WaveDriver: RF, Adaptive Optics & GPU Control

• WaveDriver is a multi-context term describing technologies that enable precise RF and optical wave control for modulators and adaptive-optics systems.
• It features an open-source RF driver covering 10 MHz–1.1 GHz with fast digital switching, 30 dB analog tuning, and high-power amplification.
• A GPU-driven architecture under the WaveDriver label achieves real-time, multi-tone RF synthesis for applications such as optical tweezer arrays.

WaveDriver is a domain-dependent term in recent arXiv literature rather than the name of a single platform. It denotes an open-source radio-frequency driver for acousto-optic and electro-optic devices (Barker et al., 2019), a laser guide star adaptive-optics concept for the Habitable Worlds Observatory (HWO) (Gerard et al., 23 Sep 2025), and, in a separate engineering framing, a GPU-driven real-time arbitrary waveform driver for large multi-tone RF synthesis (Tu et al., 2024). Within the HWO program, the name also encompasses a photonic-lantern focal-plane wavefront-sensing and integral-field-spectroscopy subsystem that has been fabricated and characterized in laboratory settings (Sengupta et al., 25 Jun 2026).

1. Scope and nomenclature

The recent literature uses the name WaveDriver in at least two distinct scientific contexts. In atomic, optical, and laboratory instrumentation, it refers to a flexible RF driver optimized for acousto-optic modulators (AOMs) and electro-optic modulators (EOMs), with wide amplification bandwidth, analog AM/FM capability, and fast digital gating (Barker et al., 2019). In exoplanet instrumentation and space observatory systems engineering, it refers to a formation-flying laser guide star (LGS) spacecraft coupled to an onboard adaptive-optics (AO) system for HWO, intended to support picometer-level wavefront stability (Gerard et al., 18 May 2026).

A closely related usage appears in the photonic-lantern work conducted for the HWO effort. There, WaveDriver is the project context for a natural-guide-star focal-plane wavefront sensor (WFS) and integral field spectroscopy (IFS) unit built from an array of seven 19-port photonic lanterns, characterized by digital off-axis holography at 632.8 nm (Sengupta et al., 25 Jun 2026). A separate high-performance computing paper is framed in the supplied material as “WaveDriver: a GPU-driven real-time arbitrary waveform driver,” describing a GPU-plus-DAC architecture for additive RF synthesis in optical tweezer experiments (Tu et al., 2024).

The principal misconception is therefore terminological: WaveDriver is not a single canonical device class. In the cited literature, it names several systems whose common feature is precise control or sensing of RF waveforms, optical wavefronts, or both.

2. Open-source RF WaveDriver for AOMs and EOMs

The RF WaveDriver presented in 2019 is a flexible, open-source driver for optically active devices such as AOMs and EOMs (Barker et al., 2019). Its stated design motivation is to cover disparate device classes with one driver: low-power AO in the visible, high-power low-frequency AO in the near-IR, resonant free-space EO, and broadband fiber EO. The design combines a wide RF bandwidth of 10 MHz–1.1 GHz, high output power, fast digital switching, analog power tuning over roughly 30 dB, and integrated as well as external RF-source selection.

The hardware uses a two-PCB architecture with total PCB area below 255 cm², both boards implemented on FR4 and connected by a short SMA jumper. The control and preamplifier board contains RF source selection through SW2, choosing between an integrated voltage-controlled oscillator (VCO) and an external RF input; an HMC8410 preamplifier providing about 20 dB gain; an ADC-20-4 directional coupler tapping about 1% of the RF to “RF Mon”; an F2255 voltage-variable attenuator (VVA) providing about 30 dB linear-in-dB AM range; and RF switches SW1, SW3, and SW4 for digital power gating and cross-talk suppression. Control electronics include a microcontroller, DACs for VVA and VCO setpoints, optocouplers, instrumentation amplifiers, op-amp buffers, and a front-panel user interface. Proper biasing of the HMC8410 is enforced by an HMC920LP5E controller implementing gate/drain sequencing and constant-IDD bias.

The high-power amplifier board uses an HMC1099 GaN power amplifier as the output stage, prematched to 50 $\Omega$, with saturated output power above 10 W and power-added efficiency of at least 50% under operating conditions. Gate/drain bias sequencing and interlocks are implemented with an LTC2924. Thermal management is handled by a backside BGA heatsink, an optional 24 V fan, more than 100 thermal vias, and copper spreading. The ordering of the switches and VVA is explicitly chosen to maximize extinction and AM dynamic range while accepting insertion loss through SW1, the coupler, and the VVA.

The electrical formulation is standard 50 $\Omega$ RF practice. The paper states

$$P = \frac{V_{rms}^2}{R}, \qquad V_{rms} = \sqrt{PR},$$

with examples of approximately $7.07$ V rms and $20$ Vpp at 1 W, and approximately $15.8$ V rms and $44.7$ Vpp at 5 W. Gain is expressed as $G(f)=|S_{21}(f)|^2$, return loss as $RL=-20\log_{10}|S_{11}|$, and impedance mismatch by

$$\Gamma = \frac{Z_L-Z_0}{Z_L+Z_0}, \qquad VSWR = \frac{1+|\Gamma|}{1-|\Gamma|}.$$

These definitions are operationally important because the amplifiers are prematched to 50 $\Omega$0 and the design requires 50 $\Omega$1 terminations at all ports to protect the MMICs (Barker et al., 2019).

3. RF performance, modulation, and operating constraints

The reported RF coverage is at least 1 W from 10 MHz to 1.1 GHz for 1 mW input drive, extending to about 1.2 GHz with higher input power (Barker et al., 2019). At higher drive, the driver provides at least 5 W from 20 MHz to 100 MHz. For the integrated VCO path with a 3 dB fixed attenuator after the VCO, typical mean output powers are 5.3 W at 40 MHz, 4.6 W at 80 MHz, 3.7 W at 200 MHz, 1.9 W at 400 MHz, and 1.4 W at 800 MHz. The dynamic range of continuous analog power control is about 30 dB, limited at the top end by gain compression in the high-power amplifier.

Analog amplitude modulation is implemented through the F2255 VVA and shows a small-signal $\Omega$2 dB bandwidth of about 70 kHz, with bandwidth increasing modestly with modulation depth. Analog frequency modulation is applied through the VCO tuning voltage $\Omega$3, with overall driver buffering supporting instrumentation-input bandwidths up to about 10 MHz, while the actual FM response is set by VCO characteristics. Digital switching performance is substantially faster: the envelope begins falling about 15 ns after the TTL edge, reaches $\Omega$4 dB in about 25 ns, falls below the measurement floor of $\Omega$5 dB within about 40 ns, and achieves final extinction above 90 dB across the full band and above 110 dB below roughly 400 MHz. The extinction ratio is defined as

$\Omega$6

The MASWSS0178 RF switches have about 20 ns intrinsic switching time, and the full system shows about 15 ns propagation delay from optocouplers and logic to onset. Turn-off is faster than turn-on, so practical digital modulation is limited to about $\Omega$7 MHz at 50% duty cycle because of asymmetric rise and fall envelopes. Harmonic distortion increases with output power and is most pronounced below 400 MHz; the reported harmonics are generally below $\Omega$8 dBc within the operating band and below $\Omega$9 dBc above about 700 MHz. For narrowband modulators, device-limited filtering ameliorates harmonics, whereas broadband fiber EOM use cases may require external bandpass filters.

Thermal and safe-operation limits are central to the design. The dissipated power is written as

$$P = \frac{V_{rms}^2}{R}, \qquad V_{rms} = \sqrt{PR},$$0

and the junction temperature estimate is

$$P = \frac{V_{rms}^2}{R}, \qquad V_{rms} = \sqrt{PR},$$1

At approximately 5 W RF output and about 50% PAE, dissipation is of order a few watts in the power amplifier. With forced-air cooling, the HMC1099 remains within about $$P = \frac{V_{rms}^2}{R}, \qquad V_{rms} = \sqrt{PR},$$2C of ambient at all input powers. Without a fan, the heatsink rises about $$P = \frac{V_{rms}^2}{R}, \qquad V_{rms} = \sqrt{PR},$$3C above ambient at $$P = \frac{V_{rms}^2}{R}, \qquad V_{rms} = \sqrt{PR},$$4 mA with no RF and about $$P = \frac{V_{rms}^2}{R}, \qquad V_{rms} = \sqrt{PR},$$5C above ambient at roughly 5 W heat, so the case temperature can approach $$P = \frac{V_{rms}^2}{R}, \qquad V_{rms} = \sqrt{PR},$$6C in warm rooms. The paper therefore warns never to operate without a 50 $$P = \frac{V_{rms}^2}{R}, \qquad V_{rms} = \sqrt{PR},$$7 load, never to sequence amplifiers without a load connected, and to follow the prescribed turn-on procedure: rocker OFF during AC mains transitions, rocker ON to sequence rails, and only then assert $$P = \frac{V_{rms}^2}{R}, \qquad V_{rms} = \sqrt{PR},$$8.

The design package is explicitly open-source and includes schematics, PCB layouts, BOMs, firmware, microcontroller code, and DDS-shield design files. The stated cost is about \$P = \frac{V_{rms}^2}{R}, \qquad V_{rms} = \sqrt{PR},$91,085 total per unit in a production run of 40 including enclosure and auxiliaries (Barker et al., 2019).

4. WaveDriver as an HWO laser guide star adaptive-optics architecture

In the HWO literature, WaveDriver is a mission concept in which an LGS spacecraft formation-flies with HWO and feeds an onboard AO system intended to stabilize both low-order wavefront modes and primary-mirror segment piston, tip, and tilt (Gerard et al., 23 Sep 2025). The central requirement is a system-level wavefront stability of 10 pm rms over 10 minutes per spatial mode, denoted $7.07$0, and the papers frame this as more than a 100$7.07$1 tighter wavefront-stability target than JWST (Gerard et al., 18 May 2026). The motivation is HWO’s Tier 1 Contrast Stability Technology Gap and the need to maintain approximately $7.07$2 contrast for Earth-analog exoplanet imaging.

The small-error contrast argument is stated in the form

$7.07$3

with $7.07$4 the rms wavefront error and $7.07$5 a coronagraph-dependent factor (Gerard et al., 23 Sep 2025). At visible wavelengths, picometer-level residuals therefore correspond to very small phase errors and quasi-static speckle intensities that must remain sufficiently stable for closed-loop suppression and post-processing. The HWO WaveDriver architecture operates jointly with internal segment metrology and adds an external bright beacon for high-temporal-bandwidth sensing and correction.

The control studies compare an optimal leaky integrator, an optimal LQG controller with an AR(2) disturbance model, and a machine-learning predictive controller implemented on FPGA (Gerard et al., 23 Sep 2025). The standard state-space form is

$7.07$6

with quadratic cost

$7.07$7

The literature emphasizes that the dominant limitation is not measurement noise but bandwidth error imposed by finite exposure and actuation rates. In the representative cases studied, the primary-mirror metrology actuator loop is limited to 10 Hz, and the papers report that this alone cannot achieve $7.07$8 for worst-case integrated disturbances when the spectral break lies in the Hz range (Gerard et al., 18 May 2026).

The principal quantitative findings are consistent across the first-phase and first-year summaries. A Zernike WFS is the most sensitive of the tested LGS WFS options for segment modes; an LGS of apparent magnitude $7.07$9 is required to reach $20$0 for primary-mirror segment stabilization; and a second-stage, high-temporal-bandwidth DM loop is needed to close the gap left by the 10 Hz metrology loop (Gerard et al., 23 Sep 2025). For $20$1 pm rms with high-frequency slope $20$2, the studies report that $20$3 is achievable with latency $20$4 ms and frame period $20$5 ms, provided a high-bandwidth DM stage is used (Gerard et al., 18 May 2026). The same framework is used to argue that WaveDriver can relax requirements on passive thermal and structural stability above about 1 Hz, on metrology and segment-actuator bandwidth, and on low-order control based solely on natural guide stars.

The low-order sensing case is more nuanced. One formulation reports that a perfectly achromatic Zernike WFS on a bright on-axis star of $20$6 yields photon-noise-limited residual $20$7 pm rms averaged over the first 15 Zernike modes (Gerard et al., 23 Sep 2025). A later synthesis gives an achromatic upper-limit comparison of $20$8 pm rms for $20$9 averaged over the first 15 Zernike modes (Gerard et al., 18 May 2026). In both formulations, the implication is the same: low-order stabilization to $15.8$0 is difficult with natural guide stars alone, and the WaveDriver beacon plus complementary natural-guide-star sensing is introduced to address that regime.

5. Photonic-lantern wavefront sensing and integral field spectroscopy in WaveDriver

A photonic lantern is central to the HWO WaveDriver concept because it can support simultaneous focal-plane wavefront sensing and spectroscopy (Sengupta et al., 25 Jun 2026). In the proposed architecture, the photonic lantern sits at the science focal plane. Its multimode input collects starlight, and the lantern adiabatically maps that input to $15.8$1 single-mode outputs. The same outputs are then used in two ways: their intensities encode wavefront information for low-order sensing, and they can be spectrally dispersed for IFS. The arrangement is motivated by the need to minimize non-common-path error while retaining science throughput.

The fabricated “Livermore lantern” is a monolithic glass device containing seven photonic lanterns, each with 19 single-mode ports, for a total of $15.8$2 ports (Sengupta et al., 25 Jun 2026). Two layouts were made. The “A” design is a standard 19-core hexagonal pattern with uniform port sizes; the “B” design alternates three core sizes, using a depressed-well structure to broaden the single-mode operation band at the expense of throughput. The reported structural parameters include numerical aperture $15.8$3, $15.8$4, $15.8$5, $15.8$6, cladding diameter $15.8$7m, taper factor 7.14, taper length 1 cm, and central wavelength design near 633 nm. The A-design core diameter is $15.8$8m with $15.8$9m spacing, while the B design uses core diameters of 4.04, 4.60, and $44.7$0m with $44.7$1m spacing.

The experimental characterization uses digital off-axis holography to measure principal modes by backward injection through the single-mode ports (Sengupta et al., 25 Jun 2026). The recorded intensity obeys

$44.7$2

where $44.7$3 is the unknown object field at the multimode end and $44.7$4 is a tilted reference field. Off-axis tilt shifts the cross terms in Fourier space; masking one side lobe and inverse transforming retrieves a complex field proportional to the object. The transfer-matrix description is equally standard: at a single wavelength, an $44.7$5-port lantern is represented by an $44.7$6 complex matrix relating modal amplitudes at the multimode input to single-mode output amplitudes, while principal modes correspond to the multimode fields generated by backward injection from each single-mode port.

The measured mode set is then used to analyze WFS response. Any multimode input field can be expanded as

$44.7$7

with coefficients estimated by projection onto the measured principal modes. The paper also defines an $44.7$8 coupling matrix through overlap integrals of the retrieved complex fields, distinguishing this correlation matrix from the transfer matrix itself. Experimentally, the seven lanterns show appreciable device-to-device variation. Some A and B lanterns exhibit strong upper-left clustering in the correlation matrix, diagonal banding is weak, and several lanterns show short-range adjacency structures in which three to four neighboring ports correlate.

Characterization was performed at HeNe 632.8 nm. Absolute throughput and insertion loss were not reported because exposure times and neutral-density settings varied during acquisition. For sensitivity estimation, all modes within a lantern were instead normalized to simulated peak-throughput values of 0.946 for the A design and 0.905 for the B design. The resulting photon-noise sensitivity curves show peaks and troughs at roughly shared spatial frequencies across lanterns. The main sensitivity peak shifts slightly to higher spatial frequencies than in simulation, which the authors associate with the experimental system favoring a slightly smaller f-number than assumed. Three A lanterns show consistent, higher sensitivity; “righta” is lower in sensitivity and shows strong inner-ring correlations; and the B lanterns generally underperform simulation because the simulated B design did not include the depressed well. Despite these variations, the reported conclusion is that all lanterns retain sensitivity to the expected low-order aberration modes and support the expected closed-loop WFS capability of WaveDriver’s natural-guide-star channel (Sengupta et al., 25 Jun 2026).

6. GPU-driven arbitrary waveform generation under the WaveDriver label

In the supplied engineering framing, WaveDriver also denotes a GPU-driven architecture for real-time arbitrary waveform generation, built from an NVIDIA GPU and a commercial high-speed DAC card on a desktop PC (Tu et al., 2024). The implementation uses a Quadro RTX 6000 with 24 GB GDDR6 and a Spectrum Instrumentation M4i.6622-x8 card with four 16-bit DAC channels. The synthesis model is additive and data-parallel: single-frequency tones are generated independently and then summed, a structure the paper describes as “embarrassingly parallel.”

The general waveform model is

$44.7$9

with time-varying amplitude, instantaneous frequency, and phase per tone (Tu et al., 2024). For static circular buffers, the reported quantization condition is

$G(f)=|S_{21}(f)|^2$0

to avoid discontinuities at chunk boundaries. For dynamic operation, the paper discusses linear chirps and a minimum-jerk chirp profile used for optical tweezer transport.

The dataflow keeps wavetables and sum buffers in GPU memory and transfers final waveform chunks by RDMA to the DAC FIFO. The optimal RDMA payload is reported as 2 MiB per transfer. With four channels, the SCAPP SDK multiplexes this into four 0.5 MiB sub-chunks, giving a per-channel chunk length of $G(f)=|S_{21}(f)|^2$1 samples. With all four channels active, the sustained aggregate RDMA rate is 2.8 GB/s, limited by the card’s PCIe 2.0 path. Typical operation uses 560 MB/s per channel, corresponding to 280 MS/s at 16-bit resolution.

Two dynamic-control pathways are reported. In “Playback,” the full dynamic waveform is precomputed in GPU memory and then streamed to the DAC. In “Streaming,” each upcoming chunk is computed inside the time budget between RDMA transfers. The measured capabilities are arbitrary simultaneous chirping of 1000 Nyquist-limited single-frequency tones at 280 MS/s for a limited duration of 35 ms in Playback mode, and continuous Streaming of 194 tones at 50 MS/s or 20 tones at 280 MS/s (Tu et al., 2024). For static waveform computation, the GPU achieves a 586-fold speedup over the CPU baseline in double precision and a 3389-fold speedup in single precision.

The intended application is control of acousto-optic deflectors for large optical tweezer arrays. In that setting, each RF tone generates one diffraction order, frequency controls deflection angle, amplitude controls trap depth, and chirps move tweezers across the field. The architecture is therefore a waveform engine rather than a modulator driver in the 2019 sense and is conceptually separate from the HWO LGS architecture. Its inclusion under the WaveDriver label reflects a shared emphasis on high-rate, high-dimensional control of wave-based hardware rather than a shared hardware lineage (Tu et al., 2024).