Device-Aware Weighted Gerchberg-Saxton and Flicker-Bounded Phase-Continuous Sequences for Holographic Optical Trap Arrays

Phase-fixed weighted Gerchberg-Saxton (WGS) is the practical method for large arrays of optical traps, but it optimizes the hologram on an ideal model of the spatial light modulator (SLM), and the sequences that move traps interpolate between holograms that were never optimized for the frames in between. We test two hypotheses against the strongest baselines on the same scalar Fourier model. First, keeping the WGS weights but replacing the phase update by conjugate-gradient steps through the pixel-crosstalk kernel and the phase quantizer of the device (device-aware WGS) lowers the trap-intensity spread on the device. On a 10 x 10 lattice behind a crosstalk device, the relative standard deviation falls from a median of 5.21 percent for phase-fixed WGS at 30 iterations (5.21 percent at 100) to 0.205 percent, a median factor of 26 over 5 seeds (range 16 to 33) at a median 0.995 of the baseline's efficiency; the factor is 7.9 with 4-bit phase and a fill factor of 0.85 added, and 2.0 on a smaller 8 x 8 grid with flicker added and a fixed 30-iteration budget (the ratio of the means of 3 seeds); over the 5 geometries of the studio's benchmark on that device the median factor is 1.73 (0.97 to 4.69), and 2 of them meet both thresholds. A control that removes the device model from the loop gains nothing (5.14 percent), so the gain comes from the model. Second, optimizing every intermediate hologram of a moving-trap sequence with the trap phases locked to a smooth trajectory (flicker-bounded sequence) keeps the minimum transient intensity of the moving traps at a median 0.920 of their mean (range 0.915 to 0.932) over 5 seeds at 13 frames (below 0.9 at seven frames), against 0.654 (0.551 to 0.709) for linear phase interpolation. On longer moves, with the frames that keep the peak step per frame at most 0.2 Fourier units, the bound is met on 6 and 6 of 8 seeds of two lattice moves and on 5 of 6 seeds of a rearrangement case; the seeds that miss it miss it on the phase path their end phases prescribe, and at an equal number of frames on the rearrangement linear phase interpolation leads. Both results hold in a model of a simulated device and in a transient model that we define; neither is a hardware measurement. Interactive studio and reproducible pipeline (Apache-2.0): https://github.com/fsantibanezleal/CEFOP_DinHot . Interactive studio: https://dinhot.fasl-work.com .