A load-coupled DMDc method for long-term aerodynamic load prediction in controlled fluid–structure interaction systems

In the world of aerospace and energy engineering, there is a constant, invisible battle between the air rushing past a structure and the structure itself trying to hold its ground. When a flexible wing, a wind turbine blade, or a floating platform moves, it pushes against the air, and the air pushes back. This two-way conversation is called fluid–structure interaction. The problem is that the air does not just push back with a steady, predictable force; it swirls, separates, and changes its mind in complex, nonlinear ways that depend entirely on how the structure is moving at that exact moment. For engineers who want to control these systems—perhaps to stop a wing from shaking violently or to steer a robot through water—this unpredictability is a major hurdle. To design a controller that works, they need a model that can predict the future: if I move the wing this way, what will the air do next? Traditional computer simulations are incredibly accurate but so slow that they cannot run fast enough to be used for real-time control. Scientists have tried to build simpler, faster models, but these often fail when the system gets complicated or when the control inputs change, because they struggle to capture the messy, nonlinear dance between the fluid and the structure over long periods of time.

A team of researchers at Xi'an Jiaotong University has developed a new method to solve this specific problem, aiming to create a fast, accurate model that can predict aerodynamic loads and structural motion over long durations without losing its way. They focused on a controlled system where a flat-plate airfoil, similar to a small wing, is made to bend and twist by piezoelectric actuators—materials that change shape when an electric voltage is applied. The challenge was to teach a computer model to watch the air and the wing interact, learn the rules of their conversation, and then predict what would happen next under various control commands, even when those commands changed the system's behavior in complex ways. The researchers found that standard methods, which assume the system behaves in a simple, linear fashion, quickly lose accuracy when the wing moves significantly or when the control inputs shift. These older models would drift away from reality, predicting the wrong amount of force or the wrong timing for the wing's movement, making them useless for long-term control tasks.

To fix this, the team introduced a new approach they call load-coupled dynamic mode decomposition with control. Instead of treating the air's push as a simple, static reaction, their method actively incorporates the nonlinear effects of the aerodynamic loads into the model's internal state. Imagine the model as a student trying to learn a complex skill; previous methods might have tried to memorize the steps in a straight line, but this new method allows the student to look at the result of the previous step—the actual force the air exerted—and use that information to correct the next step. The researchers used a mathematical technique to break down the complex flow of air and the movement of the wing into their most important, simplified patterns. They then added a special mapping tool that takes the current state of the wing and the control voltage and predicts the aerodynamic load coefficients. Crucially, these predicted loads are fed back into the model to correct the evolution of the wing's motion, ensuring that the model stays consistent with the physical laws of how displacement and velocity are related. This creates a closed loop where the prediction of the force and the prediction of the motion constantly refine each other.

The researchers tested this new method against high-fidelity computer simulations, which serve as the gold standard for accuracy but are too slow for real-time use. They trained the model on data generated by a specific type of input called a frequency sweep, where the voltage applied to the wing was slowly varied to excite the system across a wide range of frequencies. Once trained, they challenged the model to predict the system's behavior over a continuous period of 140 seconds under four different types of control inputs: the same frequency sweep, a series of constant-frequency sine waves, a complex mix of multiple frequencies at once, and smooth step pulses that turned the voltage on and off. The results were striking. While other advanced methods began to drift or produce large errors after a few seconds, the new method maintained high accuracy throughout the entire 140-second test. The errors in predicting the aerodynamic forces and the wing's tip displacement remained remarkably low, never exceeding 7.33% for forces and 1.25% for displacement across all test cases. In many instances, the new method reduced prediction errors by more than half compared to the best-performing existing techniques.

The study also included a detailed investigation to understand exactly which parts of the new method were responsible for its success. By systematically removing different components, such as the nonlinear mapping tool or the constraints that keep the wing's motion physically consistent, the researchers showed that each part played a vital role. The nonlinear mapping was essential for capturing the subtle, complex variations in the aerodynamic forces, while the kinematic constraints ensured that the wing's movement remained realistic. Furthermore, they found that training the model using errors from long, multi-step predictions, rather than just single steps, was critical for preventing the model from drifting apart over time. Without this long-term training focus, the model would fail to sustain accurate predictions over the extended duration required for practical control applications. The findings demonstrate that by coupling the prediction of loads directly with the evolution of the system's state and enforcing physical consistency, it is possible to create a reduced-order model that is both fast enough for real-time use and accurate enough to handle the complex, nonlinear realities of fluid–structure interaction. This work provides a robust tool for engineers who need to predict and control flexible structures in the air and water, offering a path toward more efficient and stable designs for everything from aircraft wings to energy harvesters.