ORCID Number

0000-0002-0067-3948

Date of Award

Spring 5-4-2026

Embargo Period

8-31-2026

Access Type

Dissertation - Open Access

Degree Name

Doctor of Philosophy in Electrical Engineering & Computer Science

Department

Electrical Engineering and Computer Science

Committee Chair

Eduardo Rojas

Committee Chair Email

rojase1@erau.edu

Committee Advisor

Eduardo Rojas

Committee Advisor Email

rojase1@erau.edu

Committee Co-Chair

Tianyu Yang

Committee Co-Chair Email

yang482@erau.edu

First Committee Member

Eduardo Rojas

First Committee Member Email

rojas1@erau.edu

Second Committee Member

Tianyu Yang

Second Committee Member Email

yang482@erau.edu

Third Committee Member

M. Ilhan Akbas

Third Committee Member Email

akbasm@erau.edu

Fourth Committee Member

Xuanhui Wu

Fourth Committee Member Email

xuanhui.wu@mnsu.edu

Fifth Committee Member

Qi Cheng

Fifth Committee Member Email

qi.cheng@okstate.edu

College Dean

James W. Gregory

Abstract

This dissertation studies three physical-layer optimization problems that share a common difficulty: the forward model, whether an electromagnetic solver or a statistical channel model, is either too expensive to evaluate at design scale or too unstable to hold across a single processing block. The first two problems are treated in their own right, as problems in antenna design and in multi-user detection over time-variant channels; the third carries the methodology onto a coherent optical link, where size, weight, and power constraints make a small-satellite deployment the natural target.

The antenna work targets a tri-layer pixelated unit cell fabricated on Avient ABS550 by fused-deposition modeling, operated from 8 to 12 GHz. The cell has a 4×4 conductor pattern on its top and bottom faces and one of two middle-layer geometries (metal plate or centric ring), which together cover the 36-state, 10-degree-quantized phase library used for beam forming. Characterizing every pattern in HFSS is infeasible, so a one-dimensional convolutional network with cross-attention is trained to predict the scattering parameters from the binary pixel layout and frequency. Locality is preserved through Hilbert ordering, and passivity, causality, and amplitude smoothness are enforced through additional loss terms. Across 933 held-out geometries, the surrogate keeps the passivity-violation rate under 12% and the mean resonance-frequency error under 0.5 GHz, and it runs roughly 180,000 times faster than the full-wave solver. A 509-element circular aperture of diameter 256 mm, populated by the surrogate and fed by a WR-90 pyramidal horn, reaches a realized directivity of 14.02 dBi at 10 GHz in HFSS; the same cell under Floquet-port excitation shows a peak directivity of 27.77 dBi. A WR-90 waveguide measurement model is also developed, and the associated de-embedding analysis links the surrogate, the Floquet simulation, and the planned physical measurement.

The multi-user-detection work addresses BSS receivers in channels whose mixing matrix drifts within one block. Five data-injection strategies are considered: no injection, linear 2x/4x/8x upsampling, sign reversal, cubic spline, and Hermite/PCHIP. Rather than fixing one of these at design time, an Echo State Network estimates the channel variation index directly from the received mixture, without pilot symbols, and a LinUCB contextual bandit or a Deep Q-Network selects the injection mode at each block using a context vector built from the estimate, the block size, and their coherence product. On the evaluation grid the improved ESN achieves an 85.8% estimation success rate, and the DQN agrees with the Monte Carlo oracle on 71.2% of cells while staying within 0.2 to 0.5 dB of the oracle's SIR. The agreement holds in the extrapolation regions of the (delta, N) grid that were not seen during training.

The two methods come together in the third study, which targets coherent optical links for non-terrestrial networks. On a bench-top 1550 nm receiver, two 1 Gb/s co-channel carriers are recorded through a single telescope under emulated atmospheric scintillation with normalized intensity variance of approximately 0.99, and the Sign Reversal injection is applied because it preserves the centrosymmetry of the BPSK constellation while doubling the effective sample count. The demixing matrix is re-estimated on a duty cycle rather than at every block. With this scheme the receiver loses less than 1 dB in SIR relative to full-rate FastICA and converges in roughly 3.5 times fewer iterations, keeping the DSP core within the sub-watt envelope of a 6U to 16U FPGA payload. Before the optical link can close, the two terminals have to find each other, and this is where the transmitarray returns: the same reconfigurable aperture is repurposed as a millimeter-wave sensing beam that drives an agile beaconless laser alignment (ABLBA) controller, narrowing the field of uncertainty into which the optical spiral search is launched. Acquisition time is analyzed in closed form against inter-CubeSat distance, position variance, and antenna half-power beamwidth, and the millimeter-wave-guided scheme yields reductions over a pure optical scan that grow with propagation distance.

Download

Available for download on Monday, August 31, 2026

Share

COinS