Slot Sb2Se3-Tunable Directional Couplers: a Compact, Reconfigurable Building-Block for Neuromorphic Photonics

Neuromorphic photonic processors need synaptic elements that are ultra-compact, low-loss, broadband, and non-volatile. We present and numerically validate a 2×2 slot directional coupler (DC) whose silicon slot region is symmetrically clad with the phase-change material (PCM) Sb₂Se₃. In our concept, the PCM forms the inner rails of the DC while the outer rails are standard Si waveguides; toggling Sb₂Se₃ between amorphous and crystalline states perturbs the super-mode splitting and flips the bar/cross transfer, enabling non-volatile weight programming. The PCM geometry is inspired our prior work on optimal transitions between different substrate materials in heterogeneous Photonic Integrated Circuits, which we adapted here to minimize scattering at the Si/PCM interfaces. Design proceeded from eigenmode/FDE analysis to 3D FDTD. Super-mode extraction was used to set an initial coupling length, then the full device was optimized in FDTD. The stand-alone coupler achieves an active length as short as 2.3 µm, yielding a footprint suited for high synapse density. Across the 1500–1600 nm band the transmission on the selected output remains ≥ 0.80 with weak wavelength dependence (flat within a few percent in our sweeps), while the non-selected port stays near zero, giving >10 dB simulated extinction without additional phase trimming. The short length directly translates to low propagation loss and relaxed phase-error sensitivity. To assess system-level feasibility, we simulated meshes of up to 15 such DC building blocks and executed basic matrix operations. We emulate programming by assigning the measured complex refractive indices of Sb₂Se₃ in its amorphous and crystalline states; switching the PCM flips each DC between bar-dominant and cross-dominant states. The meshes show numerically stable behavior for representative linear transforms, with performance limited primarily by insertion loss of passive sections and (possibly in future fabricated versions) by mesh calibration and not by bandwidth. Because the device is broadband around 1550 nm and non-volatile, it avoids continuous bias power typical of thermo-optic or carrier-based weights. These characteristics make it a promising reconfigurable building-block for dense neuromorphic PIC inferences.