By Marie
Programmable photonic processors, chips that manipulate light through arrays of interconnected waveguides, promise faster, more energy-efficient computing for both classical and quantum applications. But they have a persistent problem: fabrication imperfections mean every chip behaves slightly differently, and calibrating them to perform arbitrary computations is notoriously difficult.
A new preprint from researchers at the Universitat Politecnica de Valencia, Universite Paris-Saclay, and Queen’s University presents a tandem neural network framework that substantially improves calibration precision on these devices, and, crucially, generalizes to arbitrary optical operations rather than only the ones the network was trained on.
The problem
Photonic processors use mesh networks of Mach-Zehnder interferometers (MZIs), optical elements that split and recombine light, to perform programmable unitary transformations on light signals. A 3×3 mesh uses 6 MZIs; a 4×4 mesh uses 10. Each MZI requires precise current tuning to achieve the desired phase shift. Fabrication variations, thermal crosstalk, and environmental drift mean that the same current setting produces different results on different chips, and even on the same chip at different times.
Standard calibration approaches sample currents uniformly across a fixed range. On coherent MZI meshes, this produces a heavily biased distribution of realized optical operations, concentrated around a small subset of all possible transformations. A network trained on such data performs well on familiar operations but fails on unfamiliar ones.
The approach
The team led by Jose Roberto Rausell-Campo designed a tandem neural network (TNN) with two components. A forward network learns the mapping from applied currents to the actual optical transformation realized on the chip, effectively a differentiable model of the physical hardware. An inverse network then learns the reverse mapping: given a desired transformation, predict the currents needed to realize it.
The key innovation is in how training data is generated. Rather than sampling currents uniformly, the researchers derived the correct current distribution from random matrix theory using Haar measure principles, the unique uniform distribution over the unitary group. This “architecture-aware sampling” (AAS) method generates training data that covers the full space of possible optical transformations.
A fully physics-agnostic variant, “optimized sampling” (OS), uses differential evolution to search for current settings that produce specific target transformations without any knowledge of the chip’s internal topology, at the cost of significantly longer data acquisition.
The results
On a 3×3 MZI mesh, AAS improved calibration precision from approximately 4.0 bits (uniform baseline) to 6.31 bits when tested on random unitary matrices, a gain of roughly 2.3 bits. OS achieved 5.9 bits. On a 4×4 mesh, AAS reached 5.79 bits and OS 5.58 bits, compared to approximately 4.0 bits for uniform sampling.
The critical finding is that AAS and OS show minimal deviation between performance on training-distribution data and random test data, meaning they achieve true generalization. The uniform baseline collapses to approximately 4 bits on random unitaries while appearing competitive on distribution-matched data.
The framework was also validated on a 2×2 universal gate for coherent detection, achieving near-perfect simultaneous control of amplitude and phase, the first experimental demonstration of a black-box neural network for phase prediction in programmable photonic circuits.
Why it matters
In quantum optical computing, photonic circuits require high-fidelity unitary transformations for operations like Boson sampling and linear optical quantum computing. Fabrication errors cause infidelities that scale with circuit size. A 2-bit precision improvement directly translates to higher gate fidelities.
In classical optical computing, photonic neural networks for tasks like image classification, the framework was tested on realistic workloads including a spiking neural network on a spiral dataset and ResNet-50, Inception-V3, and MobileNet-V3 on CIFAR-10. Inception-V3 with AAS/OS showed approximately 7% degradation from 32-bit digital baseline, while uniform sampling exceeded 40% degradation.
The caveats
The preprint has not yet undergone peer review. The most significant practical limitation is data acquisition time: current instrumentation takes roughly 0.54-0.72 seconds per matrix measurement. For AAS on a 4×4 mesh, this translates to 45-300 hours of data collection; for OS, 162-1,600 hours. The authors note this could be reduced approximately 1,000-fold by operating within thermal time constants or using electro-optic actuators, but these improvements have not yet been demonstrated. The calibration is a one-time cost per device, inference is fast, but slow acquisition limits practical adoption for larger meshes.
Disclosure: Based on arXiv preprint 2601.04122, version 2 (May 2026), which has not undergone peer review.
Sources
Rausell-Campo JR, Melati D, Shastri B, Perez-Lopez D, Capmany J. “Universal Neural Network Based Calibration and Control of Programmable Classical and Quantum Photonic Integrated Processors.” arXiv:2601.04122 (v2, May 2026). DOI: 10.48550/arXiv.2601.04122