muirSEAL: Infrared Photonic Lantern Testbed
- muirSEAL is a near-infrared testbed that validates photonic lantern wavefront sensing for segmented-telescope aberrations.
- It employs a segmented deformable mirror and dual optical paths to simulate low-order Zernike aberrations and segment piston errors.
- Bench tests show a peak injection throughput near f/4 and highlight challenges in alignment and nonlinear segment response.
Searching arXiv for the cited muirSEAL and SEAL 2.0 papers to ground the article in the latest preprints. {"query":"arXiv (Jensen-Clem et al., 3 Sep 2025) SEAL 2.0 reflective multi-wavelength rebuild muirSEAL", "max_results": 5} {"query":"(Sengupta et al., 27 Aug 2025) photonic lantern imaging wavefront sensing muirSEAL", "max_results": 5} {"query":"(Cuevas et al., 29 Aug 2025) Miniature Infrared SEAL segmented telescopes photonic lantern", "max_results": 5} Miniature Infrared SEAL (muirSEAL, “miniature IR SEAL”) is a near-infrared laboratory testbed within the broader Santa Cruz Extreme Adaptive Optics Laboratory (SEAL) program at the UCSC Laboratory for Adaptive Optics. It is installed inside SEAL’s enclosure but is otherwise separate from the main bench, and it was created as a dedicated all-infrared platform for experimentally validating photonic lantern imaging and focal-plane wavefront sensing at the lantern’s intended wavelength. In the published descriptions, its core use cases are low-order Zernike aberrations, segment piston or phasing offsets, and other structured aberrations relevant to high-contrast imaging on segmented telescopes (Jensen-Clem et al., 3 Sep 2025, Sengupta et al., 27 Aug 2025, Cuevas et al., 29 Aug 2025).
1. Origin within the SEAL 2.0 program
SEAL first saw light in 2021 as a transmissive, visible-wavelength adaptive-optics testbed. In 2024, it was rebuilt with 13 custom off-axis parabolic mirrors, preserving the original architecture while converting the relay to a reflective optical train. The stated consequences of that rebuild are removal of chromatic aberration, increased throughput, reduced mechanical footprint, and clean operation in both the near-infrared and the visible. Within that rebuilt environment, muirSEAL is described as the “miniature infrared Santa cruz Extreme AO Laboratory,” co-located with SEAL but functioning as an independent mini-bench optimized for infrared astrophotonics work (Jensen-Clem et al., 3 Sep 2025).
The immediate motivation for muirSEAL was methodological rather than architectural. Preliminary closed-loop photonic-lantern results were obtained on the main SEAL platform in 2023, but the lantern under test was designed for 1.55 µm while SEAL was operating at 0.633 µm. That wavelength mismatch caused the outputs to appear multi-modal and made it difficult to determine whether discrepancies arose from the lantern itself or from the mismatch between the lantern and the testbed. muirSEAL was therefore built as a dedicated infrared bench to test photonic lanterns under better matched conditions and to extend the study to a broader set of sensing scenarios (Sengupta et al., 27 Aug 2025).
A recurring point of clarification is that muirSEAL is not a replacement for the main SEAL bench. It is a technology-development offshoot of SEAL 2.0: not a science instrument, but a focused experimental platform for advancing photonic-lantern sensing and the associated hardware and software required for on-sky deployment (Jensen-Clem et al., 3 Sep 2025).
2. Optical architecture and instrumentation
muirSEAL is illuminated by a 1.55 µm laser source, specifically a Thorlabs KLS1550. The input beam passes through a 1 cm diameter iris and then reflects from a 169-actuator IrisAO deformable mirror. That segmented DM is used both to simulate atmospheric wavefront errors and to emulate primary-mirror segment phasing. The beam is intentionally undersized relative to the DM, so that the illuminated pupil covers only four rings of the DM’s eight-ring segmented aperture; this avoids two dead actuators and yields a controlled subaperture. In the wavefront-generation pipeline, the illuminated region is used to generate the Zernike basis by expressing each phase screen in terms of piston, tip, and tilt of each segment using HCIPy (Sengupta et al., 27 Aug 2025).
After reflection from the DM, the beam is divided into two paths. One path reimages the point spread function for direct monitoring, while the other injects light into the photonic lantern. A Thorlabs Elliptec four-position slider changes the focusing lens used to feed the lantern, and a linear translation stage moves the lantern input end to the correct focal plane for each lens. This permits injection studies at multiple beam speeds, specifically f-numbers from f/2 to f/12. Both the lantern output and the PSF-monitoring arm are reimaged onto a Goldeye CL-030 VSWIR TEC1 detector (Sengupta et al., 27 Aug 2025).
Within the larger SEAL 2.0 ecosystem, muirSEAL sits alongside a broader segmented-aperture and high-contrast-imaging infrastructure. The main SEAL bench includes a segmented IrisAO primary mirror simulating a 37-segment telescope aperture, additional deformable mirrors for turbulence and correction, visible wavefront sensors including a Shack-Hartmann WFS, a four-sided non-modulated pyramid WFS, and a vector Zernike WFS, and visible science paths including direct PSF imaging and a vector vortex coronagraph. muirSEAL constitutes the infrared and astrophotonics branch of that program (Jensen-Clem et al., 3 Sep 2025).
3. Photonic lanterns as the defining sensing element
The defining wavefront-sensing element associated with muirSEAL is the photonic lantern. In the astrophotonics framework used by the SEAL papers, a lantern can function both as an imaging element and as a wavefront sensor by transforming multimode input light into a set of single-mode outputs whose relative intensities or modal content encode phase information. This makes the device a focal-plane sensor rather than a conventional slope-based pupil sensor, and the stated attraction is precisely that such a sensor can remain sensitive to phase discontinuities and segment-misalignment signatures that are difficult for many traditional wavefront sensors to detect (Jensen-Clem et al., 3 Sep 2025, Cuevas et al., 29 Aug 2025).
The principal lantern used for wavefront sensing in the imaging-and-reconstruction study is the earlier CREOL photonic lantern, designed for 1.55 µm. It is a 19-port lantern with numerical aperture 0.11, corresponding to an optimal focal ratio of
It was tested at six f-numbers: f/2, f/3, f/4, f/6, f/8, and f/12 (Sengupta et al., 27 Aug 2025).
A second device, fabricated by Lawrence Livermore National Laboratory, broadens the scope from single-lantern characterization to comparative multi-lantern imaging. This LLNL device consists of seven 19-port photonic lanterns in one structure. Four are standard lanterns with equal output-port diameters, and three are hybrid lanterns with some output ports of differing sizes. Unlike the 1.55 µm CREOL lantern, the LLNL device is designed for 0.633 µm. The stated significance is that the multi-lantern geometry enables simultaneous imaging of the same PSF into multiple lanterns, which should eventually allow extraction of higher-order modal information and comparisons across manufacturing variants (Sengupta et al., 27 Aug 2025).
In the broader SEAL 2.0 roadmap, this photonic-lantern work is adjacent to, but distinct from, photonic coronagraphy and AstroPIC-style integrated photonic circuits. The programmatic distinction is explicit: muirSEAL is focused on photonic-lantern wavefront sensing rather than on focal-plane injection into an integrated photonic chip (Jensen-Clem et al., 3 Sep 2025).
4. Reconstruction models, calibration procedures, and sensing regimes
One major experimental program on muirSEAL evaluates a linear reconstruction pipeline based on modal interaction matrices. At each f-number, the system uses a 37-mode Zernike basis chosen to match the number of DM segments imaged in the illuminated pupil. Because the DM is off-center and excludes two dead actuators, the basis is constructed directly in terms of segment pistons, tips, and tilts using HCIPy, and its suitability is checked by inspecting the PSF shape under application of each mode (Sengupta et al., 27 Aug 2025).
For the 19-port lantern measurements in that study, the outputs are extracted by aperture photometry. Approximate port centers are identified manually, refined by center-of-mass calculations on image cutouts, and then measured using circular masks typically around 6 pixels in radius, corresponding to about 30 µm, with slight adjustment depending on alignment. The measurement vector is the summed intensity in each port mask, normalized by the total intensity across all 19 measured channels. A linear model is then built by applying positive and negative pokes in each Zernike mode to form an interaction matrix, which is inverted to a command matrix using a singular value threshold of 1/30. Reconstruction performance is assessed via linear sweeps in each Zernike mode, and comparison is truncated to the first 9 modes for consistency with earlier SEAL laboratory work. The dynamic-range criterion is the first “overturning,” defined as the first point where the reconstructed coefficient decreases even though the applied coefficient is still strictly increasing (Sengupta et al., 27 Aug 2025).
A complementary line of work on segmented-mirror phasing uses muirSEAL to reconstruct piston offsets from lantern measurements. In that study, piston aberrations are injected in the range from rad to $1$ rad, approximately . The laboratory implementation differs from the aperture-photometry pipeline: the full photonic-lantern image is used as the wavefront-sensing signal rather than a reduced vector of per-port photometries. The bench interaction matrix is built using only the six innermost segments of the segmented mirror and a poke amplitude of 0.1 radians (Cuevas et al., 29 Aug 2025).
The associated theoretical framework linearizes the lantern response around a flat reference state. In simulation, the authors construct a push-pull interaction matrix by applying tiny positive and negative piston offsets with amplitude , recording the normalized output at each of the 19 ports, and forming a command matrix from the resulting responses. In the linear regime, the relation is written as
where is the photonic-lantern measurement vector and is the segment-piston vector. The estimated phase is then obtained by multiplying the normalized, flat-subtracted lantern signal by the command matrix (Cuevas et al., 29 Aug 2025).
Because the linear regime is limited, the segmented-piston study also investigates nonlinear reconstruction. The nonlinear model is implemented in PyTorch as a multi-layer neural network with two hidden layers, 2000 neurons in the first hidden layer, 100 neurons in the second hidden layer, and ReLU activations. Training uses 45,000 measurements generated by randomly applying piston offsets in the interval to $1$ radians to segments 1 through 9, optimized with mean squared error loss and the Adam optimizer at learning rate 0.001. The authors report that the network can capture nonlinear segment response, but it struggles in several cases to predict when a segment has no piston offset; they suggest adding a direct linear component so that the network learns deviations from the linear approximation rather than the full mapping (Cuevas et al., 29 Aug 2025).
5. Experimental performance and observed limitations
The most fully quantified validation result on muirSEAL concerns lantern injection throughput as a function of input f-number. The experiment compares total counts through the photonic lantern to total counts at the input by temporarily moving the detector to the lantern input plane, while adjusting tip, tilt, and focus at each f-number to maximize coupling. The resulting throughput curve shows a peak near f/4, consistent with the lantern’s numerical aperture, together with a double-peak structure with peaks at f/4 and f/8 and a sharp roll-off around f/4. The best measured injection efficiency is 31.6%, substantially below the 97–98% coupling efficiencies previously reported for near-IR single-wavelength lanterns. The expected value was roughly 76% throughput at f/4, based on a design prediction of about 79% from geometric scaling multiplied by the previously characterized 97% intrinsic efficiency, so the measured result is about a factor of 2.5 lower than expected. The stated interpretation is that the discrepancy is likely dominated by pre-lantern optics alignment and possibly input positioning differences rather than intrinsic lantern loss, with strong sensitivity to tip and tilt and a likely dependence on improved x/y placement and more complete coupling maps (Sengupta et al., 27 Aug 2025).
Wavefront-sensing behavior on muirSEAL exhibits usable sensitivity but not uniform linearity. The reported trends are wide linear ranges in some modes, strong crosstalk except near zero aberration, and mode- and f-number-dependent dynamic range without a strong universal trend. The study notes a possible reduction in dynamic range at f/3 and a possible increase in dynamic range for coma or tricoma-like modes, specifically modes 6 through 9. The largest f-numbers, f/8 and f/12, appear to perform better at large wavefront errors, although those points are also described as generally beyond the practically relevant dynamic range. A central limitation of this first campaign is large reconstruction crosstalk, worse than in earlier SEAL work; suggested causes include the comparatively lower-order DM on muirSEAL, alignment differences across the six f-number configurations, or lantern-to-lantern differences (Sengupta et al., 27 Aug 2025).
For segmented-piston reconstruction, the laboratory validation shows that the photonic lantern can encode piston information in a real segmented-aperture optical setup. The measured bench curves display wider linear ranges than the simulation, minimal cross-talk at small aberrations, and a linear range of about 0.5 radians, corresponding to 123.35 nm. The same study explicitly compares that regime with a prior Keck result from Salama et al. (2024), where a total RMS of 128 nm was reported after closed-loop control with a vector-Zernike sensor, and argues that the photonic-lantern system may therefore operate in a comparable correction regime (Cuevas et al., 29 Aug 2025).
Those bench results are nonetheless partial and include important caveats. The interaction matrix uses only the six innermost segments, so the demonstration is not a full-aperture phasing result. Segment-to-segment variation is significant, and the response for segment 4 shows a broad dynamic range but also an immediate inversion near zero aberration, so that a positive poke can be reconstructed as negative and vice versa. The interpretation given is that the lantern does encode segment information, but that uncharacterized noise may be affecting the bench data. Further study of the lantern’s noise properties and of the signal-to-noise ratio is therefore identified as necessary (Cuevas et al., 29 Aug 2025).
The initial imaging of the LLNL lantern provides a different sort of validation. The single-mode end was imaged in the laboratory and compared qualitatively with a simulation generated using the lightbeam Python package. The comparison is explicitly qualitative only because diffraction-limited conditions were not independently verified, but the laboratory image and the simulation were reported to be broadly similar in structure. The principal discrepancy was in the relative brightness of the ports, which was attributed likely to optical alignment imperfections (Sengupta et al., 27 Aug 2025).
6. Scientific role, operational context, and future development
muirSEAL’s scientific role is to serve as a bridge between simulation-level photonic-lantern concepts and deployable focal-plane wavefront sensing for segmented telescopes. The most explicit target problem is segment piston or co-phasing error, one of the most difficult aberration classes in high-contrast imaging for future large segmented telescopes. In this sense, muirSEAL occupies a distinctive niche within SEAL 2.0: classical wavefront sensors on the main bench address visible-light pupil-plane or focal-plane sensing, whereas muirSEAL develops an infrared astrophotonic alternative that is compact, focal-plane based, and directly tied to guided-wave optics (Jensen-Clem et al., 3 Sep 2025, Cuevas et al., 29 Aug 2025).
This role is reinforced by the broader SEAL 2.0 control and calibration environment. SEAL now includes Catkit2 and CACAO as real-time control frameworks, in addition to its original Keck-heritage RTC. Catkit2 is described as a service-oriented, low-latency distributed control architecture, while CACAO provides a common shared-memory format and greater flexibility. Within that environment, the adaptive-optics loop was successfully closed using the pyramid wavefront sensor at about 15 Hz, the PWFS readout reached up to 200 Hz after pupil-size reduction, and the non-modulated PWFS startup procedure yielded a residual error of about 18 nm rms after closed-loop correction. The rebuilt infrared PSF branch achieved a measured Strehl ratio of 98% after closing the loop with the visible PWFS, corresponding to an estimated NCPA of about 30 nm rms between the IR PSF and the visible PWFS. These are SEAL 2.0 system-level results rather than standalone muirSEAL metrics, but they define the multi-wavelength calibration environment in which infrared lantern sensing is being developed (Jensen-Clem et al., 3 Sep 2025).
The platform also has a telescope-deployment dimension. The SEAL 2.0 paper states that muirSEAL supports hardware and software development for the photonic lantern deployed at Lick Observatory, which places the work in an on-sky development context rather than a purely laboratory one (Jensen-Clem et al., 3 Sep 2025).
A second point of clarification concerns scope. The broader SEAL program encompasses bulk-optic coronagraphy, classical wavefront sensing, focal-plane sensing, and astrophotonic sensing. muirSEAL is complementary to this wider program but narrower in mission: it focuses on photonic-lantern wavefront sensing rather than on integrated photonic coronagraphy, and it remains a development testbed rather than a finalized instrument (Jensen-Clem et al., 3 Sep 2025).
The near-term technical agenda follows directly from the reported limitations. The papers identify the need for improved alignment control, better understanding of coupling sensitivity, extension to nonlinear reconstructors and closed-loop control, more faithful simulations of the laboratory setup, and explicit study of noise propagation in the lantern signal. Taken together, these directions suggest a staged maturation path: first demonstrating that segment information and low-order aberrations are encoded in photonic-lantern outputs, then determining where linear models suffice, and finally establishing whether nonlinear reconstruction can support robust operation in the larger-aberration regime relevant to future space- and ground-based segmented-mirror telescopes (Sengupta et al., 27 Aug 2025, Cuevas et al., 29 Aug 2025).