Calibration-assisted stress-optic phase programming for a 3D-printed optical neural-network frontend

Abstract

We report a calibration-assisted stress-optic phase-programming method for a 3D-printed optical neural-network (ONN) frontend. Full-vector stress-optic modeling gives a full-stress effective-index shift of Δ ⁢ 𝑛 e f f = − 2 . 1 2 8 4 × 1 0 − 4 at 850 nm, setting the phase-length rule for Mach–Zehnder interferometer (MZI) phase shifters. A matched MPB modal proxy reproduces the modeled phase sensitivity with a ratio of 0.999,396, giving phase errors below 0 . 0 0 1 2 ⁢ 𝜋 for the evaluated 4- and 8-node phase sets. MZI measurements across three annealing conditions show a monotonic reduction in fringe-shift magnitude; a separately defined ± 5 % gain-error envelope is used only for sensitivity analysis. The shortest 𝑠 = 1 design is tolerance-limited, but an 𝑠 d e s i g n = 0 . 8 5 process window restores a simulated 100% pass rate under ideal one-time per-node gain estimates, with a 17.65% length penalty. Mapping the resulting phase noise into a 64-mode MNIST optical-front-end benchmark gives 83.44% ideal accuracy and calibrated recovery of the ideal numerical reference. Mode-field maps and a stage-resolved projected | 𝐸 | 2 propagation map visualize the physical basis and propagation path while preserving the distinction between transverse modal fields and system-level power interpolation.