Photon-Counting Noise and Ornstein–Uhlenbeck Heater Drift in In-Situ Training of Quantum Photonic Neural Networks: An Equal-Budget Benchmark of Derivative-Free and Finite-Difference Optimizers

Data and code for the article "Photon-Counting Noise and Ornstein–Uhlenbeck Heater Drift in In-Situ Training of Quantum Photonic Neural Networks: An Equal-Budget Benchmark of Derivative-Free and Finite-Difference Optimizers" by J. Bhuvana Gowri and T. Kavitha (Vel Tech Rangarajan Dr. Sagunthala R&D Institute of Science and Technology, Chennai, India), submitted to IEEE Photonics Journal. The study simulates a six-layer, four-mode quantum photonic neural network trained as a Bell-state analyzer, with realistic coupler errors and propagation loss, an effective Kerr phase of π/10, heater phase drift modeled as an Ornstein–Uhlenbeck process, and photon counting modeled as binomial sampling of 250 or 2,500 photon pairs per input. Seven optimizer variants (BOBYQA, standard and constant-gain SPSA, finite-difference Adam with nudges of 0.01, 0.1 and 0.3 rad, and random search) are compared with an equal budget of 30,000 measurements on 30 simulated chips and 26 noise settings (5,460 runs). Contents:- code/: simulator (based on the quotonic package), measurement interface, optimizers, exact reference trainer, tuning, main study, statistical analysis, pilot studies.- tests/: validation tests (Hong–Ou–Mandel interference, Clements mesh, linear-optics limit, Kerr element, loss limit, reproduction of published results, drift and counting models, optimizers, scoring, finite-difference noise model, statistics code).- results/: all main runs, statistics (Wilcoxon tests with Holm correction, Wilson intervals), exact reference runs, tuning runs, pilot results, noise-model data and the printed output of the validation tests.- figure_data/: the data of every figure as CSV files and one Excel workbook.- figure_scripts/: scripts that rebuild the figure data from the results and draw the figures.- audit/: independent recomputation of all reported numbers from the raw run files. Software: Python with NumPy 2.4.6, JAX 0.10.2, quotonic 1.0.1 and NLopt 2.11.0 (see requirements.txt and README.md for instructions to reproduce the results).