TL;DR
This paper introduces a data-driven framework using the Expectation-Maximization (E-M) algorithm to calibrate and adapt wavefront correction systems in real-time. By treating unmeasurable electric fields as "hidden variables," the approach allows high-contrast telescopes to automatically correct for model inaccuracies and environmental drifts, significantly speeding up the search for faint exoplanets.
What Problem Does This Paper Address?
To image an Earth-like exoplanet, astronomers must suppress the blinding glare of its host star—a feat comparable to spotting a firefly next to a lighthouse from miles away. Instruments called coronagraphs do this by creating "dark holes" in the starlight’s point spread function.
However, these dark holes are incredibly fragile. Small optical imperfections (aberrations) and thermal drifts create "speckles" of light that drown out the planet. Current correction systems rely on model-based controllers, but the underlying physics models (Fourier optics) are often biased. We don't perfectly know the gain of every actuator on a deformable mirror (DM), nor can we perfectly measure the "non-common-path errors" that occur deep inside the instrument. Because the model is slightly wrong, the correction process is slow and often hits a performance floor.
Methods
The authors reformulate Focal Plane Wavefront Correction (FPWC) as a stochastic optimization problem. They use an Expectation-Maximization (E-M) algorithm to bridge the gap between the theoretical model and the actual hardware.
- The Hidden Variable Problem: You can't measure the complex electric field at the focal plane directly; you can only measure intensity (the square of the field). The electric field is thus a "hidden variable."
- The E-Step (Expectation): The algorithm uses a Kalman filter and a Rauch smoother to estimate the most likely state of the hidden electric fields based on noisy camera images.
- The M-Step (Maximization): Using the estimates from the E-step, the algorithm updates the system's Jacobian matrix (the map of how DM movements affect the light) and noise covariance matrices. This can be done analytically or via stochastic gradient ascent.
- Adaptive Control: Unlike traditional "offline" calibration, this framework can run in a reinforcement learning-like loop, adapting the model while the telescope is actively performing wavefront correction.
Datasets
The researchers validated their approach using two primary sources:
- Numerical Simulations: A simulated high-contrast system with 10nm RMS wavefront aberrations and 20% biased DM actuator gains. They generated a training set of 4,000 random DM commands and 16,000 resulting images.
- Experimental Data: Real-world testing was conducted at Princeton’s High Contrast Imaging Lab (HCIL). The testbed uses a Shaped Pupil coronagraph and two Boston Micromachines MEMS DMs (952 actuators each).
Key Insights and Ideas
- Physical Meaning of Regularization: The authors show that the "tuning parameters" used in standard controllers (like Tikhonov regularization) are mathematically equivalent to the magnitude of Jacobian uncertainties. Tuning the controller is actually an exercise in system identification.
- Gradient vs. Analytical: In simulations, analytical solutions are precise. However, in real-world experiments, stochastic gradient ascent proved more robust. Analytical updates tended to overfit to camera noise, while gradient methods reached the same ultimate contrast as the "perfect" physics model.
- Efficiency of Active Learning: The adaptive control (reinforcement learning) approach was more efficient than random exploration. Because the controller "chooses" to move actuators that aren't blocked by the coronagraph mask, the data it collects is more informative for improving the model where it matters most.
Key Results
- Model Accuracy: The E-M algorithm successfully reduced Jacobian errors in simulation from ~9% to ~2%.
- Speed and Contrast: In lab experiments, the identified model reached a contrast of better than $3 \times 10^{-7}$ in just 4 to 5 iterations, significantly faster than the biased "physics-only" model.
- Self-Correction: The adaptive framework successfully corrected for initially underestimated noise levels (process and observation noise), allowing the Kalman filter to better balance model predictions against new measurements.
Limitations and Future Work
- The Linearity Assumption: The current model assumes a linear relationship between DM commands and electric field changes. This holds for small corrections (dark hole maintenance) but fails when large DM strokes are needed to correct major aberrations.
- Scalar Regularization: To simplify the math, the authors assumed actuator errors are independent (scalar regularization). In reality, actuators are coupled, and identifying a full "regularization matrix" remains a high-dimensional challenge.
- Next Steps: The authors plan to integrate Extended Kalman Filters (EKF) and Neural Networks to handle system nonlinearities and explore more complex inter-actuator couplings. This would move the system closer to "self-maintaining" telescopes for future space missions like HabEx or LUVOIR.
Summary updated