Adaptive Learning of Periodic Solutions for Nonlinear Systems by Physics-informed Gaussian Process
This paper proposes a novel physics-informed Gaussian process (PIGP) framework that utilizes tailored periodic kernels and an adaptive collocation strategy to efficiently and accurately compute periodic solutions for nonlinear dynamical systems, demonstrating superior computational performance compared to the harmonic balance method.
Original paper licensed under CC BY 4.0 (https://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
In the world of engineering, many structures do not behave in simple, straight lines. When a bridge sways in the wind, a turbine blade spins at high speed, or a building shakes during an earthquake, the forces involved often change in complex ways depending on how much the object moves or how fast it is going. These are known as nonlinear systems. Unlike a simple spring that stretches the same amount no matter how hard you pull, these systems can twist, snap, or settle into strange, repeating patterns that are difficult to predict. Finding the specific, repeating motion—called a periodic solution—that a system will settle into is crucial for engineers. If they cannot predict these motions, they cannot design structures that are safe or efficient. For decades, scientists have relied on established mathematical tools to find these repeating patterns, but these tools often struggle when the system becomes too complex or too "jagged" in its behavior, requiring immense computing power to get a precise answer.
A team of researchers from Sun Yat-sen University and the University of Liege has developed a new approach to solve this problem, one that blends the laws of physics with a type of machine learning called a Gaussian process. Instead of forcing the system into a rigid mathematical box, their method, which they call a physics-informed Gaussian process, treats the physical laws governing the system as a set of clues. Imagine trying to guess the shape of a hidden object by feeling its surface at a few specific points; the more points you feel, the clearer the shape becomes. In this new method, the researchers use the physical equations of motion as the "feelings" at specific moments in time. They start with a rough guess of how the system moves and then use a smart, iterative process to refine that guess. The system learns from its own errors, adjusting its understanding of the motion until it perfectly matches the physical rules.
What makes this approach particularly powerful is its ability to learn where it needs to look. The researchers discovered that not all moments in time are equally important for understanding the motion. In systems where the movement is smooth, the information is spread out evenly. However, in systems where the motion is rough or sudden—like a part that hits a stop or a material that changes stiffness abruptly—the most critical information is concentrated in tiny, specific moments. The new method automatically detects these tricky moments and focuses its computational effort there, adding more "sensing points" exactly where the uncertainty is highest. This adaptive strategy allows the computer to find the solution using far fewer data points than traditional methods, making the process significantly faster and more efficient.
The researchers tested their new method on several different types of systems, ranging from simple single-mass oscillators to complex, multi-story building models. They compared their results against a widely used standard technique known as the harmonic balance method, which has been the go-to tool for decades. In every test, the new method proved to be highly accurate, capable of handling both smooth motions and those with sudden, jagged changes. Perhaps most impressively, when the researchers asked both methods to achieve the same level of precision, the new physics-informed approach consistently finished the job in less time. In some cases, it was more than twice as fast as the traditional method. This speed advantage comes from the method's ability to adapt; it does not waste time calculating details where they are not needed, but instead directs its resources to the parts of the problem that are hardest to solve.
Beyond just speed and accuracy, the new method offers a unique benefit that traditional tools lack: it provides a built-in measure of confidence. When the method calculates a solution, it also produces a map of uncertainty, showing exactly how sure it is about the result at every moment in time. This is like having a weather forecast that not only tells you the temperature but also gives you a clear range of how much that temperature might vary. For engineers designing critical infrastructure, knowing the limits of their calculations is just as important as the calculations themselves. The researchers found that as they added more sensing points to the system, this uncertainty map shrank, confirming that the solution was becoming more reliable.
While the method shows great promise, the researchers are careful to note its boundaries. They found that for systems with extremely sharp, sudden changes, the solution can sometimes wobble slightly near those sharp points, a challenge that even the traditional methods face. However, by demonstrating that their approach can handle a wide variety of difficult problems with greater efficiency and built-in reliability, the team has opened a new path for analyzing complex mechanical systems. Their work suggests that by letting machine learning learn directly from the laws of physics, rather than just from data, engineers can solve some of the most stubborn problems in dynamics with a clarity and speed that was previously out of reach.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.