How a 0.1% Error Neural Operator Tamed a Chaotic Equation

0.1% — that’s the typical relative error achieved by a Fourier neural operator in approximating a feedback control law for one of the most unstable equations in fluid dynamics, the Kuramoto–Sivashinsky (KS) equation. More strikingly, this single neural network stabilizes every variant of the system it was trained on — including configurations where classical control methods fail completely due to unstable double eigenvalues. This isn’t just incremental progress in control theory. It’s a blueprint for how machine learning can master systems that defy traditional engineering: chaotic, nonlinear, and sensitive to infinitesimal changes. And because the KS equation models real-world phenomena like flame fronts, thin film flows, and plasma instabilities, this work opens a path toward stabilizing systems once considered too unruly to control.

The implications ripple outward. Imagine suppressing turbulence in microfluidic devices, preventing instabilities in fusion reactors, or smoothing out traffic flow on smart highways — all using a single, reusable AI controller trained once, then deployed across a wide class of physical systems. That future is now mathematically possible.

The Science

The Kuramoto–Sivashinsky equation describes how small perturbations grow and interact in systems with competing diffusion and anti-diffusion effects. In its simplest form, it reads:

$$\frac{\partial u}{\partial t}+\frac{{\partial }^{4}u}{\partial x^4}+\lambda \frac{{\partial }^{2}u}{\partial x^2}+u\frac{\partial u}{\partial x}=0$$

where $u(x,t)$ represents the state (like a flame front’s displacement), $\lambda$ controls the strength of anti-diffusion (which amplifies instabilities), and the fourth-order term provides stabilizing dispersion. When $\lambda$ is constant, the system’s behavior is relatively well understood. But in real applications — such as a flame propagating through a non-uniform medium or a liquid film with variable viscosity — $\lambda$ varies in space. That variation introduces complexity that breaks most existing control strategies.

The challenge is controllability: can we steer the system to zero using only boundary inputs? For constant $\lambda$, Coron and Lü (2015) showed that a single boundary input — curvature control $u_{xx}(0,t)$ — suffices except at a discrete set of $\lambda$ values where the linearized system develops repeated unstable eigenvalues. At these points, the system becomes uncontrollable through one input alone — a phenomenon known as loss of controllability due to eigenvalue collision.

Bhan, Krstic, and Shi tackle this head-on. They consider the KS equation with a spatially varying anti-diffusion coefficient $\lambda(x) \in W^{1,\infty}(0,1)$ and two boundary controls:

• $U_1(t) = u(0,t)$: boundary displacement
• $U_2(t) = u_{xx}(0,t)$: boundary curvature

Their key insight, inspired by Heymann’s Lemma from finite-dimensional control theory, is to assign these inputs distinct roles. The displacement input $U_1$ is used not for stabilization, but for pre-feedback — a preliminary control action that modifies the system’s spectral properties to ensure controllability. Once the system is made controllable via this pre-feedback, the curvature input $U_2$ takes over, applying a Fredholm backstepping transformation to achieve rapid stabilization.

Fredholm backstepping differs from the more common Volterra approach in that it constructs the control law using global spectral data — eigenvalues and eigenfunctions — rather than local kernel PDEs. This makes it powerful but also harder to approximate numerically, since the gain depends on the full spectrum of a non-self-adjoint operator.

The researchers prove that two inputs are not only sufficient but necessary when the plant has an unstable double eigenvalue. They then show that a simple exponential pre-feedback of the form

$$U_1(t)=\epsilon {\int }_0^1{e}^{-\theta y}u(y,t)dy$$

is enough to split repeated eigenvalues and restore controllability, regardless of $\lambda(x)$. This pre-feedback reshapes the operator’s spectrum so that the curvature input can then stabilize the system at any desired decay rate using the Fredholm transformation

$$U_2(t)={\int }_0^{1}h(y)u(y,t)dy$$

where the gain $h(y)$ is constructed from the eigenfunctions $\tilde{\chi}_n$ and coefficients $\tilde{\kappa}_n$ solving an infinite coupled system (Equation 1.6 in the paper).

What They Found

The paper delivers two major results, both supported by rigorous analysis and numerical validation.

First, the authors prove that their two-input feedback law stabilizes the nonlinear KS equation locally at any prescribed exponential rate $\omega > 0$, for any spatially varying $\lambda(x)$. This removes the restrictive condition $\lambda \notin \mathscr{N}_\lambda$ that plagued earlier single-input designs. The stabilization holds even when the original system has an unstable double eigenvalue — a case where single-input control fails completely.

Second, and perhaps more transformative, they prove that the mapping from system parameters $(\lambda, a, \varepsilon, \theta)$ to the control gain $h(y)$ is continuous on compact classes of admissible designs. This continuity is nontrivial: the Fredholm gain depends on eigenvalues that can collide, eigenfunctions that vary non-smoothly, and an infinite-dimensional coupled system. Yet, by controlling the spectral data and the tails of the modal series, they establish uniform continuity — which enables neural operator approximation.

They train a Fourier neural operator (FNO) on a compact class $\mathscr{U}_{B,d}$ of coefficients and design parameters. The FNO learns to map $\lambda(x)$ and scalars $(a,\varepsilon,\theta)$ directly to the gain function $h(y) \in L^2(0,1)$. Once trained, this single operator achieves:

• Typical relative $L^2$ gain errors of 0.1%
• Successful stabilization of all held-out test cases, including those with unstable double eigenvalues
• Real-time evaluation, bypassing the need to solve eigenvalue problems online

Neural Operator Achieves High Accuracy in Gain Approximation

Label	Value
FNO	0.1107
4-mode truncation	15.2
8-mode truncation	8.3
16-mode truncation	3.1

Source: Figure 5(b) in Bhan et al. (2026). FNO achieves sub-0.12% error, outperforming modal truncations.

The figure above shows the performance of the FNO. Panel (a) compares the exact gain (computed from 1024 modes) with the neural approximation and a crude 4-mode truncation. The FNO matches the reference gain almost perfectly, while the truncated version fails to capture key features. Panel (b) shows the pointwise relative error, which remains below 0.12% across the domain — an extraordinary accuracy for a learned operator.

The computational implications are profound. Computing the exact gain requires solving a high-dimensional eigenvalue problem and inverting an infinite system — a process that can take minutes or hours per instance. The FNO evaluates the gain in milliseconds, enabling real-time control.

Stabilization Parameters in Nonlinear Closed-Loop System

Label	Value
Decay rate ω	120
Max Re(μ̃_n)	-300
a	480
d	100

Source: Figure 2 caption and Theorem 1.1. Decay rate ω = 120 s⁻¹ achieved with a = 480.

Why This Changes Things

Control theory has long faced a trade-off: elegant analytical designs versus practical implementability. The Fredholm backstepping method is mathematically powerful, but its reliance on spectral data made it seem impractical for real-world use. This paper bridges that gap.

By proving continuity of the gain map and demonstrating high-accuracy neural approximation, Bhan et al. transform Fredholm control from a theoretical construct into a deployable technology. The FNO isn’t just a faster solver — it’s a generalization engine. Trained once on a diverse set of $\lambda(x)$ profiles, it can stabilize any new system in the class without retraining.

This is a paradigm shift. Traditional control design is instance-specific: you derive a controller for one system, then start over for the next. Here, the neural operator learns the entire design process — the mapping from physics to control law — and compresses it into a single, reusable function.

The implications extend far beyond the KS equation. Many physical systems — from Navier–Stokes flows to traffic dynamics — exhibit similar instabilities and are modeled by higher-order PDEs. The Heymann–Fredholm architecture — using one input to ensure controllability, another to stabilize — could become a template for controlling systems with spectral degeneracies.

Moreover, this work demonstrates that neural operators can handle nonlocal, spectral dependencies, not just local PDE kernels. Previous neural operator successes focused on Volterra-type gains, where the kernel satisfies a local PDE. Fredholm gains are fundamentally different: they depend on global eigenstructure. Proving that such maps can be learned — and that the learned gains preserve stability — is a major theoretical advance.

Consider the example of thin liquid films. These appear in coating processes, microfluidics, and heat exchangers. They are prone to rupture due to instabilities modeled by KS-type equations. Current control methods are limited to idealized, uniform conditions. With this new approach, one could design a single AI controller that stabilizes films with arbitrary spatial variations in viscosity or thickness — enabling more efficient, reliable industrial processes.

Similarly, in plasma physics, edge-localized modes (ELMs) in fusion reactors resemble KS dynamics. Uncontrolled, they can damage reactor walls. A fast, adaptive controller based on this method could suppress ELMs in real time, bringing fusion energy closer to reality.

What’s Next

The success of the FNO raises several exciting questions. Can this approach scale to higher-dimensional systems, such as 2D KS or full Navier–Stokes? The paper’s companion works on Poiseuille flow [26] and two-fluid interfaces [27] suggest yes — but the computational and theoretical challenges grow.

Another open direction is robustness. The current result guarantees stability for sufficiently accurate gains, but how much error can the closed loop tolerate? Can the neural operator be trained to be certifiably robust, with built-in error bounds?

There’s also the question of global stabilization. The current result is local: it works for small initial disturbances. Many real-world systems experience large perturbations. Extending this to global stability — perhaps using adaptive or hybrid control — is a natural next step.

Finally, can the pre-feedback itself be learned? The current design uses a fixed exponential form. A learned pre-feedback might achieve controllability more efficiently or handle a broader class of systems.

The most immediate impact may be in digital twins and smart infrastructure. Imagine a factory where every fluidic device has a digital twin. A single neural operator, trained offline on a library of possible configurations, could generate optimal controllers on demand — reducing design time from weeks to seconds.

We are no longer limited to controlling only the systems we can fully solve. We can now control the ones we can learn.

The figure above shows the full closed-loop stabilization using the neural operator. Panel (a) displays the spatially varying $\lambda(x)$ used in the test case — a realistic profile with sharp gradients. Panel (b) shows the Fredholm kernel, computed from 32 modes. Panel (c) plots the two control inputs, which quickly drive the system to zero. Panel (d) shows the evolution of the state $u(x,t)$, collapsing from an initial sine wave to near-zero within milliseconds. This is not simulation — it’s prediction-enabled control, where machine learning and rigorous mathematics converge to tame chaos.

The era of one-size-fits-all controllers is over. The era of one-network-fits-entire-classes has begun.