⚡ electrical engineering

Adaptive Stability-Constrained Neural Differential Equations for Controlled Dynamical Systems with Unknown Inputs

This paper proposes an Adaptive Stability-Constrained Neural Differential Equation (AS-NDE) framework that jointly learns a nominal vector field and a state-input-dependent Riemannian metric to ensure exponential trajectory convergence for controlled dynamical systems with unknown perturbations, while explicitly accounting for input-dependent metric derivatives and providing a reproducible evaluation protocol with illustrative synthetic results.

Original authors: Syed Pouladi

Published 2026-08-11

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Syed Pouladi

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 are trying to teach a robot to drive a car by watching it navigate a few specific streets. You show it how to turn left at a red light and speed up on a straight road. But what happens when you send that robot out into the real world, where the wind blows harder than expected, or the driver suddenly slams on the brakes? In the world of artificial intelligence, this is the difference between "learning a pattern" and "learning a law." Scientists use tools called Neural Ordinary Differential Equations (Neural ODEs) to teach computers the rules of how things move over time, like a pendulum swinging or a motor spinning. These models are great at mimicking what they've seen, but they often fall apart when the situation changes slightly. If the robot learns to drive on a sunny day, it might crash on a rainy one because it didn't learn the underlying physics of how rain affects traction; it just memorized the sunny day. The big question for engineers is: how do we build a learning system that stays stable and safe even when the inputs (like wind or driver commands) are different from what it practiced on?

This paper introduces a new method called Adaptive Stability-Constrained Neural Differential Equations (AS-NDE) to solve exactly that problem. Think of the AI model as a student trying to learn the rules of a game. Usually, the student just tries to guess the next move correctly. If they make a tiny mistake, that mistake can grow into a huge disaster as the game goes on. The author proposes a new way of training where the student doesn't just learn the moves; they also learn a special "safety map" (called a Riemannian metric) that changes depending on the current situation. This map acts like a rubber band that pulls the student's predictions back toward the correct path whenever they start to drift. Crucially, this safety map isn't fixed; it adapts in real-time based on the controls being used, like a driver adjusting their grip on the steering wheel when the road gets slippery.

The paper shows that by using this adaptive map, the model can guarantee that two different paths (like two cars driving the same route but with slightly different wind or driver inputs) will stay close to each other. The author derived a mathematical rule proving that the distance between these paths shrinks exponentially over time, unless the difference in their inputs is huge. However, there is a very important catch: the numbers and graphs in this paper are not real results from a finished experiment. The author explicitly states that this is a "draft" with "illustrative synthetic placeholders." They have built the theory and the recipe for how to test it, but they haven't actually run the full tests on real data yet. The curves you might see in the figures are just stand-ins to show what the results could look like.

To understand how this works, imagine you are walking a dog on a leash. In a standard AI model, the leash is a rigid stick. If the dog (the prediction) wanders off a little, the stick doesn't bend enough to correct it, and the dog runs far away. In the AS-NDE model, the leash is a smart, stretchy bungee cord that gets tighter or looser depending on how fast the dog is running and where the wind is blowing. The "adaptive metric" is the brain of that bungee cord. It constantly calculates how much tension is needed to keep the dog close. The paper proves mathematically that if you use this smart bungee cord, the dog will stay close to you even if you both get hit by a sudden gust of wind, provided the wind isn't stronger than the cord can handle.

The author tested their idea on two imaginary scenarios to show how the system would work. The first was a "Duffing oscillator," which is like a spring that gets stiffer the more you pull it, a classic physics problem. The second was a "Permanent-Magnet Synchronous Motor" (PMSM), which is the kind of electric motor used in high-performance electric cars and robots. In their simulated examples, the new AS-NDE method showed much smaller errors and fewer "stability violations" (moments where the safety map failed) compared to older methods. For instance, in their placeholder data, the new method had a "violation rate" of only 2.0%, while a standard method had 61.0%. But remember, these are just numbers from a simulation designed to prove the concept works on paper.

The paper is very careful to say what it is not claiming. It does not claim to have solved the problem of controlling every physical system in the world. It does not claim that the method works perfectly on real hardware yet. In fact, the author lists several "threats to validity," warning that if the simulation is too perfect, the results might look better than they would in the messy real world. They emphasize that the "safety certificate" they created is a promise about the model's behavior, not a guarantee about the actual physical machine unless the model is proven to match the machine perfectly.

In the end, this paper is a blueprint. It lays out a rigorous, reproducible way to train AI models that are mathematically guaranteed to be stable and robust against unknown disturbances. It combines the flexibility of deep learning with the safety of control theory. The author has provided the "recipe" (the math, the training steps, and the test protocols) and a set of "mock ingredients" (the synthetic data) to show how the dish should taste. The next step, which they clearly identify as necessary, is to actually cook the meal: to run these tests on real-world data, replace the placeholder numbers with real measurements, and see if the AS-NDE model can truly keep a robot or a motor safe when the unexpected happens. Until then, the paper stands as a technically consistent and promising starting point, ready for the real-world experiments that will prove its worth.

Technical Summary: Adaptive Stability-Constrained Neural Differential Equations

Problem Statement
The paper addresses a critical limitation in learning continuous-time vector fields for controlled dynamical systems: the sensitivity of model rollouts to inputs and disturbances that differ from the training distribution. While Neural Ordinary Differential Equations (NODEs) and their variants offer flexible approximation for nonlinear system identification, a small one-step error in the learned vector field can grow rapidly when the model is rolled out under unseen control inputs or unknown perturbations. Existing stability-aware learning methods often focus on autonomous systems or fail to explicitly account for the time derivative of input-dependent metrics, leading to heuristic regularizers that may not guarantee robustness against input mismatches.

Methodology
The author proposes Adaptive Stability-Constrained Neural Differential Equations (AS-NDE), a framework designed for systems with measured controls and unmatched, unknown perturbations. The core of the methodology involves:

Joint Learning of Dynamics and Metric: The model learns a nominal vector field $f_\theta(x, u)$ and a state–input-dependent Riemannian metric $M_\phi(x, u)$ simultaneously.
Structural Positive Definiteness: The metric is parameterized via a lower-triangular matrix $L_\phi$ such that $M_\phi = L_\phi L_\phi^\top + \epsilon_M I$ , ensuring positive definiteness by construction.
Contraction Analysis with Input Derivatives: Unlike heuristic approaches, the training objective explicitly includes the total time derivative of the metric, $\dot{M}$ , which accounts for both state evolution and the time derivative of the input ( $\dot{u}$ ). The contraction residual is defined as $S_\Theta = \dot{M} + A_\theta^\top M + M A_\theta + 2\lambda M$ , where $A_\theta = \partial f_\theta / \partial x$ .
Training Objective: The loss function combines prediction error with a penalty for violating a prescribed contraction rate $\lambda$ (via a softplus function on the maximum eigenvalue of $S_\Theta$ ) and penalties to enforce metric bounds.
Incremental Input-to-State Stability (ISS) Bound: The author derives a theoretical bound on the distance between two trajectories with different inputs and disturbances. The bound shows that the Riemannian distance decays exponentially, with the steady-state error bounded by gains determined by the differences in controls and disturbances. Crucially, this derivation explicitly retains the term involving $\partial M / \partial u \cdot \dot{u}$ .

Key Contributions
The paper makes three primary contributions:

Input-Conditioned Metric: Introduction of a metric $M_\phi(x, u)$ whose positive definiteness is structural and whose total derivative is retained during training, addressing a gap in previous heuristic stability regularizers.
Theoretical Robustness Bound: Establishment of an incremental input-to-state contraction bound that quantifies trajectory separation under differing controls and unknown disturbances, providing a certificate for the learned model's sensitivity.
Reproducible Evaluation Protocol: Specification of a rigorous evaluation protocol for two benchmarks: a forced Duffing oscillator and a Permanent-Magnet Synchronous Motor (PMSM). This includes specific ablation criteria, baseline comparisons, and a clear distinction between method and evidence.

Results and Experimental Status
The paper does not present empirical results from executed experiments. The author explicitly states that no measured data or trained checkpoints were available during the drafting of this manuscript. Consequently:

All numerical curves, tables (e.g., Table I), and performance metrics (e.g., NRMSE, violation rates) are identified as illustrative synthetic placeholders.
The provided coordinates are embedded in the source code for direct replacement once experiments are conducted.
The "results" section serves as a template for a reproducible study, detailing how to split data, select baselines (e.g., plain NODE, SINDy), and report uncertainty, rather than providing evidence of the method's superiority.

Significance and Claims
The paper positions itself as a technically consistent starting point for a defensible experimental study rather than a demonstration of empirical superiority. Its significance lies in:

Theoretical Rigor: By explicitly deriving the metric's time derivative and distinguishing between sampled training penalties and formal plant certificates, the work avoids common pitfalls in stability-aware learning.
Defensibility: The author emphasizes the necessity of separating the method from unexecuted evidence. They argue that without running the prescribed experiments and replacing the placeholders, any claim of empirical benefit would be unfounded.
Practical Guidance: The paper provides a detailed roadmap for future researchers, including specific network architectures, hyperparameter ranges, and threat-to-validity checks (e.g., segment leakage, solver tolerance biases) to ensure that future evaluations of AS-NDE are rigorous and comparable.

In summary, the paper introduces a novel, theoretically grounded framework for learning stable controlled dynamics but deliberately withholds empirical claims until the proposed reproducible protocol is executed and the placeholder data is replaced with actual experimental results.

Technical Summary: Adaptive Stability-Constrained Neural Differential Equations

More like this