-
Notifications
You must be signed in to change notification settings - Fork 8
Future Directions for Advanced Adaptive Optics Control
This page describes several current and future directions for AO development. The current cacao software implements "conventional" linear control, as described below. More advanced features will be implemented.
Key challenges requiring new tools/approaches include:
- Deriving calibration on-the-fly
- Predictive control
- Sensor Fusion
- Real-time PSF estimation
- Focal plane wavefront control
Conventional AO control is linear and relies on pre-calibration.
In conventional adaptive optics (AO) control, most of the processing steps are linear. For example, input wavefront sensor (WFS) measurements are multiplied by a control matrix to yield wavefront mode coefficients, which are in turn multiplied by another matrix to yield deformable mirror (DM) commands.
Conventional AO systems require a response matrix (RM), also referred to as a interaction matrix, to be acquired prior to running the control loop. The RM is acquired or computed as part of the system calibration, and is then used to compute the control matrix (CM) linking input WFS measurements to output commands.
While much of cacao is built around this conventional AO control framework, more advanced control schemes should yield significant performance gains. Consequently, cacao is build to support further development in these areas. This page describes such modes and how they may be implemented. Much of this is at the early R&D stage and has not made it into the source code yet.
To help support advanced wavefront control approaches, cacao currently provides :
- real-time read/write access to all raw data and telemetry streams
- real-time pseudo-open loop reconstruction
- accurate timing information
- a modular software design where additional compute processes can be deployed
Pre-calibration is often inaccurate: the WFS response (RM) may depend on continuously varying atmospheric conditions, and may be very difficult to match the on-sky response with an internal light source. The most reliable way to measure the RM is to do so on-sky, while the AO loop is already running. Ideally, this can be done continuously as a non-invasive background task.
cacao supports on-the-fly calibration by use of multiple DM channels, as if the user could control multiple independent DMs. Every time one of the DM channels is modified, they are all added prior to sending the command to the physical DM. This allows AO loop control to write on one DM channel while another channel is used to measure the RM. With good timing calibration, the DM modulation required to acquire a RM can also be sent very rapidly to minimize the contribution of residual atmospheric turbulence to the RM, and so that a new RM can be quickly measured.
Predictive control addresses the time lag and WFS noise error terms that often dominate residual wavefront errors in conventional AO systems.
cacao includes a linear predictive control module based on the empirical orthogonal functions (EOFs) framework. Recent open-loop telemetry is processed to identify dominant spatial and temporal modes. Once these modes are identified, real-time telemetry is projected onto these modes to predict the future wavefront state.
The technique is purely empirical and does not require input parameters, so it can operate in the background without user supervision.
Future developments should implement more capable predictive control including non-linearities. The current software also lacks monitoring/analysis tools (for example, deriving wind speed from prediction filters), resulting in a poor understanding of performance limits and how to improve them. A more efficient implementation should be capable, for example, of adapting to varying wind speed by adopting a predictive filter tuned to the wind conditions. Machine learning approaches should be very powerful for these goals.
Modern AO systems often include multiple WFSs. For example, a laser tomography AO (LTAO) system has several laser guide stars and corresponding WFSs. An extreme-AO (ExAO) system may deploy sensors at multiple wavelengths to increase sensitivity and track chromatic terms.
Machine learning techniques should be able optimally combine multiple measurements in such systems. If a "truth" WFS is available, the EOF approach may be applicable.
While cacao can currently save telemetry from multiple sensors, tools to inspect/analyze such telemetry do not exist. Machine learning techniques exist to implement sensor fusion, and should be deployed in the future.
In some applications, knowledge of the point spread function (PSF) is essential. This is critically important for high contrast imaging system seeking to identify faint exoplanet images within a much brighter PSF. The AO system's WFS telemetry streams contain valuable information to estimate the focal plane PSF, as they record the WF errors that create PSF imperfections.
Non-linear machine learning approaches are particularly well suited to address this challenge.
Using focal plane images for wavefront sensing and calibration is especially relevant to high contrast imaging. Significant challenges come from non-linearity, chromaticity and partial coherence. We intend to support FP-WFS/C in the future.
Can be deployed on spatial or spatial+temporal dimensions. Need to skip consecutive frames to properly sample diverse conditions.
To be used on PCA coefficients for fast clustering. A kd-tree can also be built but is probably less efficient.
Looking for linear relationships between data sets.
Mutual information.
Nearest neighbor distances.
compute and control for adaptive optics (cacao) - https://github.com/cacao-org/cacao
- Real-Time OS install
- OS Performance Tuning
- Real-time OS benchmarks:
- GPU drivers and tools
- cacao Performance