples $\mathcal S$5 obtained at timestep $\mathcal S$6. For each pair, the method samples

$\mathcal S$7

forms the linear interpolation

$\mathcal S$8

and uses the target velocity

$\mathcal S$9

This is a conditional flow matching construction on short transitions rather than a rollout-based training objective (Baheri et al., 23 Sep 2025).

The training loss is

$\mathcal S$0

The first term is the usual short-transition velocity matching term. The second term penalizes violation of the degeneracy conditions when these are not imposed by construction. The final term regularizes the potentials and metric operator (Baheri et al., 23 Sep 2025).

The paper represents $\mathcal S$1 and $\mathcal S$2 with small MLPs taking $\mathcal S$3 as input, where $\mathcal S$4 is a Fourier time embedding. The metric operator is parameterized as

$\mathcal S$5

which guarantees positive semidefiniteness by construction. The skew operator $\mathcal S$6 is typically fixed rather than learned, often using canonical symplectic blocks in mechanical coordinates, so antisymmetry is likewise enforced by design (Baheri et al., 23 Sep 2025).

Two enforcement regimes are described. In hard degeneracy, the range of $\mathcal S$7 is restricted so that $\mathcal S$8 holds by construction, and $\mathcal S$9 is aligned so that $$\frac{d\mathcal{H}(u)}{dt}=0, \qquad \frac{d\mathcal{S}(u)}{dt}\le 0.$$0. In soft degeneracy, the terms

$$\frac{d\mathcal{H}(u)}{dt}=0, \qquad \frac{d\mathcal{S}(u)}{dt}\le 0.$$1

are added as penalties. The experimental system emphasizes Hard–MCFM (Baheri et al., 23 Sep 2025).

This parameterization means that MCFM does not learn an unconstrained vector field directly. It learns a conservative potential, a dissipation potential, and a PSD dissipative geometry, from which the field is assembled. The paper explicitly contrasts this with an equally expressive unconstrained neural flow

$$\frac{d\mathcal{H}(u)}{dt}=0, \qquad \frac{d\mathcal{S}(u)}{dt}\le 0.$$2

which has no built-in mechanism to prevent nonphysical energy injection (Baheri et al., 23 Sep 2025).

4. Structure-preserving inference and guarantees

A defining feature of MCFM is that the structured parameterization is paired with a matched structure-preserving sampler. The sampler uses a Strang splitting between the Hamiltonian subflow

$$\frac{d\mathcal{H}(u)}{dt}=0, \qquad \frac{d\mathcal{S}(u)}{dt}\le 0.$$3

and the metric subflow

$$\frac{d\mathcal{H}(u)}{dt}=0, \qquad \frac{d\mathcal{S}(u)}{dt}\le 0.$$4

One Strang step of size $$\frac{d\mathcal{H}(u)}{dt}=0, \qquad \frac{d\mathcal{S}(u)}{dt}\le 0.$$5 is

$$\frac{d\mathcal{H}(u)}{dt}=0, \qquad \frac{d\mathcal{S}(u)}{dt}\le 0.$$6

The conservative half-steps are handled by a symplectic, time-reversible integrator, while the dissipative step is a proximal or implicit-Euler metric step (Baheri et al., 23 Sep 2025).

The general implicit metric update is

$$\frac{d\mathcal{H}(u)}{dt}=0, \qquad \frac{d\mathcal{S}(u)}{dt}\le 0.$$7

Under smoothness assumptions, this satisfies

$$\frac{d\mathcal{H}(u)}{dt}=0, \qquad \frac{d\mathcal{S}(u)}{dt}\le 0.$$8

In the damped pendulum experiment, the metric step is implemented as

$$\frac{d\mathcal{H}(u)}{dt}=0, \qquad \frac{d\mathcal{S}(u)}{dt}\le 0.$$9

which is described as a dissipative shrink along the momentum coordinate (Baheri et al., 23 Sep 2025).

The discrete-time guarantees mirror the continuous ones. For a symmetric symplectic integrator of order $$v_\theta(x,t) = J(x,t)\nabla H_\theta(x,t) - G_\theta(x,t)\nabla \Phi_\theta(x,t). \tag{1}$$0, backward error analysis gives near-conservation of the learned invariant: $$v_\theta(x,t) = J(x,t)\nabla H_\theta(x,t) - G_\theta(x,t)\nabla \Phi_\theta(x,t). \tag{1}$$1 For the full Strang–prox step, the dissipation potential decreases according to

$$v_\theta(x,t) = J(x,t)\nabla H_\theta(x,t) - G_\theta(x,t)\nabla \Phi_\theta(x,t). \tag{1}$$2

The paper states that if $$v_\theta(x,t) = J(x,t)\nabla H_\theta(x,t) - G_\theta(x,t)\nabla \Phi_\theta(x,t). \tag{1}$$3 on its range and $$v_\theta(x,t) = J(x,t)\nabla H_\theta(x,t) - G_\theta(x,t)\nabla \Phi_\theta(x,t). \tag{1}$$4 is small enough, then the right-hand side is strictly negative whenever $$v_\theta(x,t) = J(x,t)\nabla H_\theta(x,t) - G_\theta(x,t)\nabla \Phi_\theta(x,t). \tag{1}$$5 (Baheri et al., 23 Sep 2025).

When a trusted physical energy is available, MCFM adds an optional projection

$$v_\theta(x,t) = J(x,t)\nabla H_\theta(x,t) - G_\theta(x,t)\nabla \Phi_\theta(x,t). \tag{1}$$6

This enforces

$$v_\theta(x,t) = J(x,t)\nabla H_\theta(x,t) - G_\theta(x,t)\nabla \Phi_\theta(x,t). \tag{1}$$7

while leaving tangential motion unchanged. The appendix then proves via the descent lemma that for an explicit step $$v_\theta(x,t) = J(x,t)\nabla H_\theta(x,t) - G_\theta(x,t)\nabla \Phi_\theta(x,t). \tag{1}$$8, one obtains $$v_\theta(x,t) = J(x,t)\nabla H_\theta(x,t) - G_\theta(x,t)\nabla \Phi_\theta(x,t). \tag{1}$$9 for sufficiently small $H_\theta$0, with strict decrease whenever the original field had a positive energy-increasing component (Baheri et al., 23 Sep 2025).

A common misconception is that MCFM is merely ordinary CFM with an added penalty. The paper explicitly argues otherwise: its contribution is a package consisting of a metriplectic parameterization, CFM-style local training on short transitions, a Strang–prox structure-preserving inference scheme, and theoretical guarantees linking all three to conservation, dissipation, and rollout stability (Baheri et al., 23 Sep 2025).

5. Empirical evaluation and observed behavior

The reported benchmark is a damped pendulum with damping parameter

$H_\theta$1

step size $H_\theta$2, $H_\theta$3 trajectories for training, and $H_\theta$4 trajectories for testing. Training examples are one-step transition pairs $H_\theta$5, and the main baseline is an equally expressive unconstrained neural flow trained with the same conditional flow matching objective but integrated at test time with RK4 (Baheri et al., 23 Sep 2025).

Over a $H_\theta$6 s rollout, the paper reports the following metrics:

• energy increase fraction,
• max energy ratio,
• final energy ratio,
• sign-error fraction,
• terminal distributional fit via sliced Wasserstein-2 (SW-2).

The qualitative findings are summarized as follows. The unconstrained flow initially tracks dissipation but later exhibits energy rebounds, positive $H_\theta$7, and phase-space collapse toward the origin in a nonphysical way. MCFM maintains phase portraits closer to ground truth, with smoother and more monotone energy decay, markedly fewer energy-increase events, and fewer positive energy-rate violations, while keeping terminal distributional quality comparable to the unconstrained baseline (Baheri et al., 23 Sep 2025).

These findings support the paper’s central claim that local transition fitting is not sufficient for physically realistic long-horizon rollout in dissipative systems. The conservative–dissipative split supplies an inductive bias in which Hamiltonian substeps move tangentially along level sets, dissipation acts only in designated dissipative directions, and the proximal step suppresses spurious energy injection (Baheri et al., 23 Sep 2025).

The same section also makes the principal tradeoff explicit. MCFM deliberately restricts the hypothesis class to metriplectic fields. This may exclude dissipative dynamics not well represented by such a decomposition, but the paper presents the restriction as the reason for improved long-horizon physical plausibility when the assumption is appropriate (Baheri et al., 23 Sep 2025).

6. Relation to adjacent flow-matching developments

Several contemporaneous developments clarify what MCFM is not, and why its structural commitments matter.

First, "Improving Flow Matching by Aligning Flow Divergence" shows that standard conditional flow matching can match local velocities while still producing inaccurate probability paths because path error depends crucially on divergence mismatch through the continuity equation (Huang et al., 31 Jan 2026). The paper derives the error PDE

$H_\theta$8

and proves the bound

$H_\theta$9

That work is not metriplectic, but it identifies a structural weakness of plain CFM: matching the vector field alone does not guarantee correct density evolution. This suggests that, within an MCFM setting, divergence-sensitive regularization is complementary rather than redundant (Huang et al., 31 Jan 2026).

Second, "Solving physics-constrained inverse problems with conditional flow matching" studies CFM as an amortized posterior sampler for inverse problems and gives a finite-data theory of overtraining-induced degeneracy, including variance collapse and selective memorization (Dasgupta et al., 14 Mar 2026). Under a linear interpolant, one overfit regime yields

$\Phi_\theta$0

so that $\Phi_\theta$1 and the learned conditional law collapses to a Dirac. Another regime produces trajectories terminating at memorized training points associated with nearby observations (Dasgupta et al., 14 Mar 2026). A plausible implication is that any MCFM variant must control not only physical structure but also terminal-time collapse mechanisms induced by finite-data overfitting.

Third, "UPMRI: Unsupervised Parallel MRI Reconstruction via Projected Conditional Flow Matching" provides a different kind of structure: projection onto the measurement-visible subspace,

$\Phi_\theta$2

a continuity-equation derivation of measurement-space dynamics, and a cyclic dual-space sampling algorithm alternating learned prior transport with explicit data-consistency projection (Luo et al., 19 Dec 2025). The paper itself is explicit that this is not metriplectic: there is no Poisson bracket, no dissipation potential, and no Hamiltonian/Casimir structure. Its relevance to MCFM lies in the fact that it is a comparison point for structured conditional flow matching under operator-induced geometry (Luo et al., 19 Dec 2025).

Fourth, "Multimarginal flow matching with optimal transport potentials" derives modified conditional FM targets from a dynamic OT action augmented by soft intermediate-marginal potentials (Kansal et al., 3 Jun 2026). Its exact sample dynamics are

$\Phi_\theta$3

and the conditional target decomposes into a standard CFM transport term plus potential-driven corrections: $\Phi_\theta$4 This is not a conservative–dissipative metriplectic split, but it provides a template for deriving structured conditional targets from variational principles (Kansal et al., 3 Jun 2026).

Taken together, these works delimit MCFM’s position. It is not merely “physics-constrained CFM,” nor merely “projected CFM,” nor merely “potential-guided FM.” Its distinctive content is the imposition of a metriplectic decomposition

$\Phi_\theta$5

together with a sampler designed to preserve the same structure in discrete time (Baheri et al., 23 Sep 2025).

7. Scope, limitations, and research significance

The demonstrated scope of MCFM is controlled and low-dimensional: the reported experiment is the damped pendulum, and the authors explicitly identify future directions including richer skew and metric parameterizations, extension to PDEs and control, and tighter long-horizon theory (Baheri et al., 23 Sep 2025). The method also relies on the assumption that the target system admits a meaningful conservative–dissipative decomposition, and the paper notes that hard degeneracy may become nontrivial in more complex domains (Baheri et al., 23 Sep 2025).

From the standpoint of general metriplectic theory, an important limitation is that conservation and monotone dissipation are guaranteed in continuous time only under exact skew/PSD structure and exact degeneracy, while discrete strict monotonicity is obtained only up to $\Phi_\theta$6 unless the trusted-energy projection is used (Baheri et al., 23 Sep 2025). In the wider metriplectic literature, relaxation to the intended constrained equilibria also depends on the nullspace structure of the dissipative operator, not solely on invariant preservation (Bressan et al., 11 Jun 2025). This suggests that future MCFM extensions may need to control not just conservation and decay, but also the stationary set of the dissipative channel.

Within conditional flow matching research more broadly, MCFM occupies a specific niche. It preserves the attractive local-regression training paradigm of CFM on short transitions, avoiding long rollout adjoints, yet seeks to address a failure mode that standard CFM does not target: physically incorrect long-horizon behavior despite acceptable local fit (Baheri et al., 23 Sep 2025). In that sense, MCFM is best understood as a structure-first formulation of conditional flow matching for dissipative systems, grounded in metriplectic dynamics and paired with a numerically compatible inference procedure (Baheri et al., 23 Sep 2025).