Optical matrix multiplier

Abstract

The invention discloses an optical matrix multiplier, which comprises a laser light source, a grating coupler, a BTO electro-optical modulation array, an MZI waveguide array, a photoelectric detector array and a control unit, and is characterized in that the laser light source generates coherent light with stable wavelength, and the coherent light enters a BTO waveguide after being coupled by the grating coupler and then enters the MZI waveguide array through multi-stage beam splitting; the excellent electro-optical effect of a BTO material is utilized, phase modulation is achieved under electrode driving so as to load matrix element information, and then linear operation of matrix multiplication is completed through interference superposition of an MZI waveguide array. And the output optical signal is converted into an electric signal by the photoelectric detector array and is processed by the control unit to obtain an operation result. The system is compact in structure, high in modulation linearity, high in response speed, low in power consumption and capable of achieving multi-channel parallel optical calculation. Compared with a traditional silicon-based or lithium niobate modulator, higher integration density and stability are achieved, and the calculation efficiency of matrix multiplication in the fields of artificial intelligence, scientific calculation, signal processing and the like is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of knowledge graph construction technology, and more specifically, to an optical matrix multiplier. Background Technology

[0002] As a core computing unit in fields such as scientific computing, artificial intelligence, signal processing, and quantum simulation, the performance of matrix multipliers directly determines the overall efficiency of large-scale data processing and model inference. From weight updates in deep neural networks, feature extraction in image recognition, and speech signal filtering, to tasks such as weather simulation, gene computing, and financial forecasting, the computational speed, accuracy, and energy consumption of matrix multiplication have always been key bottlenecks restricting the improvement of system performance.

[0003] Traditional electronic computing architectures rely on transistors and logic gate arrays to perform multiply-accumulate operations. As device sizes approach the nanoscale, electron mobility is significantly enhanced by scattering effects, leading to a sharp increase in power density per unit area. Simultaneously, frequent data transfers between memory and processing units cause the "memory wall effect," whose energy consumption exceeds that of arithmetic operations themselves. Although dedicated accelerators such as GPUs and TPUs employ parallel arrays and on-chip cache optimizations in their architecture, their energy efficiency is still limited by electron mobility and interconnect latency. The slowdown of Moore's Law makes the method of improving computing performance through transistor stacking unsustainable.

[0004] Optical matrix multipliers offer a new approach to overcoming this bottleneck. Optical signals propagate at the speed of light, enabling natural parallel superposition and zero-static-energy-consumption interferometric computation. Theoretically, the computational speed is limited only by the photodetector's response speed, far exceeding that of traditional electronic systems. Optical matrix multiplication encodes matrix elements through amplitude and phase loading of the optical field, generating a weighted summation result through interference. High-dimensional matrix operations can be performed within a monolithic integrated waveguide. This scheme not only boasts extremely high bandwidth and throughput but also allows for scaling up computational capabilities without increasing energy consumption because the optical signals do not interfere with each other.

[0005] However, existing optical matrix multiplication schemes still have several limitations. While silicon-based optical modulators have mature fabrication processes and are compatible with CMOS technology, their electro-optic coefficients are low (approximately 10 pm / V), requiring driving voltages typically in the tens of volts, resulting in insufficient modulation efficiency. Lithium niobate modulators, although possessing high electro-optic responses, are brittle, large in size, and have complex fabrication processes, hindering high-density on-chip integration. Furthermore, while some polymer electro-optic materials exhibit high electro-optic coefficients, their stability and temperature resistance are poor, leading to degradation under long-term operation. Existing MZI arrays also face engineering challenges such as phase drift, interference inhomogeneity, coupling loss, and packaging complexity when scaled up on a large scale, making it difficult to guarantee system accuracy and stability.

[0006] Therefore, there is an urgent need to design an optical matrix multiplier to solve these problems. Summary of the Invention

[0007] The purpose of this invention is to provide a method for constructing a knowledge graph for training injury classification and intervention, so as to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] The optical matrix multiplier includes: a laser source, a BTO electro-optic modulator array, a grating coupler, an MZI waveguide array, a photodetector array, and a control unit;

[0010] The laser source is used to generate coherent light with a stable wavelength;

[0011] The BTO electro-optic modulator array is used to perform intensity modulation and phase modulation on the input light to load input vector information and matrix element information;

[0012] The grating coupler is used to couple optical signals into or out of the MZI waveguide array;

[0013] The MZI waveguide array is used to perform linear matrix multiplication through optical interference;

[0014] The photodetector array is used to convert the output optical signal into an electrical signal;

[0015] The control unit is used to drive the BTO modulator and process electrical signals to obtain matrix operation results.

[0016] As a preferred technical solution of the present invention, the BTO electro-optic modulator array includes a substrate, a cladding layer, a BTO thin film layer, electrodes, and a waveguide structure.

[0017] The substrate is sapphire or silicon.

[0018] The cladding is silicon dioxide;

[0019] The BTO thin film layer has a thickness of 0.2–0.5 μm and is prepared by pulsed laser deposition or magnetron sputtering, and is etched to form a ridge or strip waveguide structure.

[0020] The electrodes are strip-shaped gold or aluminum electrodes, distributed on both sides of the waveguide, with an electrode spacing of 1–2 μm and an electrode length of 300–500 μm.

[0021] As a preferred technical solution of the present invention, the MZI waveguide array (4) is composed of multiple dual-port Mach-Zehnder interferometer units, each unit including two 3dB directional couplers and at least three BTO electro-optic modulators;

[0022] The waveguide has a width of 0.5–1 μm, a thickness of 0.2–0.3 μm, and a bending radius of not less than 50 μm;

[0023] The MZI unit has a length of 600–800 μm and a width of 50–150 μm.

[0024] As a preferred technical solution of the present invention, the grating coupler (3) adopts the same substrate material as the MZI waveguide, has a periodic ridge structure, an etching depth of 0.1–0.15 μm, a period of 0.6–0.8 μm, and a duty cycle of 0.5–0.6;

[0025] The grating coupler has a coupling efficiency of not less than 20%, a working bandwidth of not less than 50nm, and a gradient periodic structure at the edge to reduce reflection loss.

[0026] As a preferred technical solution of the present invention, the arrangement of the MZI waveguide array (4) corresponds to the matrix dimension;

[0027] For N×M dimensional matrix operations, the array contains N×(N−1) / 2+M×(M−1) / 2+min(N,M) MZI units, arranged in a triangular or rectangular pattern;

[0028] Adjacent MZI units are interconnected via waveguides, with the interconnected waveguide length error not exceeding ±0.1μm.

[0029] As a preferred technical solution of the present invention, the laser source (1) is a single-frequency laser with a wavelength range of 1550±50nm, a linewidth of less than 100kHz, and an output power stability better than ±0.5dB / hour.

[0030] The laser is followed by a grating coupler, and then the optical signal is fanned out by a multi-stage beam splitter. After the intensity and phase are modulated by a BTO modulator to load the input information, it is coupled with an MZI waveguide array.

[0031] As a preferred technical solution of the present invention, the photodetector array (5) is a germanium-silicon PIN photodiode with a responsivity of not less than 0.8A / W, a dark current of less than 10nA, and a 3dB bandwidth of more than 20GHz.

[0032] As a preferred technical solution of the present invention, the control unit (6) includes a digital-to-analog conversion module, a drive circuit and a digital signal processor;

[0033] The resolution of the digital-to-analog conversion module is not less than 16 bits;

[0034] The output voltage range of the driving circuit is -5V to +5V, with an adjustment accuracy of ±1mV, and it is used to drive the BTO electro-optic modulator.

[0035] As a preferred embodiment of the present invention, the electro-optic coefficient r of the BTO thin film is... 33 Not less than 180 pm / V, Curie temperature above 120°C, and phase modulation stability better than ±0.1 rad in the temperature range of –40°C to 85°C.

[0036] As a preferred technical solution of the present invention, the directional coupler is of type 2×2, with a coupling ratio of 3dB±0.2dB, a working bandwidth of not less than 40nm, and an insertion loss of less than 1dB.

[0037] The coupling spacing is 110nm, the coupling length is 40μm, the phase modulation range of a single MZI unit is 0–2π, the modulation speed is greater than 40GHz, and the power consumption is less than 100fJ / bit.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] 1. BTO materials have an ultra-large electro-optic coefficient (r). 33 With a speed of ≥200pm / V, far exceeding that of traditional silicon-based modulators, it can achieve precise phase modulation of 0-2π with a low drive voltage of -6V to +6V. It has high modulation linearity and significantly reduces the complexity of the drive circuit and improves the accuracy of matrix operations compared with existing high-drive-voltage, low-precision modulation schemes.

[0040] 2. The length of a single BTO MZI unit is only about 500μm, which is much smaller than the size of lithium niobate MZI material. This allows more MZI computing units to be integrated in the same chip area, greatly increasing the matrix operation scale per unit area. At the same time, the smaller unit size effectively reduces the physical volume of the overall system, making it easier to integrate with other photonic devices or electronic modules at high density, and adapting to the needs of miniaturized, portable, and high-performance computing scenarios.

[0041] 3. Relying on the high-speed transmission characteristics of light and the sub-nanosecond response speed of the BTO modulator, the upper limit of matrix operation speed depends only on the light propagation and detector response speed. The operation time of a 128×128 matrix is ​​<10ns, which far exceeds the microsecond or even millisecond delay of traditional electronic calculators, making it suitable for high-performance computing scenarios that are sensitive to speed. Attached Figure Description

[0042] Figure 1 This is a block diagram of the working system of the optical matrix multiplier provided in this invention.

[0043] Figure 2 This is the unit structure of the optical matrix multiplier provided in the example of the present invention.

[0044] Figure 3This is the core part of the 4*3 matrix multiplication of the optical matrix multiplier provided in this invention example.

[0045] Figure 4 This is a diagram showing the effect of the optical matrix multiplier provided in this invention performing two types of classification and recognition.

[0046] Wherein 101: laser source; 102: grating coupler; 103: BTO electro-optic modulation array; 104: optical MZI array; 105: photodetector array; 106: control unit; 201: BTO electro-optic modulator;

[0047] 202: 3dB directional coupler. Detailed Implementation

[0048] The following will refer to the appendices in the embodiments of the present invention. Figure 1 The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the embodiments described below are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention. In order to provide a clearer explanation and description of the technical solutions and implementation methods of the present invention, specific examples of preferred implementations of the technical solutions of the present invention are introduced below.

[0049] Example 1: As Figure 1 As shown, the optical matrix multiplier of this invention includes a laser source 101, a grating coupler 102, a BTO electro-optic modulation array 103, an optical MZI array 104, a photodetector array 105, and a control unit 106. 101: The laser source is used to generate a coherent optical signal with a stable wavelength; the grating coupler 102 is used to couple the optical signal into or out of the waveguide array; the BTO electro-optic modulation array 103 is used to load matrix element information and realize phase modulation of the optical signal; the optical MZI array 104 is used to realize linear matrix multiplication through optical interference; the photodetector array 105 is used to convert the output optical signal into an electrical signal; and the control unit 106 is used to drive the BTO electro-optic modulation array and process the output signal.

[0050] Light emitted from a coherent source enters the BTO waveguide via a grating coupler. After passing through a beam splitter, it is divided into multiple beams with consistent intensity and phase. These beams are then loaded with the input vector and ensured to maintain phase consistency via an MZI and BTO electro-optic modulator. Interference is then achieved via a directional coupler. Based on the strong linear electro-optic effect of the BTO, and with the application of an electric field through the electrode system, the refractive index of the BTO changes with the electric field, thereby modulating the phase of the optical signal and loading matrix information. After modulation, the light passes through the directional coupler again for interference transformation and returns to the BTO waveguide, thus completing the intensity changes at the two output ports. Subsequently, each of the two output ports undergoes another BTO electro-optic modulation to control the output phase of each port. After transmission through multiple units, the final calculation result is output via a grating coupler to the photodetector array, converted into an electrical signal, and processed by the control unit to obtain the matrix multiplication result.

[0051] like Figure 2 The diagram shows a schematic of an MZI waveguide array unit structure according to an embodiment of the present invention. The unit mainly includes a BTO electro-optic modulator 201 and a 3dB directional coupler 202. Electrodes are respectively disposed on the left and right sides of the BTO electro-optic modulator 201. By applying a voltage to the electrodes, a phase modulation effect of 0 to 2π can be achieved, and the geometry of the electro-optic modulator can be extended or adjusted according to application requirements. The 3dB directional coupler 202 is used to realize beam splitting and beam combining interference of optical signals, thereby completing optical interference calculations.

[0052] As shown in the figure, 203 is the cross-sectional structure of the BTO electro-optic modulator, showing the layered stacking relationship of the BTO layer, SiO2 cladding and Si substrate; 204 is a schematic diagram of the light field distribution, used to show the light field intensity distribution when the laser propagates in the MZI waveguide array.

[0053] In an exemplary embodiment, the optical MZI waveguide has a width of 1 μm and a thickness of 0.2 μm; the waveguide spacing of the directional coupler is 0.11 μm, and the coupling length is 40 μm; the grating coupler has an etching depth of 0.11 μm, a grating period of 0.71 μm, and a duty cycle of 0.52. This structure achieves efficient optical interference modulation within a relatively small chip area, and features a compact structure, controllable phase, and high integration.

[0054] like Figure 3 The diagram illustrates the core structure of the matrix multiplier in the Iris dataset, which can perform arbitrary 4*3 matrix multiplication operations on the input vector. Matrix multiplication operations of other dimensions can be extended based on this.

[0055] like Figure 4As shown, in an exemplary instance, the Iris dataset and the MNIST dataset are classified based on this matrix multiplier, with 80% being the training set and 20% being the test set. The Iris dataset is processed using only the optical matrix layer, and the weight matrix is ​​implemented by the matrix multiplier of BTO in this invention, with a matrix size of 4*3.

[0056] For the MNIST dataset, the data is first preprocessed through an embedding layer and then processed through an optical matrix layer, also implemented using a BTO matrix multiplier with a matrix size of 16*16. Simulation tests, using 150 sets of iris flower datasets and 60,000 handwritten digit images as light sources to input information into the optical MZI array, achieved classification accuracies of 97% and 82%, respectively. The simulation results demonstrate that the BTO-based electro-optic matrix multiplier proposed in this invention is entirely feasible.

[0057] This invention provides an optical matrix multiplier that utilizes the excellent electro-optic properties of BTO material to achieve high linearity, small size, high speed, and low power consumption optical matrix multiplication operations. The computation speed depends on the light transmission and detector response speed, far exceeding electronic computing. It also possesses high stability and high integration, increasing the number of integrated computing units by more than 10 times compared to lithium niobate electro-optic modulators of the same scale. It is suitable for large-scale matrix operations and neural network systems composed of matrix multiplication. This invention provides another feasible approach for the field of optical computing.

[0058] Example 2:

[0059] This embodiment provides an 8×8 optical matrix multiplier based on BTO electro-optic modulation and MZI waveguide array, and verifies it through simulation on Fashion-MNIST and a small-scale voice command feature set. The embodiment includes the following structural components: a laser source, a grating coupler, a BTO electro-optic modulator array, an MZI waveguide array, a photodetector array, and a control unit.

[0060] The system operates using a single-frequency laser in the 1550nm band with a linewidth of approximately 50kHz. The laser beam is coupled into the chip via an incident grating coupler and cascaded through three 1×2 beam splitters to obtain eight coherent beams of equal power. The grating period is approximately 0.72µm, the etching depth is approximately 0.12µm, the duty cycle is approximately 0.5–0.55, and the measured (simulated) coupling efficiency at a single port is ≥20%. The chip platform employs a Si / SiO2 / BTO / SiO2 stacked structure, with a BTO film thickness of approximately 0.30µm, crystallized using pulsed laser deposition (PLD) followed by annealing. The waveguide adopts a ridge structure, with a main waveguide width of 0.8–1.0µm and a thickness of 0.2–0.3µm.

[0061] The electrode material is Ti / Au (10nm / 200nm), with an electrode spacing of approximately 1.4–1.6µm and a single-stage phase modulation segment length of approximately 380–450µm. At least three BTO phase modulation regions are set within each MZI unit (one as an in-arm phaser and two as a common bias). The directional coupler adopts a 2×2 3dB structure with a coupling spacing of approximately 110–130nm and a coupling length of 40–45µm; the insertion loss is less than 1dB, and the operating bandwidth is ≥40nm.

[0062] The MZI array in this embodiment employs an 8×8 Clements grid arrangement, enabling linear mapping of any real matrix on the chip plane using a "rectangular grid." The MZI cells within the array are interconnected via waveguides, with the interconnect waveguide length error between adjacent cells controlled within ±0.1µm to reduce the accumulation of random phase bias. To improve amplitude-phase decoupling capability, a short "equalizing phase converter" (implemented by a BTO segment) is introduced between every two grid levels to compensate for residual phase bias from upstream devices. Eight output ports are located at the end of the array, each connected to one of eight germanium-silicon PIN photodiodes (responsivity ≥0.8A / W, dark current <10nA, 3dB bandwidth ≥20GHz). The control unit includes a 16-bit digital-to-analog converter, a ±5V precision driver, and a digital signal processor (DSP / FPGA). Voltage is applied to each BTO segment via a multi-channel DAC output, achieving 0–2π phase adjustment.

[0063] An example of the manufacturing process is as follows: A SiO2 cladding layer is thermally grown on a polished silicon substrate, followed by deposition of a 0.30µm BTO thin film and annealing for crystallization; the waveguide and coupler layout is defined using electron beam lithography and reactive ion etching to form ridge / strip waveguides; a SiO2 cladding layer is then deposited and windows are created; Ti / Au is sputtered and wet-etched to form strip electrodes; finally, the incident grating structure is etched and PI vias are created. The entire wafer undergoes end-face polishing, and a butterfly laser and array PD are mounted. After soldering control cables, the wafer is packaged.

[0064] The calibration method employs a three-step process: scanning, phase-locked loop (PLL), and perturbation. First, a 0–2π scan is performed on the two phasers of each MZI, recording the output light intensity curve to extract the π point and the 3dB point. Then, a low-amplitude sinusoidal perturbation (kHz-level jitter) is superimposed on each phaser, and the first derivative signal is extracted from the corresponding PD channel through PLL amplification, forming an online gradient cue. Finally, the control unit executes a function minimizing the objective function (output error or cross-aperture energy) to iteratively converge to the desired phase setting. This calibration process avoids the slow drift problem caused by thermal drift, with an overall time constant of approximately several hundred milliseconds. After the array stabilizes, low-frequency compensation is used to maintain the operating point.

[0065] In terms of application verification, to avoid duplication with the iris and MNIST scenarios in Example 1, this example selects two sets of data for simulation: Fashion-MNIST and small-scale voice commands (MFCC features of 10 word classes):

[0066] (1) Fashion-MNIST: The 28×28 grayscale image is first dimensionality-reduced to 64-dimensional embedding vectors by 1×1 on the digital side, and then fed into the four-level "block-stack" topology of the 8×8 grid array in this embodiment (64×10 linear mapping is achieved by splicing on-chip 8×8 matrix and off-chip blocks). The output 10-dimensional class score is read by the PD array and fed into the digital Softmax. Simulation is performed using an 80 / 20 training / test partition and a linear classifier target. Under the condition of not making ideal assumptions about chip noise (including −20dB level crosstalk and PD quantization noise), the test set accuracy is about 84% to 86%.

[0067] (2) Voice commands: 20-dimensional MFCCs and their first-order differences (40 dimensions in total) are extracted from a 1-second voice segment. On the digital side, the dimensions are reduced to 16 dimensions using PCA and then fed into an on-chip 8×8×2 level grid to achieve a 16×10 linear mapping, followed by digital Softmax. Simulations were performed using a 10-class subset of Google SpeechCommands (again, 80 / 20 partitioning). Under the same noise model as described above, the test set accuracy was approximately 88%. These two results demonstrate that, under the constraints of "real device error + readout noise," the optical matrix layer possesses usable performance for typical linear classification tasks.

[0068] To enhance system robustness, this embodiment employs a dual-path balanced detection structure at the output end for differential readout of critical channels, suppressing common-mode noise and laser power fluctuations. Phase conservation constraints and sparsity penalties are added to the control firmware to reduce the impact of invalid phase rotations on the dynamic range within the mesh. Furthermore, two levels of "buffered MZI" are added at the array boundaries to absorb residual phase errors in a closed-loop manner at the layout edges, thereby reducing mismatch during cross-block splicing.

[0069] The key difference between this embodiment and Embodiment 1 is that the array size is expanded from the 4×3 / 16×16 example to an 8×8 grid and a 64×10 mapping is achieved through block division; the MZI arrangement adopts rectangular grids (Clements) instead of a simple triangular arrangement; and a balanced detection and phase-locked loop-perturbation calibration strategy is introduced on the readout side, which significantly reduces the impact of thermal drift and power fluctuation.

[0070] The contents not described in detail in this description are existing technologies known to those skilled in the art. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An optical matrix multiplier, characterized by, The application relates to a laser source, a BTO electro-optical modulator array, a grating coupler, an MZI waveguide array, a light detector array and a control unit. The laser source is used for generating coherent light with a stable wavelength. The BTO electro-optical modulator array is used for amplitude modulation and phase modulation of input light to load input vector information and matrix element information. The grating coupler is used for coupling optical signals into or out of the MZI waveguide array. The MZI waveguide array is used for realizing linear operation of matrix multiplication through optical interference. The light detector array is used for converting output optical signals into electrical signals. The control unit is used for driving the BTO modulator and processing the electrical signals to obtain a matrix operation result. The BTO electro-optical modulator array comprises a substrate, a cladding layer, a BTO thin film layer, electrodes and a waveguide structure.

2. The optical matrix multiplier of claim 1, wherein, The substrate is sapphire or silicon material. The cladding layer is silicon dioxide. The BTO thin film layer has a thickness of 0.2-0.5 mu m, is prepared by a pulse laser deposition method or a magnetron sputtering method, and forms a ridge type or strip waveguide structure through etching. The electrodes are strip gold or aluminum electrodes, are distributed on both sides of the waveguide, have a spacing of 1-2 mu m, and have a length of 300-500 mu m. The MZI waveguide array (4) is composed of multiple two-port Mach-Zehnder interferometer units, each unit comprising two 3dB directional couplers and at least three BTO electro-optical modulators.

3. The optical matrix multiplier of claim 1, wherein, The waveguide has a width of 0.5-1 mu m, a thickness of 0.2-0.3 mu m, and a bending radius not less than 50 mu m. The MZI unit has a length of 600-800 mu m and a width of 50-150 mu m. The grating coupler (3) adopts the same substrate material as the MZI waveguide, has a periodic ridge structure, an etching depth of 0.1-0.15 mu m, a period of 0.6-0.8 mu m, and a duty cycle of 0.5-0.

6.

4. The optical matrix multiplier of claim 1, wherein, The coupling efficiency of the grating coupler is not less than 20%, the working bandwidth is not less than 50 nm, and a gradually changing period structure is arranged at the edge to reduce reflection loss. The arrangement mode of the MZI waveguide array (4) corresponds to the matrix dimension.

5. The optical matrix multiplier of claim 1, wherein, For N*M matrix operation, the array contains N*(N-1) / 2+M*(M-1) / 2+min(N,M) MZI units arranged in a triangular or rectangular shape. Adjacent MZI units are interconnected through waveguides, and the length error of the interconnection waveguide is not more than + / -0.1 mu m. The laser source (1) is a single-frequency laser, has a wavelength range of 1550+ / -50 nm, a line width less than 100 kHz, and an output power stability better than + / -0.5 dB / hour.

6. The optical matrix multiplier of claim 1, wherein, The laser is connected with the grating coupler, and the fan-out of optical signals is realized through multiple beam splitters, and then the intensity and phase of the optical signals are modulated through the BTO modulator to load input information, and then the optical signals are coupled with the MZI waveguide array. The light detector array (5) is a germanium-silicon PIN photodiode, has a responsivity not less than 0.8 A / W, a dark current less than 10 nA, and a 3dB bandwidth greater than 20 GHz.

7. The optical matrix multiplier of claim 1, wherein, The control unit (6) comprises a digital-to-analog conversion module, a driving circuit and a digital signal processor.

8. The optical matrix multiplier of claim 1, wherein, The resolution of the digital-to-analog conversion module is not less than 16 bits. ​ The output voltage range of the driving circuit is –5 V to +5 V, the adjustment accuracy is ±1 mV, and the BTO electro-optical modulator is driven.

9. The BTO electro-optic modulator array of claim 2, wherein, The electro-optic coefficient r of the BTO thin film 33 not less than 180 pm / V, Curie temperature higher than 120℃, phase modulation stability better than ±0.1 rad in the temperature range of -40℃ to 85℃.

10. The MZI waveguide array of claim 3, wherein, The directional coupler is a 2×2 type, the coupling ratio is 3 dB±0.2 dB, the working bandwidth is not less than 40 nm, and the insertion loss is less than 1 dB. The coupling distance is 110 nm, the coupling length is 40 mu m, the phase modulation range of a single MZI unit is 0-2pi, the modulation speed is greater than 40 GHz, and the power consumption is less than 100 fJ / bit.

Images