🤖 AI

Solving Inverse Problems of Chaotic Systems with Bidirectional Conditional Flow Matching

This paper introduces Bidirectional Conditional Flow Matching (Bi-CFM) and its conservation-constrained variant (CBi-CFM) to effectively solve inverse problems in chaotic systems by learning bidirectional mappings that capture stochastic evolution, respect physical laws, and achieve significant accuracy and speed improvements across diverse dynamical systems ranging from low-dimensional models to real-world million-body globular clusters.

Original authors: Peiyan Hu, Jian Zhang, Jiashu Pan, Ruiqi Feng, Tao Zhang, Zhi-Ming Ma, Yuan-Sen Ting, Gongjie Li, Tailin Wu

Published 2026-06-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Peiyan Hu, Jian Zhang, Jiashu Pan, Ruiqi Feng, Tao Zhang, Zhi-Ming Ma, Yuan-Sen Ting, Gongjie Li, Tailin Wu

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

The Big Problem: The "Shattered Vase" of Chaos

Imagine you have a beautiful, complex vase. You throw it against a wall, and it shatters into a million pieces. Now, imagine someone hands you the pile of shards and asks, "Can you tell me exactly what the vase looked like before it broke, and exactly how hard you threw it?"

In the world of physics, this is called an inverse problem. Usually, if you know the rules (gravity, force), you can predict the future (the vase breaking). But working backward from the broken pieces to the original vase is incredibly hard.

This gets even harder with chaotic systems. Think of a chaotic system like a game of "butterfly effect." If you change the starting position of a planet or a weather pattern by a tiny, invisible amount (like the flap of a butterfly's wing), the final result changes completely. Because of this, if you try to work backward using standard math, tiny mistakes in your guess get multiplied exponentially, turning a small error into a huge, wrong answer. It's like trying to un-mix a cup of coffee and milk; once they are mixed, you can't easily separate them back out.

The Solution: A Two-Way Street Map

The authors of this paper propose a new AI method called Bidirectional Conditional Flow Matching (Bi-CFM).

Instead of trying to reverse-engineer the physics step-by-step (which causes those errors to pile up), they teach an AI to learn a direct map between the "before" and the "after."

Here is the analogy:

Old Way (Backward Integration): Imagine trying to walk backward through a dark, crowded maze. You take one step, guess where you were, take another step, and guess again. By the time you reach the start, you've likely taken a wrong turn because your memory is fuzzy.
New Way (Bi-CFM): Imagine the AI has a magical, high-resolution map that shows the entire path from the start to the finish at once. It learns the relationship between the starting point and the ending point directly. It doesn't care about the messy steps in between; it just knows, "If the system ends up here, it likely started there."

Why "Bidirectional"?
The AI doesn't just learn to go from Start $\to$ Finish. It learns both ways simultaneously. It learns how the system evolves forward and how to reverse it. By checking that the forward path and the backward path match each other, the AI corrects its own mistakes, ensuring the map is accurate.

The Special Rule: The "Conservation Law"

Some physical systems have strict rules that never change, like Conservation of Energy. Imagine a bank account where you can move money around, but the total balance must always stay the same.

The authors created an upgraded version called CBi-CFM. This version forces the AI to respect these "bank rules."

If the final state has a certain amount of energy, the AI is mathematically forced to find a starting state that has that exact same amount of energy.
It does this by keeping the AI's "guessing path" glued to a specific surface (called a manifold) where the rules are always obeyed. It's like training a dog to stay on a leash; the dog can run around, but it can never leave the path defined by the leash (the conservation law).

What They Tested It On

The team tested this method on three types of challenges:

Classic Math Puzzles (Lorenz & Circuit Systems):
They used famous mathematical models that are known to be chaotic. The AI was much faster (over 100 times faster) and more accurate than traditional methods. It successfully figured out the starting conditions that led to specific chaotic patterns, even when the time gap was long.
Planetary Scattering (The "Cosmic Pinball"):
They simulated a solar system with three planets. Sometimes, planets crash into each other or get flung out of the system entirely (information loss).
- The Challenge: If a planet disappears, you can't just reverse the video to find it.
- The Result: The AI, especially the version with conservation rules (CBi-CFM), could accurately guess the original arrangement of the planets, even when one was missing in the final picture. It respected the laws of physics (energy) better than any other method.
Real-World Star Clusters (The "Ancient City"):
They looked at Globular Clusters—huge, dense balls of stars that have been evolving for 10 billion years. We can see what they look like now, but we want to know what they looked like when they were born.
- The Challenge: 10 billion years of stars bumping into each other erases the memory of the beginning.
- The Result: The AI reconstructed the initial state of these star clusters much better than the best existing computer simulations (Monte Carlo methods). It produced a "birth certificate" for these star clusters that matched the real observations of the sky.

The Bottom Line

This paper presents a new way to solve the "un-solvable" problem of figuring out the past of chaotic systems.

Speed: It is incredibly fast (over 100x faster than old methods).
Accuracy: It doesn't let small errors grow out of control.
Physics: It respects the fundamental laws of the universe (like energy conservation).

Think of it as giving scientists a time machine that doesn't just play the movie backward, but uses a smart map to instantly reconstruct the beginning of a chaotic event, whether it's a planet being flung out of a solar system or a star cluster forming billions of years ago.

Technical Summary: Solving Inverse Problems of Chaotic Systems with Bidirectional Conditional Flow Matching

Problem Statement
The paper addresses the inverse problem of chaotic systems: inferring initial conditions from final observed states. This task is fundamentally challenging due to three factors:

Ill-posedness and Non-uniqueness: Solutions may not be unique or stable; small perturbations in the final state can correspond to vastly different initial states.
Exponential Error Accumulation: In chaotic dynamics, infinitesimal errors in initial conditions amplify exponentially over time (the "butterfly effect"). Consequently, traditional numerical methods that integrate backward in time fail to recover accurate initial states, as errors accumulate rapidly, rendering the reverse trajectory divergent from the true history.
Information Loss: In systems involving collisions or ejections (e.g., planetary scattering), components of the system may vanish during evolution, making the reconstruction of the full initial state impossible using only the remaining observable components.

While deep learning has advanced the modeling of chaotic dynamics for prediction tasks, existing methods have not adequately addressed the specific challenges of inverse inference in chaotic regimes, particularly regarding long-timescale stability and physical consistency.

Methodology
The authors propose Bidirectional Conditional Flow Matching (Bi-CFM), a probabilistic generative framework designed to learn bidirectional mappings between the distributions of initial and final states.

Conditional Flow Matching (CFM) Foundation: The method builds upon CFM, which learns a velocity field $v_\theta$ that defines a flow path from a simple source distribution (e.g., Gaussian noise) to a target data distribution. Unlike traditional solvers that rely on iterative time-stepping, CFM learns an end-to-end mapping, avoiding the accumulation of numerical errors over the time horizon.
Bidirectional Modeling: To address the instability of reverse-time integration, Bi-CFM jointly learns both the forward dynamics $p(u_T | u_0)$ and the reverse dynamics $p(u_0 | u_T)$ . By enforcing mutual consistency between these two directions, the model captures the global structure of the chaotic system rather than just reproducing initial states, thereby mitigating error amplification.
Conservation-Constrained Bi-CFM (CBi-CFM): For systems governed by conservation laws (e.g., energy, momentum), the authors extend the framework to CBi-CFM. This variant enforces physical constraints by:
1. Constrained Prior: Sampling the prior distribution via random walks on the conservation manifold, followed by projection corrections to mitigate numerical drift.
2. Tangent Space Projection: Constraining the learned velocity field to lie within the tangent space of the conservation manifold. This ensures that the entire probability flow evolves strictly within the manifold defined by the conserved quantity $H$ .

Key Contributions

Algorithmic Innovation: The introduction of Bi-CFM, which utilizes bidirectional learning to solve the ill-posed inverse problem of chaotic systems without relying on time-dependent iterative integration.
Physical Consistency: The development of CBi-CFM, which integrates conservation laws directly into the generative flow, ensuring that inferred initial states satisfy physical invariants (e.g., total energy) even in the presence of information loss.
Distribution-Level Modeling: A shift from deterministic trajectory recovery to probabilistic distribution matching, acknowledging the stochastic nature of chaotic evolution under limited computational resources and the non-uniqueness of solutions.

Experimental Results
The authors evaluated the method across three distinct domains:

Classic Chaotic Systems (Lorenz, Circuit, Lorenz 96):
- Bi-CFM outperformed baselines (including Backward Integration, a standard neural backbone, and a Random baseline) across five distribution-level metrics, including Wasserstein-2 distance and KL divergence.
- While Backward Integration performed reasonably well for short time horizons, its performance degraded rapidly as the horizon increased due to error accumulation. Bi-CFM maintained stability across long horizons.
- Visualizations confirmed that Bi-CFM closely matched the ground-truth joint distributions of initial and final states, whereas other methods exhibited significant deviations or collapsed to lower-dimensional manifolds.
Three-Body Planetary Systems (Scattering and Collision):
- In a system involving planet-planet scattering and collisions (where information is lost), Backward Integration was inapplicable.
- CBi-CFM demonstrated superior adherence to energy conservation, with Relative Conservation Errors comparable to the ground-truth numerical integration errors.
- The method successfully inferred the distributions of initial features (mass, semi-major axis, eccentricity) and captured causal relationships, distinguishing between trajectories with and without scattering events.
Real-World Million-Body Globular Clusters:
- Applied to the inverse problem of inferring initial states of globular clusters from observational data (surface brightness and velocity dispersion profiles) after $\sim 10$ Gyr of evolution.
- Compared to a state-of-the-art Monte Carlo method, Bi-CFM produced final simulated profiles that better matched observational data both visually and quantitatively (lower $\chi^2$ errors).
- The method demonstrated scalability, handling high-dimensional problems ( $N \sim 10^5$ particles) with inference times comparable to lower-dimensional systems, whereas traditional numerical methods suffer from the curse of dimensionality.

Significance and Claims
The paper claims that Bi-CFM and CBi-CFM provide a scalable route to solving inverse problems in long-timescale, high-dimensional chaotic systems. The authors emphasize that their approach:

Mitigates Error Accumulation: By learning end-to-end mappings rather than integrating backward, the method avoids the exponential error growth inherent in chaotic reverse-time dynamics.
Handles Ill-Posedness: The probabilistic formulation naturally accommodates non-unique solutions by modeling the distribution of possible initial states.
Ensures Physical Validity: The conservation-constrained extension allows the model to generate physically consistent solutions in systems where conservation laws are critical.
Offers Computational Efficiency: The method achieves a speedup of more than two orders of magnitude over traditional backward integration while maintaining accuracy, making it feasible for large-scale astrophysical applications like Galactic archaeology.

The authors conclude that their work establishes a new tool for investigating the assembly history of complex physical systems, such as planetary systems and globular clusters, where traditional numerical inversion fails.