📊 statistics

Physics-Constrained Fine-Tuning of Flow-Matching Models for Generation and Inverse Problems

This paper introduces a framework that fine-tunes flow-matching generative models via differentiable PDE residual minimization and joint latent parameter optimization to generate physically consistent solutions and solve inverse problems in scientific systems.

Original authors: Jan Tauberschmidt, Sophie Fellenz, Sebastian J. Vollmer, Andrew B. Duncan

Published 2026-07-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jan Tauberschmidt, Sophie Fellenz, Sebastian J. Vollmer, Andrew B. Duncan

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 have a very talented artist who has spent years learning to paint landscapes by looking at thousands of photos. This artist (the Flow-Matching Model) is great at creating realistic-looking pictures, but they don't actually understand the laws of physics. If you ask them to paint a river, they might make it look beautiful, but the water might be flowing uphill or defying gravity because they only learned from the look of the photos, not the rules of how water behaves.

This paper presents a new way to teach this artist the rules of physics after they have already learned to paint, without having to send them back to art school for years.

Here is how the method works, broken down into simple concepts:

1. The Problem: The "Guessing Game"

In science, we often want to figure out hidden things (like the type of rock underground or the wind speed) based on what we can see (like a seismic image or a weather map). This is called an inverse problem.

The old way: You usually need a massive library of "paired" data (e.g., a picture of the rock and the exact map of what's underground) to train a computer. But in the real world, we rarely have those hidden maps. We only have the pictures.
The limitation: If you just train an AI on the pictures, it learns to mimic the texture but might ignore the physics (like conservation of energy or fluid flow).

2. The Solution: "Physics-Constraint Fine-Tuning"

The authors propose a "post-training" trick. Think of it like a tutor who steps in after the artist has finished their initial training.

The Tutor's Rulebook: The tutor has a rulebook of physics equations (Partial Differential Equations, or PDEs). These equations describe how the world works (e.g., "Water must flow downhill," or "Heat must spread out").
The Weak-Form Check: Instead of demanding the artist's painting be mathematically perfect down to the last decimal (which is hard and causes errors), the tutor uses a "soft check." Imagine the tutor sprinkling a few random "test points" over the painting. They check if the physics holds up at those specific spots. If the water is flowing uphill at a test point, the tutor gives a gentle nudge to correct it. This is called using weak-form residuals.

3. The Magic Trick: Learning the Hidden Parameters

Here is the clever part. The artist doesn't just need to fix the painting; they also need to guess the hidden settings that caused the painting to look that way.

The Scenario: Imagine the artist is painting a river. The river looks a certain way because of the slope of the land (a hidden parameter). The artist doesn't know the slope, but they need to guess it to make the water flow correctly.
The Joint Dance: The authors built a system where the artist and a "parameter predictor" dance together.
1. The artist paints the river.
2. The predictor guesses the slope of the land.
3. The tutor checks: "If the slope is this, does the water flow correctly?"
4. If not, they both adjust slightly. The artist tweaks the water, and the predictor tweaks the slope guess.
5. They repeat this until the water flows perfectly according to physics, and the slope guess makes sense.

4. The "Adjoint Matching" Engine

How do they adjust the artist without breaking their ability to paint diverse, creative scenes?

The Metaphor: Imagine the artist is walking a path. The "Adjoint Matching" method is like a GPS that gently steers the artist off the "lazy path" (where they might ignore physics) and onto a "physics-compliant path."
The Benefit: It doesn't force the artist to paint the exact same river every time. It ensures that whatever river they paint, it obeys the laws of physics. This keeps the creativity (diversity) alive while fixing the errors.

5. What They Tested

The authors tested this on four different "physics puzzles":

Darcy Flow: Figuring out how water moves through porous rock (like a sponge).
Elasticity: How materials stretch and bend under pressure.
Helmholtz: How sound waves travel through different materials.
Stokes Flow: How thick fluids (like honey) move in a container.

They also tested it on natural images (like photos of parrots). In this case, the "physics" was replaced by a "preference" (making the parrot look like Pop Art). This proved the method works not just for science, but for any situation where you want to tweak a generated image to fit a specific rule.

The Bottom Line

This paper introduces a way to take a powerful AI that generates images or data and give it a "physics conscience."

It doesn't need expensive, perfect training data with hidden labels.
It can fix "broken" physics in the generated output.
It can simultaneously guess the hidden causes (like material properties) that created the image.
It keeps the output diverse and creative, rather than making everything look the same.

In short, it turns a "pretty picture generator" into a "scientific simulator" that respects the laws of nature, all without needing to retrain the whole system from scratch.

Technical Summary: Physics-Constrained Fine-Tuning of Flow-Matching Models

Problem Statement

Deep generative models, such as diffusion and flow-matching (FM) networks, have demonstrated success in representing complex spatio-temporal physical systems. However, standard generative models often fail to strictly adhere to the underlying physical laws (e.g., conservation laws, boundary conditions) governing these systems, particularly when trained on observational data where the governing parameters (e.g., material properties, source terms) are unknown or unobserved.

Existing approaches face significant limitations:

Physics-Informed Neural Networks (PINNs): While effective for forward/inverse problems, they typically regress single solutions and do not capture the distribution of plausible outcomes required for generative tasks.
Pre-training with Physics Losses: Integrating physics constraints during pre-training (e.g., Physics-Based Flow Matching) often requires paired data of solutions and their corresponding physical parameters, which is frequently unavailable, expensive to obtain, or high-dimensional.
Inference-Time Constraints: Methods that enforce constraints at inference time (e.g., projection-based sampling) can introduce discontinuities or struggle with local constraints like boundary conditions without retraining.

The core challenge addressed by this work is how to fine-tune a pre-trained flow-matching model—trained on low-fidelity or observational data without ground-truth parameters—to enforce parameter-dependent Partial Differential Equation (PDE) constraints and jointly infer the latent physical parameters, all without requiring joint parameter-solution training data.

Methodology

The authors propose a framework for Physics-Constrained Fine-Tuning (PCFT) that combines Adjoint Matching with a Joint State-Parameter Evolution strategy.

1. Reward Definition via Weak-Form Residuals

Instead of using strong-form PDE residuals (which involve high-order derivatives and are unstable with noisy data), the method employs weak-form residuals. For a generated sample $x$ and inferred parameters $\alpha$ , the residual is defined as:
$R_{weak}(x, \alpha) = \frac{1}{N_{test}} \sum_{i=1}^{N_{test}} |\langle L_\alpha x, \psi^{(i)} \rangle_{L^2(\Omega)}|^2$
where $L_\alpha x = 0$ represents the governing PDE, and $\psi^{(i)}$ are randomly sampled, compactly supported test functions (Wendland polynomials). This formulation provides a stable, low-variance learning signal that penalizes deviations from physical laws and boundary conditions.

2. Joint Evolution of State and Parameters

Since the base model generates only the state $x$ without corresponding parameters $\alpha$ , the authors introduce a latent parameter predictor $\phi$ and evolve both the state $x$ and the parameter $\alpha$ jointly:

Inverse Predictor: A network $\phi(x_1) = \alpha_1$ is pre-trained to minimize the PDE residual on fully denoised samples from the base model.
Surrogate Base Flow: Since the true flow of $\alpha$ is unknown for the base model, a surrogate flow is constructed using the inverse predictor. At any state $(x_t, \alpha_t)$ , the predicted final parameter $\hat{\alpha}_1 = \phi(\hat{x}_1)$ (where $\hat{x}_1$ is the predicted final state) defines a target direction for the parameter evolution.
Regularization: An additional vector field $v_{reg}$ is introduced to pull the fine-tuned parameter trajectory toward the base model's estimated parameters, preserving sample-specific details and preventing over-regularization.

3. Adjoint Matching Fine-Tuning

The fine-tuning process is formulated as a stochastic optimal control problem. The goal is to steer the base generative process $v_{base}$ toward a reward-tilted distribution $p_r(x) \propto e^{\lambda r(x)}p(x)$ , where the reward $r$ is the negative PDE residual.

Control Problem: The fine-tuned vector field $v_{ft}$ is derived by adding a control term $u_t$ to the base drift.
Lean Adjoint State: To optimize the control efficiently, the method uses a "lean adjoint state" that evolves backward in time. This avoids computing gradients through the entire generative trajectory during the backward pass, reducing computational cost.
Loss Function: The objective is a consistency loss between the learned control $u_t$ and the adjoint state $a_t$ :
$\mathcal{L} = \frac{1}{2} \int_0^1 \| u_t(\tilde{X}_t) + \sigma(t) a_t \|^2 dt$
Noise Schedule: The authors introduce a scaled memoryless noise schedule $\sigma^2(t) = (1-\kappa)2\eta_t$ to stabilize training and mitigate numerical blow-ups near $t=0$ , a novel extension to the standard Adjoint Matching framework.

Key Contributions

Post-Training Physical Enforcement: A strategy to tilt the generative distribution toward PDE-consistent samples using weak-form residuals, improving physical validity while preserving the diversity of the original model.
Joint Inverse Problem Solving: A principled formulation that evolves state and latent parameters jointly via vector fields, enabling the inference of hidden physical parameters (e.g., permeability, elasticity) without paired training data.
Theoretical Grounding: The integration of Adjoint Matching with a scaled noise schedule, providing a stable, low-variance optimization framework for reward-based fine-tuning of flow-matching models.
Cross-Domain Utility: Demonstration that the framework extends beyond PDEs to natural image generation, where "physical constraints" are replaced by parametric color transformations guided by preference scores.

Experimental Results

The method was evaluated on four canonical PDE families and a natural image task:

Darcy Flow: The model successfully denoised pressure fields and recovered permeability maps. Fine-tuning reduced PDE residuals significantly. The authors demonstrated a controllable trade-off between residual reduction and distributional fidelity (measured by MMD and SSIM diversity) via hyperparameter tuning ( $\lambda_x, \lambda_f$ ).
Linear Elasticity: Under boundary condition misspecification, the method achieved low boundary errors and weak/strong residuals while maintaining modest distributional shift, outperforming Physics-Based Flow Matching (PBFM) and ECI (Extrapolation-Correction-Interpolation) baselines.
Helmholtz Equation: In a setting with model mismatch (damped vs. lossless), the joint Adjoint Matching model achieved the lowest residuals and the best distributional fidelity (lowest MMD) among all tested methods, including PBFM and ablated Adjoint Matching variants.
Stokes Lid-Driven Cavity: The method handled a severe model mismatch (non-zero forcing vs. zero forcing). While residual levels were comparable across Adjoint Matching variants, the joint model uniquely achieved low parameter-distribution discrepancies (MMD $_\alpha$ ), highlighting its ability to learn the correct parameter distribution.
Natural Images: Applied to a Latent Flow Matching model for class-conditional generation (macaws), the method successfully performed parametric color transformations to match text prompts ("Pop Art"), demonstrating that the framework can bridge generative modeling with preference-based inference.

Significance and Claims

The paper claims to bridge the gap between generative modeling and scientific inference. By enabling the fine-tuning of pre-trained models to enforce complex, parameter-dependent physical constraints without requiring paired solution-parameter data, the approach offers a data-efficient pathway for:

Simulation-Augmented Discovery: Generating physically plausible solutions and exploring hypothetical scenarios.
Inverse Problems: Inferring latent physical parameters from sparse or noisy observations.
Robustness: Maintaining sample diversity while adhering to physical laws, avoiding the mode collapse often seen in constrained optimization.

The authors position this work as a step toward "physics-aware generative modeling," where models can be adapted to specific physical regimes or constraints post-hoc, making them more versatile tools for scientific discovery and engineering applications. The code is made available to ensure reproducibility.