TL;DR
This paper introduces a data-driven approach using neural networks to model and control X-ray adaptive mirrors, overcoming the limitations of traditional linear models. By learning the complex, non-linear dynamics of these mirrors, the system achieves precise shape control below 2 nm RMS, crucial for next-generation X-ray light sources.
What Problem Does This Paper Address?
Modern X-ray light sources, like synchrotrons and free-electron lasers, are producing beams with unprecedented brightness and coherence. To fully leverage this power for advanced experiments, the quality of the X-ray wavefront must be meticulously controlled. Adaptive optics, specifically deformable mirrors, are essential for this, but they face significant challenges. These mirrors, often made with piezo-bimorph actuators, exhibit complex, non-linear behaviors like drift, creep, and hysteresis. Traditional control methods rely on linear models, which struggle to accurately capture these dynamics, especially for rapid or large shape changes. Furthermore, probing the X-ray wavefront in situ can be difficult and interrupt experiments, necessitating control methods that require minimal or no feedback (open-loop). This paper tackles the problem of accurately modeling and controlling these non-linear adaptive mirrors in an open-loop fashion.
Methods
The authors propose a two-part framework:
- Data-Driven Predictive Modeling: They employ a feedforward neural network to learn a discrete-time model of the mirror's non-linear dynamics. This model takes into account not only the current mirror shape and applied voltage but also a history of past shapes and voltages. This is crucial because the mirror's future shape depends on its past states and how it has been actuated previously. The network architecture includes multiple fully connected layers with ELU activation and a skip connection, which was found to significantly improve performance, especially when the mirror is near rest.
- Optimization-Based Control: Once the predictive model is trained, it's used to generate a sequence of voltage inputs to drive the mirror from an initial state to a desired target shape. This is achieved by minimizing a non-convex objective function using the Adam algorithm, which penalizes deviations from the target shape over a finite time horizon. This approach is similar to iterative linear quadratic regulator (iLQR) but directly minimizes the objective function without linearization.
Datasets
The study utilizes an ex situ experimental setup with a prototype piezo-bimorph mirror and a visible-light Fizeau interferometer for surface metrology.
- Training Data: A large dataset of 5964 examples was collected. This involved applying sequences of random voltage inputs to the mirror actuators and simultaneously measuring the resulting surface profiles. The data acquisition was synchronized to a 2-second time interval.
- Test Data: A separate dataset of 497 examples was used to evaluate the model's performance. This data was acquired using a similar procedure to the training data.
The paper does not introduce a new benchmark dataset but rather uses its experimentally acquired data for training and testing.
Key Insights and Ideas
- Non-linearity is King: The paper strongly emphasizes that the non-linear dynamics (drift, creep, hysteresis) of piezo-bimorph mirrors are not minor nuisances but fundamental characteristics that must be accounted for to achieve nanometer-level precision.
- History Matters: The predictive model's success hinges on incorporating a history of past states and inputs. This acknowledges that the mirror's response isn't instantaneous and depends on its previous configurations.
- Data-Driven Beats Physics-Based (for this problem): By avoiding assumptions about the mirror's physical behavior and instead learning directly from data, the neural network model can generalize better and handle complex phenomena that are difficult to model analytically.
- Open-Loop Control is Achievable: The work demonstrates that with an accurate predictive model, precise mirror shape control can be achieved without continuous, invasive feedback, which is a significant advantage for X-ray beamlines.
Key Results
- Superior Prediction Accuracy: The neural network model achieved a mean prediction error of 1.26 nm RMS, significantly outperforming a traditional linear prediction model which had a mean error of 4.20 nm RMS over the test dataset.
- Precise Shape Control: The control algorithm successfully drove the mirror to achieve and maintain desired shapes with mean RMS errors below 2 nm. Specifically, 1.70 nm after the first step and 1.91 nm after ten steps for random shapes, and 1.44 nm after one step and 1.51 nm after ten steps for cylindrical shapes.
- Effective for Large Moves: The method demonstrated good performance even for relatively large surface height changes (up to ~430 nm RMS).
- Robustness to Time Intervals: While performance was best at the training time interval (2s), the model showed reasonable predictive capabilities even for longer intervals (12s, 30s) when applied iteratively.
Limitations and Future Work
- Time Interval Dependence: The predictive performance degrades at time intervals longer than the training interval. The authors suggest that acquiring more finely sampled data, especially for periods where voltage inputs are held constant, could improve performance at longer timescales.
- Optimization Non-Convexity: While the Adam optimizer is effective, the non-convex nature of the optimization problem means that guaranteed optimality is not achieved, potentially leading to slight overshoots.
- Future Work: The authors suggest that this data-driven calibration method is simple to implement and automatable. They believe it can be applied to other types of adaptive mirrors, such as resistive-element mirrors. They also note that the data could be collected in situ with a wavefront sensor for even more direct application.
Summary updated