Generative design of stabilizing controllers with diffusion models: the Youla approach

Matteo Cercola, Donatello Materassi, Simone FormentinarXiv 2025 · DOI · arXiv

Introduction

In control engineering, the design of controllers that simultaneously satisfy performance requirements—such as rapid settling time—and robustness constraints—such as sensitivity limits—remains a significant challenge. The mapping between a controller's parameters and its closed-loop performance is highly nonlinear and often non-convex. Traditional approaches typically rely on iterative numerical optimization, heuristic tuning, or exhaustive parameter sweeps, which are computationally intensive and generally produce only a single solution.

While generative AI, particularly diffusion models, has shown remarkable success in mapping complex distributions in other fields, its direct application to control synthesis is hindered by the risk of instability. A model that naively generates controller coefficients may frequently produce unstable closed-loop systems. This paper introduces a framework that leverages the generative power of diffusion models while providing rigorous mathematical guarantees of stability by design.

Methodology: Stability via Youla-Kucera Parameterization

The core innovation of this work is the integration of Denoising Diffusion Probabilistic Models (DDPM) with the Youla-Kucera parameterization. Instead of training the model to generate standard controller coefficients $C(s)$, the authors train it to generate a stable, proper Youla parameter $Q(s)$.

According to classical control theory, for any given plant, any stable $Q(s)$ is guaranteed to result in an internally stabilizing controller. This "Youla trick" effectively embeds the physics and mathematical constraints of stability directly into the representation space of the AI. The design process involves:

  1. Conditioning: The DDPM is conditioned on the plant's dynamics and desired performance metrics (settling time and the $\mathcal{H}_\infty$ norm of the sensitivity function).
  2. Generation: The model iteratively denoises a signal to produce a valid Youla parameter $Q(s)$.
  3. Guidance: Classifier-free guidance is employed during sampling to steer the generated controllers toward the user-specified performance targets without requiring expensive optimization loops at inference time.

Experimental Setup and Dataset

To train the model, the authors generated a synthetic dataset of 200,000 stable single-input single-output (SISO) linear time-invariant (LTI) plants paired with feasible Youla parameters. A separate test set of 1,000 unseen plants was used for evaluation. The dataset tracks plant and Youla coefficients alongside normalized closed-loop metrics: settling time and robustness (sensitivity function norm).

Key Results and Discussion

The framework demonstrates a high success rate, meeting user-specified targets in 95.3% of cases on the test set. A key advantage of this probabilistic approach is "multishot generation," where the model can produce multiple diverse, valid candidate controllers for a single specification almost instantly. This provides engineers with a Pareto-optimal set of solutions to choose from.

Comparison of target performance (J) versus generated controller performance across multiple samples. The model consistently generates controllers near the target specifications.

In instances where the model fails to meet the exact target (approximately 4.7% of the time), the deviations are minimal. The median error for the $\mathcal{H}_\infty$ norm was only 0.13, and the median settling time error was 2 seconds. This suggests a graceful degradation at the Pareto frontier; when requested specifications are physically unachievable, the model outputs controllers that reside on the feasible boundary.

Distribution of the absolute error between target specifications and the best generated controller performance. The narrow distributions highlight the precision of the generative process.

Conclusion and Future Work

By shifting the generative target to the Youla domain, this research successfully bridges modern generative AI with classical control theory. The resulting framework amortizes the cost of controller design, allowing for rapid, stable, and high-performance synthesis. Future work will focus on extending this approach to multi-input multi-output (MIMO) systems, higher-order controllers, and a broader range of performance metrics such as tracking accuracy and actuator limits.

Summary updated

BibTeX
@misc{cercola2025generative,
  author = {Matteo Cercola and Donatello Materassi and Simone Formentin},
  title = {{Generative design of stabilizing controllers with diffusion models: the Youla approach}},
  year = {2025},
  eprint = {2512.15725},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2512.15725},
}
Figure