Solving Inverse Problems of Chaotic Systems with Bi... | AI Research

Modeling chaotic systems is a fundamental challenge in science and engineering because these systems are highly sensitive to initial conditions; even the smallest perturbation can grow exponentially, making it nearly impossible to trace a final state back to its origin. This paper introduces a new approach called Bidirectional Conditional Flow Matching (Bi-CFM) to solve these "inverse problems," where the goal is to infer the initial state of a system based on its final observed state.

Addressing Chaos with Probabilistic Modeling

Traditional numerical methods often fail to solve inverse problems in chaotic systems because they rely on backward integration, which becomes unstable and inaccurate as time progresses. Instead of trying to find a single, deterministic path, the researchers treat chaotic evolution as a stochastic process. By learning the probability distribution of states rather than a single trajectory, the model can better handle the inherent uncertainty and non-uniqueness of chaotic dynamics. The framework is "bidirectional," meaning it simultaneously learns both forward and reverse mappings, which helps the model maintain consistency and prevents the exponential error accumulation that plagues other methods.

Enforcing Physical Laws

Many physical systems, such as planetary orbits, are governed by strict conservation laws (like the conservation of energy). To address this, the authors developed an extension called Conservation-constrained Bi-CFM (CBi-CFM). This version ensures that the inferred initial states and the resulting probability flow strictly adhere to the system's known conservation manifold. By projecting the model's velocity field onto the tangent space of this manifold, the system ensures that the generated solutions are physically plausible and maintain energy conservation at a level comparable to ground-truth data.

Performance and Scalability

The researchers tested their method on a variety of systems, ranging from classic chaotic models like the Lorenz and Circuit systems to high-dimensional problems and complex astrophysical scenarios. Across these tests, Bi-CFM consistently outperformed traditional baselines, achieving better accuracy in distribution-level metrics while running more than two orders of magnitude faster. In the case of three-body planetary scattering—a problem where information is often lost due to collisions or ejections—the model successfully reconstructed the initial conditions.

Real-World Applications

Beyond theoretical models, the researchers applied their method to real-world observational data from globular clusters. These dense stellar systems have evolved over billions of years, making them a difficult testbed for understanding assembly history. The results demonstrated that Bi-CFM is a scalable and accurate tool for long-timescale dynamics, providing a new way to look back at the history of complex, chaotic systems in the universe.