Self-Configuring Photonic Networks (SCNs)

• Self-Configuring Photonic Networks are systems that use in situ measurements to autonomously adjust and calibrate optical transformations.
• They integrate programmable interferometric meshes, ASIC controllers, and localized feedback loops to minimize loss and compensate for drift in real time.
• SCNs enable advanced modal recovery, switching, and quantum state analysis, broadening applications across classical and quantum photonics.

Self-Configuring Photonic Networks (SCNs) are photonic systems in which the internal optical transformation is configured in situ from measurements of the physical device, rather than fixed solely by offline design or manual tuning. In the literature, the term spans self-aligning meshes of $2\times 2$ interferometric blocks that automatically align to an unknown optical field by a sequence of simple one-parameter power minimizations, programmable processors whose software layer computes optical paths and tunable states from connectivity intent, and forward-trained free-space or quantum photonic processors that learn their modal basis from measured hardware response (Miller, 2020, Xie et al., 2024, Rocha et al., 23 Jan 2025, Roques-Carmes et al., 2024). The category also has clear boundaries: some works provide enabling switching or control substrates rather than a complete SCN, while adjacent nonlinear frameworks broaden self-configuration toward driven-dissipative adaptation rather than the self-alignment of a programmable linear optical transformation (Nikolova et al., 2015, Yorke, 19 May 2026).

1. Historical emergence and conceptual scope

Early work that is relevant to SCN concentrated on programmable switching fabrics and their electronic control interfaces. An 8×8 microring-based silicon photonic switch demonstrated software controlled switching, real-time firmware controlled switching, fully non-blocking connectivity, path independent insertion loss, and close to 39 dB port isolation, but it did not demonstrate closed-loop autonomous self-configuration; it is best understood as an enabling hardware substrate (Nikolova et al., 2015). A 7×7 silicon-photonic multicore-fiber switch likewise demonstrated reconfigurable SDM switching with 57 Mach-Zehnder interferometric (MZI) structures, lowest total insertion loss of the silicon integrated circuit as low as 4.5 dB, and bit error rate performance below $10^{-9}$, while remaining a programmable hardware block rather than a complete SCN (Ding et al., 2016).

A more direct SCN trajectory appears in programmable interferometric meshes. A self-learning optical neural network chip used 48 thermo-optic phase shifters and 20 MZIs to realize multichannel optical switching, optical MIMO descrambling, and tunable optical filtering by complete self-learning, starting from a black-box device state rather than a calibrated internal model (Zhou et al., 2019). A subsequent self-configuring network of $2\times 2$ blocks showed that a feed-forward mesh can automatically align itself to an unknown coherent multimode field, recover the relative amplitudes and phases of all modal components, and, after calibration, run backward as a field generator (Miller, 2020).

Recent SCN research has broadened the target of self-configuration. The same principles have been extended to partially coherent light, where cascaded self-configuring layers diagonalize an input coherency matrix by sequentially maximizing average output power (Roques-Carmes et al., 2024), and to bipartite quantum states, where a bipartite self-configuring network learns the Schmidt decomposition by maximizing output powers or coincidence counts (Roques-Carmes et al., 2024). In parallel, a self-configuring free-space MPLC showed that high-dimensional linear diffractive processors can be trained directly on hardware by forward-only transmission-matrix measurements (Rocha et al., 23 Jan 2025), and a multidimensional integrated processor used an optical singular-value decomposition engine to sort random speckled inputs across spatial and polarization dimensions and then implement beam shaping, switching, and add-drop functions (Chen et al., 13 Apr 2026).

The conceptual scope of SCN is therefore broader than a single mesh topology. In a canonical SCN, one expects network parameters tuned automatically toward target input-output behavior, usually via local feedback and error signals, often in a programmable interferometric mesh or adaptive optical circuit, with calibration against fabrication errors and drift (Yorke, 19 May 2026). At the same time, the literature now includes software-defined, forward-trained, quantum-variational, and partially coherent extensions of that original paradigm.

2. Architectural substrates

SCNs are realized on several distinct photonic substrates, each supporting a different balance of universality, locality of control, monitoring, and dimensionality.

Substrate	Representative work	Relation to SCN
Feed-forward MZI meshes of $2\times2$ blocks	(Miller, 2020)	Automatic alignment by local detector nulling; analysis and generation of coherent multimode fields
Hexagonal programmable integrated photonic processor	(Xie et al., 2024)	Software-defined interconnects, switching, and multicast; hybrid SDN-like rather than fully autonomous closed-loop SCN
Free-space 4-plane MPLC with MEMS phase-only light modulator	(Rocha et al., 23 Jan 2025)	Self-configuring linear optical processor trained in situ by forward-only transmission-matrix measurements
Adaptive beam coupler with integrated ASIC controller	(Sacchi et al., 16 Jan 2025)	Multiple parallel local feedback loops for real-time self-configuration and drift compensation
Multidimensional SVD engine with input/output antennas and $8\times8$ mesh	(Chen et al., 13 Apr 2026)	Self-programmed in situ processing across spatial and polarization dimensions
Driven-dissipative nonlinear photonic network	(Yorke, 19 May 2026)	Nonlinear, dynamical, partially self-configuring framework adjacent to canonical SCN

The canonical integrated substrate remains the programmable interferometric mesh. In the coherent-field literature, binary-tree and diagonal-line layers built from MZIs are preferred because they support progressive local nulling: once a block is configured in an earlier column, later adjustments do not disturb that null under ideal feed-forward assumptions (Miller, 2020). In software-defined optical networking, the substrate may instead be a hexagonal arrangement of 72 Programmable Unit Cells (PUCs) and 28 optical ports connected to photodetectors, where each PUC is a $2\times 2$ MZI with bar, cross, and tunable coupler states (Xie et al., 2024). In free-space SCN, the substrate can be a 4-plane MPLC built from repeated reflections between a MEMS phase-only light modulator and a mirror, with plane spacing of approximately 6 cm and up to 32,400 optimized parameters (Rocha et al., 23 Jan 2025).

SCN substrates also differ in how directly they expose internal observability. The integrated ASIC-controlled beam coupler pairs each photonic device with monitor photodiodes and local electronic control loops (Sacchi et al., 16 Jan 2025), while a recirculating “bricks” mesh has been proposed as a substrate especially suited to dense monitoring, self-calibration, and stabilization because it can support power monitoring “in each location of the circuit” and subsequent feedback control (Gosciniak, 20 Apr 2026). By contrast, some self-configuring processors rely primarily on external instruments—power meters, cameras, polarimeters, or coincidence electronics—rather than embedded monitors (Rocha et al., 23 Jan 2025, Chen et al., 13 Apr 2026).

3. Self-configuration mechanisms and mathematical formulations

The central SCN mechanism is measurement-driven adjustment of a programmable optical transformation. In the coherent multimode case, a self-configuring layer is tuned so that all power ends up at one designated output by a sequence of local one-parameter minimizations of detector power. The underlying MZI model is written as

$$\begin{bmatrix} a_R\ a_B \end{bmatrix} = M(\phi_{\mathrm{Tot}},\Delta\theta,\Delta\phi) \begin{bmatrix} a_T\ a_L \end{bmatrix},$$

with $\Delta\phi$ controlling relative input phase alignment and $\Delta\theta$ controlling the effective split ratio; after self-configuration, the final mesh settings encode the full complex input vector (Miller, 2020).

For partially coherent light, the objective changes from coherent nulling to statistical diagonalization. The coherency matrix obeys

$$\rho_\text{out} = U_\text{PCLA}\,\rho_\text{in}\,U_\text{PCLA}^\dagger,$$

and the $10^{-9}$0-th self-configuring layer maximizes the average output power

$10^{-9}$1

thereby learning the eigenbasis in which the outputs are mutually incoherent and the output powers are the eigenvalues $10^{-9}$2 (Roques-Carmes et al., 2024). In the bipartite quantum setting, the state matrix $10^{-9}$3 is decomposed by a variational singular-value objective, with the coincidence-based criterion

$10^{-9}$4

so that the configured subnetworks become the Schmidt bases of the unknown pure state (Roques-Carmes et al., 2024).

A different mathematical route appears in forward-trained MPLCs. For plane $10^{-9}$5, the current output field is written

$10^{-9}$6

where $10^{-9}$7 is the transmission matrix from the input to plane $10^{-9}$8, $10^{-9}$9 is the diagonal phase modulation at that plane, and $2\times 2$0 is the downstream transmission matrix to the output. The target field at plane $2\times 2$1 is inferred from the measured downstream transmission matrix by

$2\times 2$2

and the phase update is obtained from

$2\times 2$3

which yields a forward-only in situ approximation to wavefront matching without shaped backward propagation through the MPLC (Rocha et al., 23 Jan 2025).

Software-defined SCN variants formulate self-configuration as a routing and resource-allocation problem on a graph abstraction of the photonic mesh. In the 72-PUC hexagonal processor, the controller assigns a weighted cost

$2\times 2$4

to each arc, then applies Dijkstra-based shortest-path routing, conflict-aware switching synthesis, or multicast tree construction with tunable-coupler compensation (Xie et al., 2024). This is not local nulling in the Miller sense, but it is still a self-configuration mechanism in which network intent is translated automatically into photonic states.

4. Monitoring, control electronics, and self-correction

SCNs require a control plane capable of setting working points, compensating drift, and maintaining operation under perturbation. A dedicated example is an 8-channel mixed-signal CMOS ASIC in AMS 0.35 µm CMOS, powered from 3.3 V, with active area about 12 mm² and per-channel power dissipation of about 10 mW. Two such ASICs were used to control a 16-channel silicon-photonic adaptive beam coupler, with each channel closing local dithering-based feedback loops around one photonic device through a TIA, gated integrator, 10-bit ADC at $2\times 2$5, 12-bit DACs, and heater drivers (Sacchi et al., 16 Jan 2025).

In that system, self-configuration is realized by lock-in extraction of local derivatives $2\times 2$6 from monitor-port optical power and integral feedback that drives those derivatives to zero. Starting from random initial heater voltages, the first stage converged in about 1.5 ms and the full 4-stage mesh in about 10 ms. The same controller also compensated static and dynamic wavefront distortions, suppressed turbulence-induced perturbations up to about 300 Hz, and supported 25 Gbit/s NRZ OOK with $2\times 2$7 and $2\times 2$8, as well as improved 50 Gbit/s PAM-4 eye diagrams (Sacchi et al., 16 Jan 2025). This established a concrete SCN control substrate for real-time stabilization and reconfiguration.

Theoretical work on balanced photonic binary tree cascades clarifies why some SCN architectures are much easier to correct than others. For a perturbation vector $2\times 2$9, the output error obeys the second-order expansion

$2\times2$0

and, for a single phase shifter, the paper shows that phase sensitivity is proportional to the optical power through that phase shifter (Pai et al., 2022). The resulting scaling laws are architecture-dependent: configuration time and error sensitivity scale as $2\times2$1 for balanced trees and as $2\times2$2 for unbalanced trees, even though both use $2\times2$3 nodes. For SCN design, that result formalizes the advantage of low-depth architectures for local self-correction (Pai et al., 2022).

A more speculative but SCN-oriented control substrate is the recirculating “bricks” mesh. It is proposed as a shifted rectangular mesh with 2 to 4 MZIs per unit cell, with ports on all four sides and support for internal monitoring through a Wheatstone-bridge-like arrangement that outputs voltage directly. The architecture is presented as compatible with self-calibration, self-configuration of whole layers, and real-time stabilization against process tolerances and thermal drift (Gosciniak, 20 Apr 2026). The paper’s status is architectural rather than experimental, but it addresses a recurrent SCN requirement: dense observability with minimal insertion loss.

5. Applications and demonstrated systems

SCN ideas have been applied to switching, mode unscrambling, filtering, beam shaping, partially coherent sensing, and quantum modal analysis. In multicore-fiber networking, a reconfigurable 7×7 SDM switch integrated input and output multicore-fiber couplers with a silicon-photonic MZI matrix and achieved 4.5 dB lowest insertion loss in bar configuration, 5.5 dB in cross configuration, crosstalk lower than $2\times2$4 and $2\times2$5 depending on configuration, and successful 1 Tb/s/core transmission over 2 km 7-core fiber with $2\times2$6 for all spatial channels (Ding et al., 2016). That work is most accurately categorized as SCN-enabling spatial switching hardware.

In general-purpose programmable integrated photonics, a hexagonal 72-PUC processor demonstrated dynamic optical interconnects, $2\times2$7 circuit switching, and $2\times2$8 multicasting on the same chip. Measured interconnect insertion loss ranged from 7.7 dB to 10.5 dB, average leakage to non-target outputs was $2\times2$9, all 720 permutations of the 6×6 switch were solved successfully, and a 1×26 multicast showed maximum output deviation of 1.31 dB (Xie et al., 2024). This was a software-defined photonic network substrate in which the control plane computed routes, conflict-free switch states, and multicast splitting ratios automatically.

A black-box self-learning ONN chip demonstrated that the same universal MZI mesh can be retrained for multiple signal-processing functions. In optical switching, crosstalk reached below $8\times8$0 at 1550 nm for one routing state and below $8\times8$1 for another; in 4-channel 10 Gbit/s NRZ MIMO descrambling, crosstalk reached lower than $8\times8$2 at 1550 nm; and in tunable filtering, the center wavelength was adjusted from 1537 nm to 1562 nm with FWHM fixed at about 20 nm (Zhou et al., 2019). The significance for SCN lies in the use of output-based cost functions rather than internal calibration.

Free-space self-configuring MPLCs extend SCN to very high-dimensional linear diffractive optics. A 4-plane device mapped a single orthogonal speckle input to $8\times8$3 with output fidelity 0.95 after 40 mask updates, achieved fidelities 0.87, 0.92, and 0.87 for a three-mode speckle-to-HG transformation, and realized a 10-mode Hermite–Gaussian sorter with average crosstalk $8\times8$4 and a 7-mode orthogonal speckle sorter with $8\times8$5 (Rocha et al., 23 Jan 2025). Optimization times ranged from 4 min for single-beam reshaping to 47 min for the 10-mode HG sorter and 64 min for the 7-mode speckle sorter. These demonstrations showed that self-configuration can absorb unknown aberrations and misalignments into the learned optical transformation.

Multidimensional integrated processors have pushed SCN into joint spatial-polarization processing. An optical SVD engine with an 8×8 Clements mesh and multidimensional antennas sorted random speckled inputs, then synthesized target beams including Gaussian, Hermite–Gaussian, Laguerre–Gaussian, and polarization-programmed outputs. Three pairs of orthogonally polarized Gaussian beams were produced with normalized Stokes-parameter deviations within $8\times8$6, while a mode/polarization-domain ROADM achieved crosstalk below $8\times8$7 for single-beam dropping and less than $8\times8$8 for selective dropping of two concurrent beams, with stable optimized crosstalk reached in approximately one minute on average (Chen et al., 13 Apr 2026). High-speed optical switching was also demonstrated with 64-Gbaud NRZ-OOK and 64-Gbaud PAM-4 eye diagrams.

Quantum and statistical extensions of SCN are equally notable. For partially coherent light, self-configuring optics can diagonalize the input coherency matrix and thereby separate mutually incoherent natural modes without full tomography (Roques-Carmes et al., 2024). For bipartite quantum states, a bipartite self-configuring network learns Schmidt modes and values by optimizing powers or coincidence counts (Roques-Carmes et al., 2024). For multimode squeezed light, a variational SCN discovers the top $8\times8$9 supermodes with $2\times 2$0 physical elements and optimization steps rather than the $2\times 2$1 burden of full covariance reconstruction, and a nonuniform frequency-bin implementation recovered the input Bloch–Messiah decomposition with fidelity 99.58% for $2\times 2$2, $2\times 2$3, and 10% loss (Karnieli et al., 20 Sep 2025).

6. Boundaries, limitations, and emerging directions

A persistent issue in the SCN literature is the distinction between genuinely self-configuring systems and enabling programmable hardware. The microring switch fabric and the 7×7 SDM switch are highly relevant because self-configuring network nodes will need low-loss, low-crosstalk, software-controlled switching fabrics, yet those works do not provide autonomous control, embedded feedback optimization, or self-healing logic (Nikolova et al., 2015, Ding et al., 2016). The software-defined 72-PUC processor goes further by automatically compiling user intent into photonic states, but its demonstrations remain primarily open-loop programming with experimental verification, not continuous telemetry-driven closed-loop optimization (Xie et al., 2024).

Measurement overhead and hardware nonidealities remain major constraints. The self-configuring MPLC depends on interferometric complex-field measurements, transmission-matrix acquisition, and a MEMS phase-only modulator with only 4-bit phase precision; the prototype’s efficiency is estimated as

$2\times 2$4

and the current method scales as $2\times 2$5 in the sampled basis size, number of modes, plane count, and optimization cycles (Rocha et al., 23 Jan 2025). Integrated controller approaches mitigate calibration burden, but their present bandwidth is tied to thermal actuators, with a reported control bandwidth around 400 Hz, and their most straightforward objective remains full steering of power to one target output rather than arbitrary multiport matrix synthesis (Sacchi et al., 16 Jan 2025).

The term SCN is also stretched by adjacent dynamical frameworks. The Reconfigurable Nonlinear Photonic Decision Network is explicitly described as a nonlinear, dynamical, partially self-configuring photonic framework in which weights evolve under

$2\times 2$6

with hysteresis, saturation, bistability, and driven-dissipative state evolution (Yorke, 19 May 2026). It shares SCN-relevant properties—internal parameters reconfigure during operation, adaptation occurs in situ, and local feedback modulates physical configuration—but it lacks the canonical SCN ingredients of an explicit programmable mesh, autonomous calibration to a target optical transformation, explicit trainable interconnects, and physical hardware demonstration (Yorke, 19 May 2026). This has made it a useful conceptual extension rather than a direct SCN blueprint.

Current research directions indicate two converging trajectories. One trajectory pursues faster and more scalable self-configuration of linear optical processors: projected 10 kHz MPLC hardware reduces optimization times to 9 s for single-beam reshaping and 88 s for a 10-mode HG sorter (Rocha et al., 23 Jan 2025), while frequency-domain quantum SCNs propose full-network hardware compressed to $2\times 2$7 or even $2\times 2$8 cavities through inverse-designed surrogate networks (Karnieli et al., 20 Sep 2025). The other trajectory seeks richer physical self-configuration, including dense monitoring, self-calibrating recirculating meshes, and nonlinear driven-dissipative adaptation (Gosciniak, 20 Apr 2026, Yorke, 19 May 2026). Taken together, these developments indicate that SCN is evolving from a narrow label for self-aligning interferometric meshes into a broader technical program centered on in situ photonic learning, calibration, and modal discovery across coherent, partially coherent, classical, and quantum regimes.