A Neural Network Computing System Based on Spin-Orbit Coupling of Organic Crystals

Abstract

This invention discloses an optical storage layer neural network computing system based on organic crystal spin-orbit coupling and its acceleration method. First, a pattern to be recognized is generated using a 405nm laser via a spatial light modulator, which vertically excites the center of a non-regular hexagonal sheet-like BPDBNA single crystal. The emitted signal undergoes an optical spin Hall effect induced by TE-TM splitting, resulting in a polarization-locked light field output along the six edges. Real-space images of the six polarization channels are acquired, and the six edges are divided into fan-shaped regions to obtain the output vector, which is then input into a shallow neural network. Only the output layer weights are trained. Experiments show that in ten-symbol recognition, the simplified configuration (output dimension 3) achieves a 30-fold speedup and 99.5% accuracy; in MNIST recognition, the complete configuration (output dimension 72) reduces the network size by 10 times, increases speed by 3 times, and achieves 93.6% accuracy. This invention has a simple structure, is easy to integrate, and provides a general solution for low-power, high-speed optical neuromorphic computing.

Description

Technical Field

[0001] This invention belongs to the field of optical neural networks and neuromorphic computing technology, specifically relating to a method and system for constructing a physical storage layer by utilizing the photon spin-orbit coupling effect in the natural resonant cavity of an organic crystal, and expanding the output dimension through polarization degrees of freedom to accelerate neural network training and reduce network size. Technical Background

[0002] The performance of artificial neural networks is highly dependent on network size, but the resulting computational bottlenecks and energy consumption challenges are becoming increasingly prominent. Reservoir computing in neuromorphic computing replaces most of the training layers in deep networks with a fixed nonlinear physical system (reservoir), training only the output layer weights, thus significantly reducing training complexity. Optical systems, due to their parallelism, high bandwidth, and low power consumption, have become an ideal platform for realizing physical reservoirs. However, the output dimension of existing optical reservoirs is usually limited by the number of spatial channels, making it difficult to effectively expand without increasing structural complexity. To process complex inputs (such as handwritten digits and natural images), it is often necessary to introduce a large number of output nodes or cascade multiple reservoirs, which weakens the simplification advantage of reservoirs. Therefore, how to significantly expand the output dimension with low overhead is a key problem that urgently needs to be solved in this field.

[0003] Spin-orbit coupling (SOC) describes the interconversion between the polarization (spin) and propagation direction (orbit) of light. In planar waveguides or microcavities, the TE-TM splitting caused by the effective mass difference between the transverse electric (TE) and transverse magnetic (TM) modes is one of the most common forms of photonic SOC and can induce the optical spin Hall effect (OSHE)—the separation of circularly polarized light into different spatial regions in real space. Although OSHE has been observed in quantum wells, liquid crystals, and organic microcavities and is widely used in topological photonics and spin optoelectronics, photonic spin-orbit coupling has not yet been used to construct or enhance optical reservoir computing systems.

[0004] Organic semiconductor crystals possess unique molecular packing structures, significant optical anisotropy, natural waveguide resonant cavity characteristics, and room-temperature stable exciton responses, providing an ideal platform for realizing photonic SOCs. In particular, some organic crystals (such as BPDBNA) can spontaneously form hexagonal sheet-like resonant cavities, eliminating the need for additional vapor-deposited mirrors; the refractive index difference between their upper and lower surfaces and air provides effective optical confinement. Based on the natural resonant cavities of such crystals, TE-TM splitting can stably generate a spin Hall effect, outputting optical signals with defined polarization states along the six edges of the crystal. This characteristic is naturally suitable for constructing multidimensional output physical reservoirs: by combining spatial partitioning with polarization channels, the output dimension can be expanded by more than an order of magnitude without increasing device size. Summary of the Invention

[0005] In view of this, it is indeed necessary to provide a neural network computing system for optical storage layers based on organic crystal spin-orbit coupling and its acceleration method to solve the problems of difficulty in expanding the output dimension and limited training speed of existing optical storage layers.

[0006] The specific technical solution of the present invention is as follows:

[0007] First, the organic molecule BPDBNA was synthesized: 4,4'-biphenyldicarboxaldehyde (420.46 mg, 2 mmol) and 2-naphthylacetonitrile (668 mg, 4 mmol) were added to a reactor with a branched tube, and high-purity argon gas was introduced to replace the air in the reactor. 20 mL of anhydrous ethanol and 1 mL of ultra-dry tetrahydrofuran were injected sequentially through a syringe, and the reactor was placed in an ice bath for 1 hour. The solution gradually changed from milky white to light green. 112 mg of potassium tert-butoxide was weighed, dissolved in a small amount of anhydrous ethanol, and slowly added dropwise to the reaction system. The branched tube reactor was transferred to an oil bath and heated to 65°C for 4 hours. After the reaction was completed, the mixture was cooled to room temperature, filtered, and the precipitate was collected. It was washed 2–3 times each with water, ethanol, and dichloromethane. The product was preliminarily dried on a hot plate and then transferred to a vacuum drying oven at 70°C for 6 hours to obtain a yellow powdery BPDBNA product with a yield of 80%.

[0008] Next, BPDBNA organic semiconductor single crystals were prepared: 10 mg of BPDBNA powder was weighed and placed in a quartz boat, which was then placed in a quartz tube inside a horizontal tube furnace; the high-temperature zone of the furnace was set to 285°C, and the low-temperature zone was set to 225°C; a hydrophobic glass substrate was placed in the low-temperature zone, and the substrate was fixed to the inner wall of the quartz tube with high-temperature tape; the system was evacuated to a background pressure of ~10⁻¹ Pa, and then high-purity argon was introduced as a carrier gas at a flow rate of 45 sccm; the heating time was 1 hour, and the holding time was 10 hours; under the transport of the carrier gas, the BPDBNA molecules sublimated in the high-temperature zone were carried to the low-temperature zone, where they condensed and nucleated due to supersaturation, growing into micro-nano single crystals; the resulting crystals were non-regular hexagonal plates with smooth and flat surfaces, a width of 30–40 μm, a length of 20–30 μm, and a thickness of 2–3 μm, with planar waveguide resonant cavities naturally formed on the upper and lower surfaces, eliminating the need for additional vapor-deposited reflectors. The crystal structure of BPDBNA organic semiconductor single crystal is monoclinic. The space group has a molecular transition dipole moment at a 77° angle to the crystal surface; the quality factor Q of the natural planar waveguide resonator is 40–60, and the TE-TM splitting intensity is 0.4–0.6 meV·μm². The dimension of the output vector can be adjusted by changing the number of spatial partitions N and the number of polarization channels to adapt to recognition tasks of varying complexity.

[0009] Finally, an optical storage layer neural network computing system was constructed: the obtained hexagonal organic single crystal was placed on a transparent substrate, and a continuous-wave laser was used as the excitation source. After beam expansion, the laser beam was incident on a spatial light modulator, and a pre-computed hologram was loaded to generate an input pattern to be recognized (such as a simple symbol or MNIST handwritten digit). The generated pattern was focused by an objective lens and perpendicularly irradiated onto the central region of the hexagonal organic single crystal, exciting the crystal to generate a photoluminescence signal. During the propagation of the light-emitting signal inside the crystal, the photon spin-orbit coupling effect induced by TE-TM splitting was used to generate an optical spin Hall effect in real space—lights with different circular polarizations were separated into different spatial quadrants, and six edge outputs had light fields with definite polarization states. The edge output signals were collected by an objective lens, and a quarter-wave plate and a linear polarizer were introduced sequentially in the detection optical path to collect horizontal (H), vertical (V), diagonal (D), anti-angle (A), and left-hand circularly polarized light, respectively. and right-handed circular polarization The image consists of six polarization channels in real space. The six edges of the crystal are divided into N sector regions (N ranges from 6 to 12) according to their spatial positions. The intensity values ​​of each region under the six polarization channels constitute the output vector, with an output dimension of 6×N. This output vector is input into a shallow neural network (the number of input layer nodes equals the output dimension, and the number of output layer nodes equals the number of categories to be identified). Only the output layer weights are trained. Supervised learning is performed using the conjugate gradient algorithm or the backpropagation algorithm to achieve the recognition and classification of the input pattern.

[0010] Using the above technical solution, in 10 simple symbol recognition tasks, when the output dimension is 3 (no polarization, three-sector integration), the training speed is improved by 30 times and the accuracy reaches 99.5%. In the MNIST handwritten digit recognition task, when the output dimension is 72 (6 polarization channels × 12 spatial partitions), the network size is reduced by 10 times, the training speed is improved by 3 times, and the accuracy reaches 93.6%. This system has a simple structure, requires no complex micro-nano fabrication, and is easy to integrate on-chip, providing a new solution for low-power, high-speed optical neuromorphic computing.

[0011] The beneficial effects of this invention are:

[0012] (1) Utilizing the natural hexagonal resonant cavity of organic crystal (BPDBNA), no additional vapor-deposited reflector is required. Optical confinement can be achieved solely by the refractive index difference between the crystal and air interface. The structure is extremely simple, the cost is low, and it is easy to integrate on a chip.

[0013] (2) By moving the position of the excitation spot, the intensity weight of each output channel on the six edges of the crystal can be independently controlled, and the polarization state of each channel is locked by the photon gauge field induced by TE-TM splitting, which is independent of the excitation position. This achieves natural decoupling of intensity modulation and phase encoding, which is convenient for training the weight of the neural network.

[0014] (3) For the first time, photon spin-orbit coupling (optical spin Hall effect) was applied to optical reservoir calculation. The multiple polarization channels (linear polarization and circular polarization) generated by the spin Hall effect significantly expanded the output dimension, increasing the output dimension by more than an order of magnitude without increasing the device size.

[0015] (4) In the task of recognizing ten simple symbols (+, <, >, ∠, ⊥, 7, L, T, X, Γ), a simplified configuration (no polarization, three-sector integration, output dimension 3) was adopted, which increased the training speed by 30 times and the recognition accuracy reached 99.5%, realizing high-speed and high-precision symbol recognition.

[0016] (5) In the MNIST handwritten digit recognition task, the full configuration (six polarization channels × twelve spatial partitions, output dimension 72) was adopted, the network size was reduced by 10 times, the training speed was increased by 3 times, and the recognition accuracy reached 93.6%, which proved the effectiveness and universality of the method for complex inputs.

[0017] (6) The output dimension of the optical storage layer computing system can be flexibly adjusted according to the input complexity (by adjusting the number of spatial partitions and the number of polarization channels), providing a tunable balance between recognition accuracy and training speed, and providing a general solution for optical neuromorphic computing in different application scenarios.

[0018] (7) The system consumes about 1.5pJ of light per recognition and has a photoluminescence lifetime of about 5ns. After training, a single symbol recognition can be completed in a submicrosecond timescale. It has the potential for low power consumption and real-time processing and is expected to be applied to edge computing, optical interconnection and intelligent sensing.

[0019] Compared with existing technologies, this invention utilizes the natural hexagonal resonant cavity of the organic crystal BPDBNA, eliminating the need for additional vapor-deposited mirrors or complex micro / nano fabrication. This results in a minimalist structure, low cost, and ease of on-chip integration. By independently controlling the intensity weights of each output channel at the excitation spot position, and leveraging the TE-TM split-induced optical spin Hall effect, the output dimension is expanded with polarization degrees of freedom (six polarization channels), increasing the output dimension by more than an order of magnitude without increasing device size. In ten symbol recognition tasks, a simplified configuration (output dimension 3) improves training speed by 30 times and achieves 99.5% accuracy. In the MNIST handwritten digit recognition task, a complete configuration (output dimension 72) reduces network size by 10 times, improves training speed by 3 times, and achieves 93.6% accuracy. Furthermore, the system's output dimension can be flexibly adjusted according to input complexity, providing a tunable balance between recognition accuracy and training speed, offering a universal solution for low-power, high-speed optical neuromorphic computing. Attached Figure Description

[0020] Figure 1 This is a synthetic route diagram for the BPDBNA molecule.

[0021] Figure 2 The morphology and crystal structure of the BPDBNA hexagonal single crystal were characterized.

[0022] Figure 3 The absorption and emission spectra of the BPDBNA hexagonal crystal are shown.

[0023] Figure 4 The mode dispersion relation is shown for the natural planar waveguide resonant cavity of the BPDBNA hexagonal crystal.

[0024] Figure 5 This is a schematic diagram of the optical spin Hall effect induced by TE-TM splitting in a BPDBNA crystal.

[0025] Figure 6 Create light path diagrams for ten specific symbols (+, <, >, ∠, ⊥, 7, L, T, X, Γ).

[0026] Figure 7 For ten specific symbols (+, <, >, ∠, ⊥, 7, L, T, X, Γ) in the three-dimensional output space The scatter plot of separation and clustering in the image.

[0027] Figure 8 This is a digital projection optical path diagram of MNIST based on a spatial light modulator (SLM).

[0028] Figure 9 The graphs show the changes in recognition accuracy versus training time under different output dimensions, as well as a comparison of network size reduction. Detailed Implementation

[0029] The following will provide a more detailed description of the optical storage layer neural network computing system and its acceleration method based on organic crystal spin-orbit coupling provided by the present invention, in conjunction with the accompanying drawings and specific embodiments.

[0030] Example 1

[0031] In this embodiment, the system is used to identify ten specific symbols (+, <, >, ∠, ⊥, 7, L, T, X, Γ), specifically including the following steps:

[0032] 1. Synthesis of BPDBNA molecule: 4,4'-biphenyldicarboxaldehyde (420.46 mg, 2 mmol) and 2-naphthylacetonitrile (668 mg, 4 mmol) were added to a reactor with a branched tube, and high-purity argon gas was introduced to replace the air; 20 mL of anhydrous ethanol and 1 mL of ultra-dry tetrahydrofuran were injected sequentially, and the reaction was carried out in an ice bath for 1 hour; 112 mg of potassium tert-butoxide was weighed, dissolved in anhydrous ethanol, and added dropwise to the reaction system; the reaction was transferred to an oil bath at 65°C for 4 hours; after cooling, the mixture was filtered and washed 2–3 times sequentially with water, ethanol, and dichloromethane; after preliminary drying on a hot plate, the mixture was dried under vacuum at 70°C for 6 hours to obtain a yellow powdery BPDBNA product with a yield of 80%.

[0033] 2. Preparation of BPDBNA hexagonal single crystals: Weigh 10 mg of BPDBNA powder and place it in a quartz boat, which is then placed inside a horizontal tubular furnace quartz tube; maintain a high temperature zone of 285℃ and a low temperature zone of 225℃, with a hydrophobic glass substrate placed in the low temperature zone; evacuate to ~ High-purity argon carrier gas (45 sccm) was introduced; the temperature was raised for 1 hour and held for 10 hours; non-regular hexagonal plate-like single crystals were deposited and grown in the low-temperature region, with a width of 30–40 μm, a length of 20–30 μm, and a thickness of 2–3 μm.

[0034] 3. Construction of the optical storage layer system: The obtained hexagonal single crystal is placed on a transparent substrate. A 405nm continuous wave laser is used as the excitation source. After beam expansion, the laser beam is incident on a spatial light modulator (SLM). A pre-calculated hologram is applied to generate light intensity patterns of ten symbols. The light is then focused by an objective lens (100×, NA=0.95) and irradiated perpendicularly onto the center of the crystal to excite photoluminescence. The emitted signal propagates in the crystal and, under the influence of the spin Hall effect induced by TE-TM splitting, outputs a polarization-locked light field from the six edges. The edge signals are collected through the objective lens. No polarizer is added to the detection optical path (no polarization configuration). The six edges are merged into three 120° sector regions according to their adjacent relationships, and the intensity is integrated for each sector. This forms a three-dimensional output vector.

[0035] 4. Neural Network Training and Recognition: The output vector of each symbol is collected after being repeatedly excited 50 times, and a scatter plot is drawn in 3D space. The clouds of different symbols are clearly separated and do not overlap. A 3×10 shallow neural network (3 inputs, 10 outputs) is constructed, and only the output layer weights are trained. The conjugate gradient algorithm is used for supervised learning. After 80 training steps, the recognition accuracy reaches 99.5%, which is 30 times faster than training directly using the original CCD image (high-dimensional input).

[0036] Example 2

[0037] In this embodiment, the system is used to recognize the MNIST handwritten digit dataset (0–9, 10 classes in total), specifically including the following steps:

[0038] 1. Synthesis of BPDBNA molecules and preparation of hexagonal single crystals: Same as steps 1 and 2 of Example 1.

[0039] 2. Construction of the optical storage layer system: Same as step 3 in Example 1, but a polarization measurement module is introduced into the detection optical path. A quarter-wave plate and a linear polarizer are added sequentially behind the objective lens to collect data for horizontal (H), vertical (V), diagonal (D), anti-diagonal (A), and left-handed circular polarization, respectively. and right-handed circular polarization The real-space image has six polarization channels. Each edge of the hexagonal crystal is further subdivided into 12 sector regions, resulting in a total of 6 (polarization) × 12 (space) = 72-dimensional output vectors.

[0040] 3. Automated Data Acquisition: The SLM was controlled by a LabVIEW program to load the MNIST digital hologram (5000 digits in total), and the CCD was controlled to acquire images of each polarization channel. The mechanical shutter blocked the excitation light for 2 seconds after each acquisition to cool the crystal. A total of 35,000 images (5000 digits × 7 polarization states, including unpolarized PL) were acquired, taking 20 hours.

[0041] 4. Neural Network Training and Recognition: A 72-dimensional output vector was input into a shallow neural network (72 nodes in the input layer and 10 nodes in the output layer). Only the output layer weights were trained, and backpropagation was used for supervised learning. After 1000 training steps, the recognition accuracy on the test set (10% of the data) reached 93.6 ± 1.1%. Compared with a shallow network that does not use a physical storage layer (input is the original MNIST image pixels, approximately 784 dimensions), the network size was reduced by about 10 times, and the training speed was increased by 3 times. The output dimension and network performance have a power-law scaling relationship: the error rate increases exponentially with the output dimension. The training time decreases as the output dimension increases exponentially by 0.68.

Claims

1. A neural network computing system based on organic crystal spin-orbit coupling in an optical storage layer and its acceleration method, characterized in that, Includes the following steps: BPDBNA organic molecules were synthesized; non-regular hexagonal plate-like single crystals of BPDBNA were prepared by physical vapor transport method, with planar waveguide resonant cavities naturally formed on the upper and lower surfaces of the crystal; a continuous wave laser was used as the excitation source, and the laser was modulated by a spatial light modulator to generate the input pattern to be identified, which was then perpendicularly irradiated onto the center of the hexagonal crystal to excite photoluminescence; when the luminescent signal propagated in the crystal, it was affected by the TE-TM splitting-induced optical spin Hall effect, and the six edges output light fields with defined polarization states; the edge output signals were collected through an objective lens, and a quarter-wave plate and a linear polarizer were sequentially introduced into the detection optical path to acquire a real space image of at least one polarization channel; the six edges of the crystal were divided into at least one sector region according to their spatial position, and the intensity values ​​of each region under each polarization channel constituted the output vector; this output vector was input into a shallow neural network, and only the output layer weights were trained to achieve the recognition and classification of the input pattern.

2. The method according to claim 1, characterized in that, The synthesis steps of the BPDBNA molecule are as follows: 4,4'-biphenyldicarboxaldehyde and 2-naphthylacetonitrile are added to a reactor, argon gas is introduced, anhydrous ethanol and ultra-dry tetrahydrofuran are injected, the reaction is carried out in an ice bath, and then an ethanol solution of potassium tert-butoxide is added dropwise. The mixture is then transferred to an oil bath at 65°C and reacted for 4 hours. After cooling, filtration, washing and drying, the BPDBNA product is obtained with a yield of 80%.

3. The method according to claim 1, characterized in that, The preparation steps of the BPDBNA non-hexagonal plate-like single crystal are as follows: 10 mg of BPDBNA powder is weighed and placed in a quartz boat, which is then placed inside a horizontal tubular furnace quartz tube; the high-temperature zone is 285°C, and the low-temperature zone is 225°C; a hydrophobic glass substrate is placed in the low-temperature zone; and a vacuum is drawn to ~ High-purity argon carrier gas is introduced at a flow rate of 45 sccm; the temperature is increased for 1 hour and held for 10 hours; the crystal width is 30–40 μm, the length is 20–30 μm, and the thickness is 2–3 μm; the upper and lower surfaces of the crystal naturally form a planar waveguide resonant cavity with a quality factor Q of 40–60 and a TE-TM splitting strength of 0.4–0.6 meV·μm².

4. The method according to claim 1, characterized in that, The natural planar waveguide resonant cavity does not require an external mirror; it achieves optical confinement solely through the refractive index difference between the crystal and air interfaces.

5. The method according to claim 1, characterized in that, The polarization channel is selected from horizontal linear polarization (H), vertical linear polarization (V), diagonal polarization (D), anti-diagonal polarization (A), and left-handed circular polarization. Right-handed circular polarization One or more of the following; when multiple polarization channels are used, each channel is acquired separately, and the output vector dimension is the product of the number of polarization channels and the number of spatial partitions.

6. The method according to claim 1, characterized in that, The spatial partitioning involves merging the six edges of the hexagonal crystal into three 120° sector regions in pairs according to their adjacent relationships, or dividing them into 6 to 12 sector regions at equal angles.

7. The method according to claim 1, characterized in that, The dimension of the output vector is adjusted by changing the number of spatial partitions and the number of polarization channels to adapt to recognition tasks of varying complexity.

8. The method according to claim 1, characterized in that, In the recognition of ten specific symbols (+, <, >, ∠, ⊥, 7, L, T, X, Γ), a non-polarization configuration is adopted, merging the six edges into three 120° sector regions in pairs, with an output dimension of 3; a 3×10 shallow neural network is constructed and trained using the conjugate gradient algorithm. After 80 training steps, the recognition accuracy reaches 99.5%, and the training speed is improved by 30 times.

9. The method according to claim 1, characterized in that, In MNIST handwritten digit recognition, six polarization channels are used. Each edge is divided into 12 sector regions, with an output dimension of 72; a 72×10 shallow neural network is constructed and trained using the backpropagation algorithm, achieving a test accuracy of 93.6±1.1%, reducing the network size by 10 times and increasing the training speed by 3 times.

Images