Nonvolatile programmable silicon photonics using an ultralow-loss Sb2Se3 phase change material

Switching of Sb₂Se₃ patch on a Mach-Zehnder interferometer

In our work, we use a standard 220-nm-thick SOI platform where 120-nm-thick rib waveguides are produced by partial etching of the silicon layer. The refractive index values of Sb₂Se₃ in its two states (n_amorph = 3.285 and n_cryst = 4.050, respectively, for amorphous and crystalline states) are close to that of silicon (n_Si = 3.48), resulting in very similar mode profiles and hence low insertion losses between the different types of waveguide configurations (see table S1). Figure 1 (C to E) illustrates the geometry of the silicon photonic platform under study and includes the mode profiles for the different configurations. From mode calculations, we found that an amorphous single pixel in an otherwise crystalline Sb₂Se₃ layer changes the effective refractive index of the waveguide, n_eff, by an amount Δn_eff = −0.072 (21).

By selective switching of the PCM in one arm of an asymmetric MZI, the optical phase shift is translated into a change of the transmission function of the device. The first demonstration is shown in Fig. 2A, in which we achieve optical phase tuning of an MZI with a 125-μm-long patch of Sb₂Se₃. Progressive tuning of the optical phase is achieved by amorphizing a series of spots in a continuous line along the PCM patch. Spectra were taken every 25 pixels, where each pixel is spaced by 1000 nm along the MZI. The figure shows a succession of vertically offset recorded spectra (gray) from the initial state (red) to the fully switched state (black), followed by a reset to the final state (blue). A detailed view shown at the bottom of Fig. 2A shows the initial, set, and reset states, with the insertion loss for each spectrum normalized to the initial spectrum. The insertion loss remains unchanged during this process, highlighting the lack of any losses induced by the amorphization or recrystallization. Changes in the effective refractive index Δn_eff along the patch are expected to result in very small reflective losses of <0.02 dB per interface (see table S1). A reversible 2π phase shift was obtained as a result of 100 separate crystallization/amorphization pulses. Regardless of the small imperfections introduced by the motorized stage (backlash) during recrystallization, the full 2π range was set and reset with a 0.1-nm offset, which results in an accumulated error within 2.5% of the full range. This measurement demonstrated that a 750-nm pixel size results in a resolution of 0.02π shift, allowing very fine phase control.

As the induced optical phase change scales linearly with amorphized distance and given the ultralow loss of the material, very large phase tuning ranges can be readily achieved by using longer PCM structures. Figure 2B shows a 10π phase shift obtained by switching a 250-μm length of PCM patch, with no measurable change to insertion loss or modulation depth. The effect of the PCM on insertion loss and extinction of the MZIs is explored in more detail in fig. S2, which shows that for PCM patch lengths of up to 250 μm, no additional rebalancing of the MZI is required to compensate for losses in the PCM. Apart from using the focused laser spot size for calibration of the switched pixel, the induced optical phase can be fine-tuned by small changes in the laser power. The phase shift was tuned to 0.04π per pixel in the experiment of Fig. 2B by a small increase in optical peak power of the diode write laser. This average phase shift per pixel was calculated as the total phase shift divided by the number of pixels. Total phase shift was obtained from the spectral shift of the fringes in the MZI response.

To complement the broadband measurements, we performed additional narrowband measurements on an MZI with a shorter, 50-μm patch length as shown in Fig. 2C. Starting at the wavelength of 1553 nm, we tuned the MZI from the maximum transmission (set to 0 dB) to the minimum (−17 dB) and subsequently reset part of this device by recrystallization of a 20-μm length of the PCM patch (red curve). The curve shows the individual levels in transmission induced by switching subsequent pixels in steps of 500 nm along the PCM. An asymmetry between the first write and reset scans is observed, which is attributed to a burn-in effect of the initial film. A second amorphization scan (second blue area) follows more closely the reset scan, where we notice a small but noticeable drop in the maximum extinction of the device from −18 to −15 dB, indicative of small changes in the material upon repeated switching of the PCM layer, changing the sensitive balance between the MZI arms (see also fig. S2). Overall, this result highlights the ability to define stable multilevel switching processes for highly dense and accurate optical phase control with very fine quantization.

Next, the endurance of the phase change was probed by repeatedly switching a single pixel of the MZI. We used a pixel at the −13.5-dB point of the MZI response curve on the steepest part of the slope, where the sensitivity to individual pixels is high. By repeatedly cycling between the amorphous and crystalline phases in a single spot, the transmission modulation gives a good indication of the lifetime of the phase change. Figure 2D shows the transmission in each phase for the first 600 switches, with red dots corresponding to the crystalline and blue dots to the amorphous material phase. A long-term drift of the device transmission, due to environmental fluctuations and material changes, was observed, and we used a 50-point moving average to normalize the drift, as shown in the bottom panel of Fig. 2D. The first 350 full set-reset cycles resulted in a constant transmission change, with a small increase seen toward the 100th cycle. This increase was observed in most tested devices and is attributed to the settling of the switching dynamics after the first few burn-in cycles of the PCM, resulting in a slightly increased responsivity. A reduction of the switching contrast below 50% of the initial value is seen after around 500 cycles. Another example of an endurance experiment is shown in fig. S4. We note that the endurance seen in our current work is a factor 10 lower than that observed for Sb₂Se₃ on planar films (21), which is attributed to the somewhat more challenging thermal environment of the PCM on the waveguide and may require further improvements in thermal design.

Digital programming of MMI using precalculated pixel patterns

Having successfully demonstrated that an optical phase shift can be induced by switching thin Sb₂Se₃ patch on top of an MZI, we moved on to demonstrate a new approach for a programmable optical router based on an MMI patch. MMIs are of interest for their small footprint as well as for their substantially reduced sensitivity to the environment when compared to MZIs. The device principle is based on the concept of wavefront shaping, where a distribution of weak perturbation pixels is used to effectively steer the light with very low loss (22,23). The perturbation induced by switching of the PCM layer is sufficient to apply this device concept in a practical scalable device configuration. Furthermore, a high level of control is achieved, allowing us to independently set and reset pixels spaced at 800-nm pitch in complex patterns.

A 1 × 2 MMI device of 33 μm × 6 μm, covered with a 23-nm crystalline Sb₂Se₃ film, was simulated using a two-dimensional (2D) finite-difference time-domain approach, and a pixel perturbation pattern was optimized using an iterative scheme (22). In this scheme, we switched every pixel of the MMI individually and investigated the resulting MMI outputs. Switched pixels were kept if they improved the target response and were reset when they did not result in an improved target response. Through this iterative process, the optimum combination of pixels in the region between the input and output ports was determined, and we therefore maximized the transmission from one of the outputs. Repeating this optimization two times resulted in a pattern with high single-port transmission. Figure 3 (A and B) shows the calculated intensity in the device before and after patterning.

The designed pattern was subsequently written onto an experimental device. Figure 3 (C and D) shows the fabricated MMI with the same dimensions and cladding as the simulated ones before phase change (C) and after the perturbation pattern was written (D). A residual structure can be seen in the MMI in the unperturbed state (Fig. 3C), which is caused by the crystalline domains of Sb₂Se₃ in the crystalline (unswitched) state. In Fig. 3D, a very good agreement is seen in the pattern registration, with a small ~0.5° tilt present in the experimental pattern due to limitations in the alignment of the setup. Each pixel was written sequentially in columns and rows from the input to the output.

Figure 3 (E and F) shows the simulated (dashed curves) and experimental (solid lines) transmission in the top and bottom outputs for both the set (amorphization) and reset (crystallization) steps, where the single-port transmissions were obtained by normalizing to the sum of the two output ports. The absolute MMI single-port transmissions before writing are −30.6 dB (top) and −30.8 dB (bottom) compared to −25.9 dB for a straight waveguide. After writing the full pixel pattern, a 92%:8% splitting ratio was found between the top and bottom outputs, respectively. Running the final pattern from the 2D optimization sweep in a full 3D simulation for the fully patterned MMI provides a 97%:3% splitting ratio, as shown by the dots in Fig. 3E. The 3D simulation results indicate that optimization using a 2D effective-index model provides a good approximation despite the complex 3D geometry of the perturbation positioned on top of the waveguide. The 5% difference between experimental and simulated results is attributed to experimental limitations in pattern registration onto the device, which are as yet not fully understood. The device reset was done pixel-wise in the same way as the write sequence, i.e., from start to end of the pattern; therefore, the curves in Fig. 3 (E and F) are not the inverse of each other. We point out that the reset was not perfect in all areas. The thermal properties at the edges of the MMI are different due to the surrounding insulating ZnS:SiO₂ capping layer, which results in different switching dynamics. Eventually, the material breaks down at the edges of the MMI, which results in nonreversible switching. This can be mitigated by implementing smoother interfaces and also by using a variable laser power depending on the local topology.

Iterative optimization of MMI transmission

Next to patterning the MMI device using a calculated pixel pattern, we tested an alternative scheme of iterative optimization of the actual device. In the iterative approach, we switched individual pixels and tested their effectiveness on the target transmission. For every pixel, a test was performed to see whether switching increases the transmission at the output port. If a pixel contributed positively to the target, its switched state was preserved; otherwise, it was erased. Whereas a single-pass optimization already resulted in a notable power splitting between the two ports, additional passes are shown to further improve the target transmission function.

Figure 4 (A and B) shows results for bottom (blue) and top (red) output ports where only the accepted pixels are shown, and every data point corresponds to a switched or reset pixel in the MMI. Figure 4 (C and D) shows the same data on a logarithmic scale and normalized to the original port intensities of the unperturbed device, T₀. Raw port output signals are subject to intrinsic variations in grating efficiency and collection fibers, therefore normalizing the signal looks at relative changes compared to the unperturbed device. The entire device was optimized in one pass (pass 1), followed by a second pass, which started again at the beginning of the device and which reoptimized the first 20 μm of the MMI (pass 2).

Figure 4C shows an increase of single-port transmission of the bottom output by 2 dB, accompanied by a 6-dB reduction of the top output. The total transmission, defined as the sum of the unnormalized T_top + T_bottom, shows an overall 0.5-dB increase in insertion loss caused by writing the pattern. The changes induced by writing the pixel pattern are reversible, and the original device transmission is retrieved by resetting all pixels in the device as shown in Fig. 4 (B, D, and F). Movies of writing (both passes 1 and 2) and resetting the device are included in the Supplementary Materials. A series of three videos showing the complete iterative writing process and the device reset are presented in the Supplementary Materials. Both precalculated and iterative optimization patterns result in a target performance exceeding 8-dB extinction between the transmission from the two ports. The impact of the patterning on the total device throughput is not strongly affected, with only up to 0.5 dB additional loss for all devices under study. In some cases, an improvement in device transmission was seen, as shown for example in fig. S6. Such a small improvement in overall device throughput can be attributed to the effect of iterative wavefront shaping, which compensates for some of the imperfections of the original device, as seen in earlier studies (22).

Reversibility of write and reset of MMI pixel patterns

To show that the pixel pattern writing approach is both reconfigurable and free-form, the same pattern was written into an MMI multiple times, with a full reset between each pattern for both the simulated pattern and the mirror image of the pattern in the vertical direction, which, due to symmetry, guides light toward the bottom output. Figure 5A shows the MMI before (1) and after (2) the first write cycle. To improve the visibility of any changes with respect to the original MMI patterns, images of all set and reset states were processed by differencing with the original MMI, as shown in Fig. 5B. A good repeatability was shown between each pattern, with the reset state showing no memory of the patterns written. After several switching cycles, incremental damage to the Sb₂Se₃ film occurred at the edges of the MMI. The damage appears from the corners of the device, where the lower local thermal conductivity of the surrounding ZnS:SiO₂ capping layer and the abrupt interfaces result in a higher maximum temperature, making the Sb₂Se₃ film more prone to delamination or void formation. These effects can be compensated by a variety of methods such as reducing the pulse power for these areas or by reducing the actively written area of the MMI. Figure 5B shows the transmission of both outputs for all the perturbations when writing and resetting the patterns shown in Fig. 5A. A repeatable transmission change was seen for both orientations, which was not strongly affected by the increasing presence of the damaged areas (see Fig. 5A). The splitting ratio was higher for the inverse pattern, suggesting either a small misalignment between the pixel pattern and the MMI or a nonsymmetrical design due to fabrication tolerances in the Sb₂Se₃ film.