A photonic accelerator system

Abstract

This document describes photonic accelerator system comprising a pixelated programmable photonic slab having a plurality of addressable pixels, each pixel possessing a trained programmable optical property ω that controls propagation of optical signals within the pixel. The pixelated programmable photonic slab is configured to apply a target optical computation function to a plurality of received optical signals, whereby the computation function is based on the collective interaction of the trained programmable optical properties ω of the pixels.

Description

A PHOTONIC ACCELERATOR SYSTEMCROSS REFERENCE TO RELATED APPLICATION

[0001] This application claims the benefit of priority to Singapore patent application no. 10202401862R which was filed on 24 June 2024, the contents of which are hereby incorporated by reference in its entirety for all purposes.TECHNICAL FIELD

[0002] This application relates to a photonic accelerator system comprising a pixelated programmable photonic slab having a plurality of addressable pixels, each pixel possessing a trained programmable optical property ID that controls propagation of optical signals within the pixel. The pixelated programmable photonic slab is configured to apply a target optical computation function to a plurality of received optical signals, whereby the computation function is based on the collective interaction of the trained programmable optical properties of the pixels. The processed optical signals resulting from the computation function are then provided to a plurality of optical photodetectors, each configured to convert a corresponding received processed optical signal into an electrical output signal. A computing module is then configured to receive the electrical output signals from the optical photodetectors and to generate a computational result based on the received signals.BACKGROUND

[0003] The computational demand for artificial intelligence (Al) is increasing at an unprecedented rate, with computing power required by Al systems doubling approximately every 3.5 months. This rapid growth far outpaces the capabilities of modem electronic processors Photonic computing, implemented through hardware known as photonic accelerators, is widely regarded as a promising technology to complement electronic systems and support the exponential growth in Al computing requirements. This promise arises from two key advantages offered by photonic accelerators.

[0004] First, it was found that photonic accelerators are able to deliver exceptionally high compute throughput due to the approximately 1,000-fold greater optical bandwidth compared to electrical bandwidth. Second, they offer superior energy efficiency by eliminating the metal interconnect bottlenecks that are inherent in electronic systems. As a result, anticipated performance targets for photonic computing include compute throughputs on the order ofpetaMAC / s and energy efficiencies reaching femtojoules per MAC, where MAC denotes the critical multiply-accumulate operation fundamental to Al workloads.

[0005] To date, several proof-of-concept approaches have been demonstrated by those skilled in the art in the pursuit of practical photonic computing solutions. These include coherent detection techniques that employ Mach-Zehnder interferometer meshes, broadcast- and-weight architectures based on micro-ring resonator arrays, photonic in-memory computing using photonic crossbar arrays, and time-and-wavelength interleaving enabled by microcombs, among others. However, despite the efforts of those skilled in the art, commercial success has been limited primarily due to fundamental challenges in scaling these technologies to achieve the throughput and energy efficiency required for widespread commercialization.

[0006] Specifically, the main limiting factor to wider adoption is the scalability of photonic accelerators. Current photonic accelerators are constrained to relatively small scales, which directly constrains their achievable computational throughput and energy efficiency. To attain high throughput, the input size, output size, and data rate of the photonic accelerator has to be scaled proportionally, while optimal energy efficiency requires enabling a single photon to participate in as many computational nodes as possible. However, most existing proof-of- concept systems operate at scales smaller than 10x10, delivering only about 1 teraMAC / s throughput and achieving energy efficiency of approximately 1 picojoule per MAC, both metrics falling short by a factor of 1,000 relative to expected performance targets. This limited scalability primarily results from significant cumulative errors and optical crosstalk that arise when large numbers of discrete photonic components are controlled within photonic integrated circuits.

[0007] The scalability challenge faced by conventional photonic accelerators arises from their reliance on the precise control of numerous interconnected discrete components. As the number of components increases, the complexity of control grows significantly, leading to increased optical crosstalk and cumulative error accumulation. These factors degrade computational accuracy and present a fundamental barrier to scaling such architectures to meet the high throughput and energy efficiency demands of modern Al workloads.

[0008] As a result, those skilled in the art are constantly looking for new architectures and design methodologies that can overcome these scalability limitations while maintaining high computational accuracy and energy efficiency. In particular, there is significant interest in developing photonic accelerator systems that simplify control complexity, reduce crosstalk and error accumulation, while being able to be integrated into large scale solutions to meet the growing computational demands of Al systems and other such similar data-intensive applications.SUMMARY

[0009] In one aspect, the present application discloses a photonic accelerator system comprising a pixelated programmable photonic slab having optical input terminals for receiving a corresponding optical signal. The disclosed pixelated programmable photonic slab comprise a plurality of addressable pixels, each pixel having a trained programmable optical property m that controls propagation of optical signals within the pixel, wherein the trained programmable optical properties m of all the pixels collectively apply a target optical computation function ffOrward.l(Jthe optical signals propagating through the photonic slab to produce corresponding processed optical output signals. The photonic accelerator system further comprises a plurality of optical photodetectors with each optical photodetector receiving a processed optical output signal from a corresponding output optical terminal of the photonic slab and generating a corresponding electrical output signal; and a computing module configured to receive the electrical output signals from the plurality of optical photodetectors and generate a computational result based on the received electrical output signals.

[0010] In embodiments of the one aspect, the programmable optical property m associated with the target optical computation function ffOrward is trained using an in-situ training process comprising the steps of applying a training set of input optical signals to the pixelated programmable photonic slab; measuring corresponding output signals y from the pixelatedprogrammable photonic slab; computing an error vector based on a derivative of an outputloss, wherein the output loss is computed based on a deviation between the output signals andrespective target output signals; determining a loss gradient U( ) with respect to the programmable optical property m of each of the addressable pixels using a differentiable model updating the programmable optical property m of each of theaddressable pixels based on the loss gradient — to minimize the output loss; and iteratively repeating the steps of applying the set of input optical signals through the updating of the programmable optical property of each of the addressable pixels until the computed loss value is less than or equal to a predetermined threshold.

[0011] In embodiments of the one aspect, the differentiable model comprises an analytical model obtained by solving Maxwell’s equations to derive a deterministic relationship between the output signals, the programmable optical properties of all the addressable pixels, and the modulated optical signals received by the pixelated programmable photonic slab.

[0012] In embodiments of the one aspect, the differentiable model comprises an empirical model obtained by applying randomly generated input optical signals and programmable optical properties of the addressable pixels to the pixelated programmable photonic slab; measuring corresponding output signals from the pixelated programmable photonic slab; and training a deep neural network using the randomly generated input optical signals, the randomly generated programmable optical properties and the measured corresponding output signals as target outputs, wherein the trained deep neural network model learns a differentiable functional relationship between the input optical signals, the programmable optical properties of the addressable pixels and the output optical signals.

[0013] In embodiments of the one aspect, the system further comprises a plurality of optical modulators, each optical modulator configured to receive an input optical signal and generate a corresponding modulated optical signal, wherein each modulated optical signal is provided to a corresponding optical input terminal of the pixelated programmable photonic slab such that the optical signals propagating through the photonic slab comprise the modulated optical signals.

[0014] In embodiments of the one aspect, the system further comprises a first set of demultiplexing arrays, each demultiplexing array configured to receive an input optical signal and demultiplex the input optical signal into a plurality of discrete wavelength components. The system also includes a plurality of modulator arrays, each modulator array optically coupled to a respective demultiplexing array and configured to modulate each of the discrete wavelength components provided by the respective demultiplexing array, and a plurality ofmultiplexing arrays, each multiplexing array optically coupled to a respective modulator array and configured to recombine the modulated discrete wavelength components to generate a corresponding modulated optical signal. In embodiments, each modulated optical signal is provided to a corresponding optical input terminal of the pixelated programmable photonic slab such that the optical signals propagating through the photonic slab comprise the modulated optical signals. In embodiments, the system further comprises a second set of demultiplexing arrays arranged between the output optical terminals of the pixelated programmable photonic slab and the plurality of optical photodetectors with each demultiplexing array configured to receive the processed optical output signal from a corresponding output optical terminal of the photonic slab and demultiplex the processed optical output signal into a plurality of discrete processed wavelength components. In this embodiment, the plurality of optical photodetectors are configured to receive the discrete processed wavelength components in place of the processed optical output signals such that each optical photodetector generates a corresponding electrical output signal based on a respective discrete processed wavelength component.

[0015] In another aspect, the present application discloses a method for performing optical computation using a pixelated programmable photonic slab comprising the steps of receiving, at each optical input terminal of the pixelated programmable photonic slab, a corresponding optical signal, the pixelated programmable photonic slab comprising a plurality of addressable pixels, each pixel having a trained programmable optical property m that controls propagation of optical signals within the pixel, wherein the trained programmable optical properties m of all the pixels collectively apply a target optical computation function fOrWar<i to the optical signals propagating through the photonic slab to produce corresponding processed optical output signals. The disclosed method further comprises the steps of receiving, using a plurality of optical photodetectors, the processed optical output signals from corresponding output optical terminals of the pixelated programmable photonic slab; generating, using each of the optical photodetectors, a corresponding electrical output signal based on a respective processed optical output signal; and receiving, using a computing module, the electrical output signals from the plurality of optical photodetectors and generating a computation result based on the received electrical output signals.BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Various embodiments of the present disclosure are described below with reference to the following drawings:Figure 1 illustrates a top view of an embodiment of a photonic accelerator system in accordance with embodiments of the present disclosure;Figure 2 illustrates a top view of another embodiment of a photonic accelerator system in accordance with embodiments of the present disclosure;Figure 3 illustrates a front view or cross-sectional view of a pixelated programmable photonic slab in accordance with embodiments of the present disclosure;Figure 4 illustrates a perspective view of the pixelated programmable photonic slab in accordance with embodiments of the present disclosure;Figure 5 illustrates a top view of the pixels of the pixelated programmable photonic slab in accordance with embodiments of the present disclosure;Figure 6 illustrates a perspective view of the pixels of the pixelated programmable photonic slab in accordance with embodiments of the present disclosure;Figure 7a illustrates a top view showing the propagation of optical signals through an unprogrammed pixelated programmable photonic slab;Figure 7b illustrates a top view showing the propagation of optical signals through a pixelated programmable photonic slab whose pixels’ optical properties have been programmed in accordance with embodiments of the present disclosure;Figure 8 illustrates an optical programming setup for programming the optical properties of the pixels of the pixelated programmable photonic slab in accordance with embodiments of the present disclosure;Figure 9 illustrates a flowchart that sets out the process or method for obtaining optimal optical properties of the pixels of the pixelated programmable photonic slab in accordance with embodiments of the disclosure; andFigure 10 illustrates a flowchart that sets out the process or method for implementing a photonic accelerator system in accordance with embodiments of the disclosure.DETAILED DESCRIPTION

[0017] The following detailed description is made with reference to the accompanying drawings, showing details and embodiments of the present disclosure for the purposes of illustration. Features that are described in the context of an embodiment may correspondingly be applicable to the same or similar features in the other embodiments, even if not explicitlydescribed in these other embodiments. Additions and / or combinations and / or alternatives as described for a feature in the context of an embodiment may correspondingly be applicable to the same or similar feature in the other embodiments.

[0018] In the context of various embodiments, the articles “a”, “an” and “the” as used with regard to a feature or element include a reference to one or more of the features or elements. The term “and / or” includes any and all combinations of one or more of the associated listed items.

[0019] In the context of various embodiments, the term “about” or “approximately” as applied to a numeric value encompasses the exact value and a reasonable variance as generally understood in the relevant technical field, e.g., within 10% of the specified value.

[0020] As used herein, “comprising” means including, but not limited to, whatever follows the word “comprising”. Thus, use of the term “comprising” indicates that the listed elements are required or mandatory, but that other elements are optional and may or may not be present.

[0021] As used herein, “consisting of’ means including, and limited to, whatever follows the phrase “consisting of’. Thus, use of the phrase “consisting of’ indicates that the listed elements are required or mandatory, and that no other elements may be present.

[0022] It should be noted that although the terms first, second and third are used herein to describe various elements, these elements should not be limited by these terms as these terms are meant to only distinguish one element from another element. Thus, the first element described herein could be termed as a second element without departing from this disclosure.

[0023] In the context of various embodiments, the term “coat” means to enclose something completely to form a barrier around it. Thus, the use of the term “coat” indicates that something is on all sides of another thing.

[0024] In the context of various embodiments, the term “disposed on" relates to the placement or deposition of one material or layer onto the surface of another and may involve one or more types of deposition techniques.

[0025] In the context of various embodiments, the term “around” or “adjacent” means to be in the proximity or location of something and does not necessarily mean that two objects have to be in contact.

[0026] In the context of various embodiments, the directional terms mentioned herein, such as “above” and “below” or “upper” and “lower” refer to directions as described with reference to the drawings. Therefore, the directional terms are only used for illustration and are not meant to limit the present disclosure.

[0027] As used herein, a “layer” refers to a material portion including a region having a particular thickness. The layer may extend over the entirety of the structure or may cover only part of the structure as defined in the description. For example, a layer may be located between two horizontal planes; may be located between, or at, a top surface and a bottom surface of the structure. The layer may also extend horizontally, vertically, and / or along the surface of the structure.

[0028] Additionally, for the sake of brevity, extensive explanations of conventional techniques of fabricating semiconductor devices and integrated circuits are not described in detail herein. The tasks and processes described herein may also be integrated into a more comprehensive procedure with extra steps of features that are not elaborated upon in this document. Specifically, certain processes of fabricating semiconductor devices are well known to one skilled in the art hence, such processes will be omitted entirely.

[0029] Tn embodiments of the disclosure, a photonic slab is treated and programmed as a photonic accelerator, with its computing functionalities being enabled through pixel-level programming and in-situ training. To realize this concept, a complete system of the photonic accelerator system architecture is illustrated in Figure 1, which generally comprises four main functional modules: light routing / distribution modules, data loading modules, data processing modules, and data readout modules.

[0030] It is noteworthy that conventional photonic accelerators are typically constructed by interconnecting numerous discrete photonic components, each requiring precise control of itsindividual functionality. This complex control architecture introduces significant optical crosstalk and cumulative errors, which degrade computational accuracy and ultimately hinder scalability.

[0031] In contrast, embodiments of the present disclosure implement a photonic accelerator using a monolithic photonic slab featuring a large number of input and output ports to enable massively parallel interconnections. The photonic slab incorporates inherent pixel-level programmability to emulate synaptic plasticity, and in-situ training techniques are employed to support robust and adaptive learning. This integrated approach overcomes the scalability limitations associated with managing large assemblies of discrete photonic components and has the potential to fully realize the computational advantages of photonics, thereby allowing photonic accelerators designed in accordance with embodiments of the disclosure to achieve peta-scale throughput and femtojoule-per-operation energy efficiency.

[0032] Figure 1 illustrates a top view of a schematic of photonic accelerator system 100 in accordance with embodiments of the present disclosure. As shown, photonic accelerator system 100 comprises frequency comb source 102, modulator array 104, pixelated programmable photonic slab 106, and photodetector array 108. In embodiments of the disclosure, the photonic accelerator system 100 may be designed to perform high-throughput optical computations, such as matrix-vector multiplication, by leveraging wavelength-division multiplexing, pixellevel optical programmability, and in-situ training of pixelated programmable photonic slab 106.

[0033] In embodiments of the disclosure, frequency comb source 102 may be configured to generate a plurality of coherent optical signals at distinct wavelengths. These optical signals may then be divided into a plurality of channels, e g., N channels, using cascaded photonic splitters 103 and subsequently distributed along a set of input optical paths to be provided to modulator array 104.

[0034] Each modulator in modulator array 104 is configured to modulate an incoming optical signal with data, thereby producing a plurality of modulated optical signals. These modulated optical signals are then provided to corresponding input terminals 105 of pixelated programmable photonic slab 106.

[0035] In embodiments of the disclosure, pixelated programmable photonic slab 106 comprises a pixelated, programmable optical medium that includes a plurality of addressable pixels, each having a programmable optical property m. The modulated optical signals enter the slab via input terminals 105 and propagate through the slab. During propagation, the optical signals interact with the trained pixelated structure, which collectively applies a target optical computation function, such as f forwardl(Jthe propagating optical signals within the pixelated programmable photonic slab 106. The optical signals processed by the target optical computation function ffOrward are then emitted from the output terminals 107 of the slab. The detailed workings of the pixelated programmable photonic slab 106 will be described in greater detail in the later sections with reference to Figures 5 - 7.

[0036] The processed optical output signals are then received by photodetector array 108, where each photodetector of photodetector array 108 is optically coupled to a corresponding output terminal 107 of the pixelated programmable photonic slab 106. Each photodetector then converts the received optical signal into a corresponding electrical output signal. These electrical signals may then be transmitted to a computing module (not shown) which subsequently performs post-processing or downstream computation based on the output of the photonic accelerator.

[0037] Figure 2 illustrates a top view of a schematic of another embodiment of a photonic accelerator system 200, which expands upon the architecture shown in Figure 1 by incorporating wavelength-division multiplexing (WDM) components for enhanced data throughput and channel parallelism. The photonic accelerator system 200 includes frequency comb source 102, photonic splitters 103, a set of demultiplexing arrays 202, modulator arrays 204, multiplexing arrays 206, pixelated programmable photonic slab 106, demultiplexing arrays 208, and photodetector arrays 210.

[0038] In this embodiment, frequency comb source 102 generates a set of coherent optical signals at multiple discrete wavelengths. These wavelength-multiplexed signals are distributed by photonic splitters 103 to the set of demultiplexing arrays 202. Each demultiplexer of demultiplexing arrays 202 is configured to receive an input optical signal whereby each received optical signal is then demultiplexed into a plurality of discrete wavelengthcomponents, e g., K discrete wavelength components, before each of these discrete wavelength components are modulated by the respective modulator arrays 204, which apply data-encoding modulation to each individual wavelength channel.

[0039] The modulated discrete wavelength components from each respective modulator array 204 are then recombined by a corresponding multiplexing array 206. Each multiplexing array 206 outputs a modulated WDM signal comprising the recombined modulated wavelength components, which is then provided to a corresponding optical input terminal 105 of the pixelated programmable photonic slab 106. As described above, pixelated programmable photonic slab 106 operates as the core computational engine, where the modulated optical signals propagate through a pixelated medium comprising a plurality of addressable pixels. Each pixel has a trained programmable optical property m , and the collective pixel configuration implements a target optical computation function, ffOrWardonthe modulated inputs. The processed optical signals are then emitted via output terminals 107.

[0040] In this embodiment, the processed optical output signals at the output terminals 107 are received by a set of demultiplexing arrays 208. Each demultiplexer of demultiplexing arrays 208 separates its received processed optical output signal into individual wavelength components, which are then directed to photodetector array 210. Each photodetector in the photodetector array 210 converts a respective demultiplexed output optical signal into a corresponding electrical output signal and these electrical signals may be provided to a downstream computing module for post-processing or used as final output in an Al inference task.

[0041] In an exemplary configuration of embodiments of the disclosure, the number of input ports N and output ports M of pixelated programmable photonic slab 106 may be set to 128 each, respectively. It should be noted that these values are provided solely as an example, and in principle, the values of N and M may be selected based on application requirements and system design constraints.

[0042] In a scenario where wavelength-division multiplexing (WDM) is not applied (i.e., the embodiment illustrated in Figure 1 and with K — 1), and with the data loading rate defined as 50 GHz, the resulting compute throughput may then be calculated as N X M X K X Data Rate — 128 X 128 X 1 X 50 GHz = 0.8 petaMAC / s. Under the assumption that theanalogue-to-digital converter (ADC) and digital-to-analogue converter (DAC) are the dominant sources of power consumption, the energy per MAC operation may then be computed to be approximately 2000 pJ / 128— 16 fJ / MAC.

[0043] In a further scenario where wavelength-division multiplexing is applied with 32 distinct wavelengths (i.e., the embodiment illustrated in Figure 2 and with K — 32), the compute throughput increases proportionally to 128 X 128 X 32 X 50 GHz= 26 petaMAC / s, while the energy consumption remains at 16 fJ / MAC, assuming similar power sources are the dominant power sources. This level of performance significantly exceeds that of conventional electronic accelerators, such as NVIDIA’s latest H100 Tensor Core GPU, which delivers approximately 2 petaMAC / s at an energy cost of 400 fJ / MAC.

[0044] A cross-sectional view of pixelated programmable photonic slab 106 that functions as the computational engine of the photonic accelerator system in accordance with embodiments of the present disclosure is illustrated in Figure 3. In embodiments of the disclosure, the pixelated programmable photonic slab 106 comprises phase-change material layer 304 disposed on substrate 302, with an oxide layer (not shown) coating phase-change material layer 304. Phase-change transitions of phase-change material layer 304 may be induced by exposure to external stimuli such as optical pulses or thermal energy. In some embodiments, these phase transitions are reversible, enabling dynamic and reprogrammable optical functionality within the pixelated programmable photonic slab 106.

[0045] In embodiments of the disclosure, phase-change material layer 304 may comprise a low-loss phase-change material layer, such as, but not limited to antimony selenide (SbzSej). The Sb2Ses thin film layer may have a thickness on the order of tens of nanometres and may be deposited on a 220 nm thick silicon substrate layer. In this exemplary configuration, the majority of the optical signals propagating through the pixelated programmable photonic slab 106 will be confined to silicon substrate layer 302 while remaining evanescently coupled to the Sb2Se3 film 304. Such a configuration combines the low optical loss characteristic of silicon with the non-volatile, reprogrammable optical properties of Sb2Se3. One skilled in the art will recognize that while specific substrate configurations are described herein, it will be appreciated that other substrate materials and layer combinations may also be employed without departing from the scope of the present disclosure.

[0046] Figure 4 illustrates a perspective view of pixelated programmable photonic slab 106 in accordance with embodiments of the present disclosure. It can be seen that input terminals105 are formed on one side of the photonic slab 106 to receive optical signals, which then propagate through substrate 302 of photonic slab 106 while remaining evanescently coupled to the overlying phase-change material layer 304. Output terminals 107 are formed on the opposing side of the photonic slab 106 to emit processed optical signals after the computation function is applied to the optical signals propagating within the photonic slab 106.

[0047] Figure 5 illustrates a top-down schematic view of pixelated programmable photonic slab 106 while Figure 6 illustrates a perspective view of pixelated programmable photonic slab106 in accordance with embodiments of the present disclosure. As shown, pixelated programmable photonic slab 106 is subdivided into a plurality of discrete, addressable pixels 502 arranged in a two-dimensional array. Each pixel 502 corresponds to a localized region within a vertically stacked structure that includes phase-change material layer 304, e.g., a Sb2Se3 layer, which is disposed on substrate layer 302, e g., silicon substrate layer, and coated with an oxide layer. These pixels collectively define the programmable optical domain of the photonic slab 106, wherein optical signals propagating laterally through substrate layer 302 of photonic slab 106, along the x-axis of photonic slab 106, interact with the pixelated regions 502 via evanescent coupling.

[0048] Each addressable pixel 502 is characterized by a trained programmable optical property m, which in some embodiments comprises a programmable refractive index value. In embodiments of the disclosure, the programmable optical property m of each pixel may be locally modified by inducing phase transitions within the phase-change material layer using optical, thermal, or electrical stimuli. By configuring the profile of the programmable optical property m of each pixel across the pixel array, pixelated programmable photonic slab 106 may be programmed or configured to implement a desired optical computation function, such as a matrix-vector multiplication. One skilled in the art will recognize that pixelated programmable photonic slab 106 may comprise any number, size, shape or arrangement of pixels and that the illustration shown in Figures 5 and 6 are provided for exemplary purposes only and are not intended to be limiting.

[0049] Figure 7a illustrates the propagation of optical signals through pixelated programmable photonic slab 702 which is in an unprogrammed state, in accordance with embodiments of the present disclosure. Photonic slab 702 comprises a two-dimensional grid of addressable pixels 502, similar to the configuration shown in Figure 5. In this example, photonic slab 702 has not undergone pixel -specific training or programming, and as such the programmable optical property a> of each pixel remains in a default uniform optical state.

[0050] As illustrated, when a plurality of optical signals 704a, 704b, and 704c, enter the photonic slab 702, these optical signals will propagate laterally along the X-axis of photonic slab 702. Due to the uniform distribution of the programmable optical properties of the pixels across the pixel array, the optical signals pass through the photonic slab 702 with minimal modulation or computational interaction. This unprogrammed state serves as a baseline configuration, where the photonic slab 702 functions as a passive optical transmission medium rather than performing any optical computation.

[0051] Figure 7b illustrates the propagation of optical signals through pixelated programmable photonic slab 712 that has been programmed in accordance with embodiments of the present disclosure. Photonic slab 712 includes a two-dimensional array of addressable pixels 502, where each pixel exhibits a distinct, trained optical property, such as a spatially varying refractive index, resulting from in-situ training.

[0052] As shown, optical signals 714a, 714b, and 714c enter photonic slab 712 and propagates laterally along the X-axis of photonic slab 712. Unlike the baseline configuration depicted in Figure 7a, the pixel array in this case has been selectively programmed to modulate the optical path of each signal. Due to the programmed variation in optical property of individual pixels, certain regions of photonic slab 712 cause the optical signals to bend, split, or interfere, thereby performing an optical computation. For example, optical signals 714a and 714c are directed toward a common region in the photonic slab 712 where they interact, indicating a form of signal merging or interference-based processing. This demonstrates how a trained photonic slab can apply a desired computation function, such as matrix-vector multiplication, by manipulating light purely through passive, reconfigurable material states.

[0053] One skilled in the art will recognize that the optical signal interactions illustrated in Figure 7b are provided by way of example only and are not intended to be limiting. Other patterns of interaction, routing, or computational behaviour may result from alternative pixel programming configurations within photonic slab 712.

[0054] Figure 8 illustrates an exemplary optical setup that may be used for modifying the optical property a> of individual pixels within a pixelated programmable photonic slab using optical excitation. In this embodiment, the photonic slab comprises a silicon substrate 02 and an overlying phase-change material layer 304 that defines the pixel array. The setup is configured to induce localized phase transitions, in selected pixels of phase-change material layer 304, thereby changing their optical property and enabling pixel-level programmability.

[0055] As illustrated, it can be seen that white light source 810 provides broadband illumination directed downward through beam splitter 804 and beam splitter 806 toward focusing lens 808. Simultaneously, pump light source 812, such as a pulsed or continuous- wave laser, is used to direct a focused beam to beam splitter 806 which in turn directs the focused beam to focusing lens 808 to concentrate the optical energy onto specific regions of phase-change material layer 304. Pump light source 812 is operatively coupled to a function generator 814, which controls key pulse characteristics such as duration, intensity, and repetition rate Camera 802 is positioned above the setup to capture reflected illumination from the surface of phase-change material layer 304, providing real-time optical feedback for monitoring and alignment.

[0056] By adjusting the pump beam’s parameters and focal position, the setup can selectively expose regions of phase-change material layer 304 to controlled optical energy, triggering phase-change transitions at selected pixel locations, such as at location 916. This allows the optical property m of individual pixels to be written, erased, or reprogrammed with precise control. The change resulting from the phase transition alters the programmable optical property ® of the pixel, thereby contributing to the overall computational configuration of the photonic slab.

[0057] In accordance with embodiments of the present disclosure, the programmable optical property m of each pixel within a pixelated programmable photonic slab may be optimized ortrained using a hybrid physical-digital training process. The overall training methodology is illustrated in Figure 9 and enables the photonic slab to implement a desired optical computation function, such as classification or inference, by adjusting the pixel-specific optical properties across the photonic slab.

[0058] The training process 900 begins with an input training dataset x, which may represent any form of optical signals being provided to the photonic slab. As the input training dataset x is applied to the photonic slab at step 902, it undergoes a transformation defined by a forward function fforward- This forward function models the light propagation behaviour within the photonic slab and is governed by the physical configuration of the photonic slab. The forward function ffOrwara will not be analytically expressed in this disclosure as it is inherently defined by the photonic slab’s optical characteristics.

[0059] The parameter space m, which encodes the programmable optical properties of the individual pixels, determines how the input training dataset x interacts with the forward function ffOrward- The resulting optical output y is therefore a function of both the input and the programmed pixel properties, expressed as y = ffOrward(.x’M) The optical output y may be measured by process 900 at step 904.

[0060] A loss function L — \y — y arget I is then computed by process 900 and this loss function L is used to quantify the deviation between the system output y and the expected or target output ytarget- The error vector ^, which is a derivative of the loss function L with respect to the output y is then computed at step 906 by process 900. However, since training involves adjusting the pixels’ properties m, process 900 will establish the sensitivity of the output y with respect to changes in the pixels’ properties u>. This is achieved through a digital differentiable model, which estimates — dc .

[0061] By applying the chain rule of differentiation, the loss gradient with respect to the programmable optical properties may then obtained at step 908 by process 900 as: dL dL dy da> dy da>where this loss gradient indicates how each pixel’s programmable optical property to should be updated in order to minimize the output loss L, thereby aligning the output y more closely with the target output y tar et-

[0062] At step 910, process 900 updates the programmable optical property a> of each addressable pixel based on the computed loss gradient, with the objective of minimizing the output loss L. This updating step may be performed using a gradient descent algorithm or other optimization techniques suitable for physical systems and is omitted for brevity.

[0063] If process 900 determines at step 912 that the output loss L exceeds a predetermined threshold, process 900 then returns to step 902 and all the steps are iteratively repeated, beginning with the application of the input optical signals at step 902 and continuing through to the updating of the pixel properties at step 910, until the computed loss value L falls below a predetermined threshold. When this happens, process 900 proceeds to step 914 where all the optical properties of the pixels in the photonic slab are set based on the current set of optical properties.

[0064] In embodiments of the disclosure, the digital differentiable model used to estimate dy— may be obtained based on an analytical model. In this method, Maxwell’s equations are solved to obtain a deterministic relationship between the input x, programmable optical properties a> of the pixels, and the output y. Since Maxwell’s equations are differentiable with respect to the material properties, this formulation enables direct computation of the output and its gradient. This model is exact under known initial conditions and provides an interpretable, physics-based differentiable model.

[0065] In other embodiments of the disclosure, the digital differentiable model used to y estimate — may be obtained based on an empirical model (black-box approach). In this approach, a dataset is generated by applying randomly sampled inputs x and programmable optical properties to to the photonic slab and measuring the resulting outputs y. A deep neural network is then trained with inputs x and optical properties to as inputs and output y as the output targets Once trained, the neural network approximates the behaviour of the photonic slab and serves as a differentiable function, enabling the calculation of OCi) during training. In lembodiments of the disclosure, the neural network may comprise a multilayer perceptron (MLP), a convolutional neural network (CNN), a transformer-based network, or other networks, where the choice of neural network architecture may be selected based on the nature of the input data, the complexity of the optical transformation, and the desired training efficiency.

[0066] A process for performing optical computation using a pixelated programmable photonic slab in accordance with embodiments of the disclosure is illustrated in Figure 10. Process 1000 begins at step 1002 by receiving, at each optical input terminal of a pixelated programmable photonic slab, a corresponding optical signal. The pixelated programmable photonic slab comprises a plurality of addressable pixels, each pixel having a trained programmable optical property o> that controls propagation of optical signals within the pixel, wherein the trained programmable optical properties of all the pixels collectively apply a target optical computation function ffOrWardtothe optical signals propagating through the photonic slab to produce corresponding processed optical output signals. Process 1000 then proceeds to step 1004 where process 1000 receives, using a plurality of optical photodetectors, the processed optical output signals from corresponding output optical terminals of the pixelated programmable photonic slab. At step 1006, process 1000 generates, using each of the optical photodetectors, a corresponding electrical output signal based on a respective processed optical output signal. At step 1008, process 1000 then receives, using a computing module, the electrical output signals from the plurality of optical photodetectors and generates a computation result based on the received electrical output signals.

[0067] In embodiments of the disclosure, the programmable optical property m associated with the target optical computation function ffOrward is trained using an in-situ training process. Process 1000 performs the training process by applying a training set of input optical signals to the pixelated programmable photonic slab, measuring corresponding output signals y from the pixelated programmable photonic slab and computing an error vector based on a derivative of an output loss, wherein the output loss is computed based on a deviation between the output signals and respective target output signals. Process 1000 then determines a loss gradient with respect to the programmable optical property m of each of the addressablepixels using a differentiable modeland the error vector updates the programmableoptical property m of each of the addressable pixels based on the loss gradientto minimizethe output loss, and iteratively repeats the steps of applying the set of input optical signals through the updating of the programmable optical property of each of the addressable pixels until the computed loss value is less than or equal to a predetermined threshold.

[0068] Numerous other changes, substitutions, variations, and modifications may be ascertained by the skilled in the art and it is intended that the present application encompass all such changes, substitutions, variations, and modifications as falling within the scope of the appended claims.

Claims

CLAIMS1. A photonic accelerator system comprising: a pixelated programmable photonic slab having optical input terminals for receiving a corresponding optical signal, the pixelated programmable photonic slab comprising: a plurality of addressable pixels, each pixel having a trained programmable optical property m that controls propagation of optical signals within the pixel, wherein the trained programmable optical properties m of all the pixels collectively apply a target optical computation function ffOrwardt0the optical signals propagating through the photonic slab to produce corresponding processed optical output signals; a plurality of optical photodetectors, each optical photodetector receiving a processed optical output signal from a corresponding output optical terminal of the photonic slab and generating a corresponding electrical output signal, and a computing module configured to receive the electrical output signals from the plurality of optical photodetectors and generate a computational result based on the received electrical output signals.

2. The photonic accelerator system according to claim 1, wherein the programmable optical property a> associated with the target optical computation function ffOrWard 'strained using an in-situ training process comprising the steps of: applying a training set of input optical signals to the pixelated programmable photonic slab; measuring corresponding output signals y from the pixelated programmable photonic slab; computing an error vector based on a derivative of an output loss, wherein the outputloss is computed based on a deviation between the output signals and respective target output signals;determining a loss gradientwith respect to the programmable optical property m of eachof the addressable pixels using a differentiable modeland the error vectorupdating the programmable optical property m of each of the addressable pixels based on the loss gradientto minimize the output loss; and iteratively repeating the steps of applying the set of input optical signals through the updating of the programmable optical property of each of the addressable pixels until the computed loss value is less than or equal to a predetermined threshold.

3. The photonic accelerator system according to claim 2, wherein the differentiable model comprises an analytical model obtained by solving Maxwell’s equations to derive a deterministic relationship between the output signals, the programmable optical properties of all the addressable pixels, and the modulated optical signals received by the pixelated programmable photonic slab.

4. The photonic accelerator system according to claim 2, wherein the differentiable model comprises an empirical model obtained by: applying randomly generated input optical signals and programmable optical properties of the addressable pixels to the pixelated programmable photonic slab; measuring corresponding output signals from the pixelated programmable photonic slab, and training a deep neural network using the randomly generated input optical signals, the randomly generated programmable optical properties and the measured corresponding output signals as target outputs, wherein the trained deep neural network model learns a differentiable functional relationship between the input optical signals, the programmable optical properties of the addressable pixels and the output optical signals.

5. The photonic accelerator system according to any one of claims 1 to 4, wherein the pixelated programmable photonic slab comprises: an antimony selenide (Sb2Se3) layer disposed on a silicon substrate and coated with an oxide layer, wherein each of the plurality of addressable pixels is defined by a corresponding region of a layer stack comprising the oxide layer, the SbzSes layer, and the silicon substrate.

6. The photonic accelerator system according to any one of claims 1 to 5, wherein the trained programmable optical propertyof each addressable pixel comprises a programmable refractive index value.

7. The photonic accelerator system according to any one of claims 1 to 6 further comprising: a plurality of optical modulators, each optical modulator configured to receive an input optical signal and generate a corresponding modulated optical signal,wherein each modulated optical signal is provided to a corresponding optical input terminal of the pixelated programmable photonic slab such that the optical signals propagating through the photonic slab comprise the modulated optical signals.

8. The photonic accelerator system according to any one of claims 1 to 6 further comprising: a first set of demultiplexing arrays, each demultiplexing array configured to receive an input optical signal and demultiplex the input optical signal into a plurality of discrete wavelength components; a plurality of modulator arrays, each modulator array optically coupled to a respective demultiplexing array and configured to modulate each of the discrete wavelength components provided by the respective demultiplexing array; and a plurality of multiplexing arrays, each multiplexing array optically coupled to a respective modulator array and configured to recombine the modulated discrete wavelength components to generate a corresponding modulated optical signal, wherein each modulated optical signal is provided to a corresponding optical input terminal of the pixelated programmable photonic slab such that the optical signals propagating through the photonic slab comprise the modulated optical signals.

9. The photonic accelerator system according to claim 8 further comprising: a second set of demultiplexing arrays arranged between the output optical terminals of the pixelated programmable photonic slab and the plurality of optical photodetectors, each demultiplexing array configured to receive the processed optical output signal from a corresponding output optical terminal of the photonic slab and demultiplex the processed optical output signal into a plurality of discrete processed wavelength components, wherein the plurality of optical photodetectors receives the discrete processed wavelength components in place of the processed optical output signals such that each optical photodetector generates a corresponding electrical output signal based on a respective discrete processed wavelength component.

10. The photonic accelerator system according to any one of claims 1 to 9, wherein the target optical computation function ffOrWard comprises a matrix-vector multiplication operation.

11. A method for performing optical computation using a pixelated programmable photonic slab comprising: receiving, at each optical input terminal of the pixelated programmable photonic slab, a corresponding optical signal, the pixelated programmable photonic slab comprising: a plurality of addressable pixels, each pixel having a trained programmable optical property m that controls propagation of optical signals within the pixel, wherein the trained programmable optical properties of all the pixels collectively apply a target optical computation function ffOrwardt0the optical signals propagating through the photonic slab to produce corresponding processed optical output signals, receiving, using a plurality of optical photodetectors, the processed optical output signals from corresponding output optical terminals of the pixelated programmable photonic slab; generating, using each of the optical photodetectors, a corresponding electrical output signal based on a respective processed optical output signal; and receiving, using a computing module, the electrical output signals from the plurality of optical photodetectors and generating a computation result based on the received electrical output signals.

12. The method according to claim 11, wherein the programmable optical property a> associated with the target optical computation function ffOrWarstrained using an in-situ training process comprising the steps of: applying a training set of input optical signals to the pixelated programmable photonic slab; measuring corresponding output signals y from the pixelated programmable photonic slab;computing an error vector based on a derivative of an output loss, wherein the outputloss is computed based on a deviation between the output signals and respective target output signals;determining a loss gradientwith respect to the programmable optical property m of each of the addressable pixels using a differentiable modeland the error vectorupdating the programmable optical property m of each of the addressable pixels based on the loss gradientto minimize the output loss; and iteratively repeating the steps of applying the set of input optical signals through the updating of the programmable optical property of each of the addressable pixels until the computed loss value is less than or equal to a predetermined threshold.

13. The method according to claim 12, wherein the differentiable model comprises an analytical model obtained by solving Maxwell’s equations to derive a deterministic relationship between the output signals, the programmable optical properties of all the addressable pixels, and the modulated optical signals received by the pixelated programmable photonic slab.

14. The method according to claim 12, wherein the differentiable model comprises an empirical model obtained by: applying randomly generated input optical signals and programmable optical properties of the addressable pixels to the pixelated programmable photonic slab; measuring corresponding output signals from the pixelated programmable photonic slab; training a deep neural network using the randomly generated input optical signals, the randomly generated programmable optical properties and the measured corresponding output signals as target outputs, wherein the trained deep neural network model learns a differentiable functional relationship between the input optical signals, the programmable optical properties of the addressable pixels and the output optical signals.

15. The method according to any one of claims 11 to 14, wherein the pixelated programmable photonic slab comprises: an antimony selenide (Sb?Se^) layer disposed on a silicon substrate and coated with an oxide layer, wherein each of the plurality of addressable pixels is defined by a corresponding region of a layer stack comprising the oxide layer, the SbzSes layer, and the silicon substrate.

16. The method according to any one of claims 11 to 15, wherein the trained programmable optical property a> of each addressable pixel comprises a programmable refractive index value.

17. The method according to any one of claims 11 to 16, whereby before the step of receiving, at each optical input terminal of a pixelated programmable photonic slab, the corresponding optical signal, the method further comprises the steps of:receiving, using a plurality of optical modulators that are each respectively optically coupled to the optical input terminals of the pixelated programmable photonic slab, a plurality of input optical signals; generating, using each of the plurality of optical modulators, a corresponding modulated optical signal, providing each modulated optical signal to a corresponding optical input terminal of the pixelated programmable photonic slab such that the optical signals propagating through the photonic slab comprise the modulated optical signals.

18. The method according to any one of claims 11 to 16, whereby before the step of receiving, at each optical input terminal of a pixelated programmable photonic slab, the corresponding optical signal, the method further comprises the steps of: receiving, using a first set of demultiplexing arrays, a plurality of input optical signals; demultiplexing, using each demultiplexing array, a respective input optical signal into a plurality of discrete wavelength components; modulating, using a plurality of modulator arrays that are each optically coupled to a respective demultiplexing array, the corresponding discrete wavelength components received from each demultiplexing array to generate modulated discrete wavelength components; recombining, using a plurality of multiplexing arrays that are each optically coupled to a respective modulator array, the modulated discrete wavelength components to generate corresponding modulated optical signals; and providing each modulated optical signal to a corresponding optical input terminal of the pixelated programmable photonic slab such that the optical signals propagating through the photonic slab comprise the modulated optical signals.

19. The method according to claim 18, whereby in place of the steps of receiving the processed optical output signals from the output optical terminals, and generating the corresponding electrical output signal, the method further comprises the steps of: receiving, using a second set of demultiplexing arrays arranged between the output optical terminals of the pixelated programmable photonic slab and the plurality of optical photodetectors, the processed optical output signals from the output optical terminal of the photonic slab;demultiplexing, using each demultiplexing array from the second set of demultiplexing arrays, a corresponding processed optical output signal into a plurality of discrete processed wavelength components; receiving, using the plurality of optical photodetectors, the discrete processed wavelength components; generating, using each of the plurality of optical photodetectors, a corresponding electrical output signal based on a respective discrete processed wavelength component; and providing the electrical output signals to the computing module.

20. The method according to any one of claims 11 to 19, wherein the target optical computation function ffOrWard comprises a matrix-vector multiplication operation.

Images