A 3D cloud structure prediction system based on quantum-enhanced neural operators
The 3D cloud structure prediction system using quantum-enhanced neural operators solves the accuracy and computational complexity problems of traditional models in 3D cloud structure prediction, achieving efficient nonlocal feature extraction and high-precision prediction, and supporting domestic deployment.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- EARTH SYST NUMERICAL PREDICTION CENT OF CHINA METEOROLOGICAL ADMINISTRATION
- Filing Date
- 2025-12-12
- Publication Date
- 2026-07-03
AI Technical Summary
Traditional physical models and deep learning models suffer from insufficient accuracy, high computational complexity, locality bias, and insufficient nonlinearity and multi-scale interaction in predicting three-dimensional cloud structures. Existing quantum enhancement models lack a physical basis and fail to effectively utilize the complex physical problems of quantum entanglement modes and cloud evolution.
A three-dimensional cloud structure prediction system based on quantum-enhanced neural operators is adopted, including an input and preprocessing layer, a spatiotemporal encoder, a topology-aware quantum entanglement layer, a dynamic quantum-classical fusion unit, and a three-dimensional deconvolution decoder. Through quantum bit network mapping and classical temporal feature fusion, non-local feature extraction and efficient prediction are achieved.
It significantly improves the accuracy and structural fidelity of 3D cloud structure prediction, reduces computational complexity, supports end-to-end training and operational deployment, and has high computational efficiency and accuracy.
Smart Images
Figure CN121746590B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of quantum computing, and more specifically to a three-dimensional cloud structure prediction system based on quantum-enhanced neural operators. Background Technology
[0002] Spatiotemporal prediction of three-dimensional cloud structure is a key task in numerical weather prediction. Traditional weather forecasting models rely on physical equations (such as cloud microphysical parameterization methods) to simulate cloud formation and evolution. However, these methods face the following problems at high resolution:
[0003] 1) Insufficient accuracy: Traditional physical models are difficult to accurately describe the vertical structure and complex three-dimensional dynamics of clouds, especially in high-resolution predictions where there are large errors.
[0004] 2) High computational complexity: The physical model has a large computational cost, which limits the application of real-time weather forecasting, especially in large-scale weather simulation.
[0005] Deep learning methods (such as Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) have been widely applied in weather forecasting and have achieved significant progress. However, traditional deep learning models still have the following limitations:
[0006] 1) Locality bias: CNN and RNN rely on the extraction of local features and cannot effectively capture the nonlocal spatiotemporal dependencies in 3D cloud fields.
[0007] 2) Insufficient nonlinear and multi-scale interaction: Cloud evolution has strong nonlinear and multi-scale coupling, and traditional methods are difficult to handle these complex interactions at the same time.
[0008] Some studies have attempted to combine quantum computing with deep learning, such as Quantum Long Short-Term Memory (QuantumLSTM) networks, but these methods have the following limitations:
[0009] 1) Inefficient interaction between quantum and classical computing: Classical quantum encoding (such as amplitude encoding or angle encoding) often cannot maintain structure in high-dimensional inputs, thus undermining the advantages of quantum computing.
[0010] 2) Not optimized for specific physical problems: Existing quantum enhancement models lack physical foundations and fail to utilize the topological structure of specific problems, resulting in an inability to effectively link quantum entanglement modes with complex physical problems such as cloud evolution. Summary of the Invention
[0011] To address the aforementioned problems, the purpose of this invention is to provide a three-dimensional cloud structure prediction system based on quantum-enhanced neural operators, which outperforms all existing mainstream models in terms of accuracy, structure fidelity, and computational efficiency.
[0012] This invention provides a three-dimensional cloud structure prediction system based on quantum-enhanced neural operators, comprising:
[0013] An input and preprocessing layer is used to receive the original three-dimensional cloud field and convert the original three-dimensional cloud field into a three-dimensional cloud field with a number of channels;
[0014] A spatiotemporal encoder is used to compress the three-dimensional cloud field into a low-dimensional latent variable space and output the latent variables;
[0015] Topology-aware quantum entanglement layer (TEQL) is used to map the latent variables to a qubit network and output quantum-enhanced features;
[0016] The Dynamic Quantum-Classical Fusion Unit (DQCFU) is used to fuse the quantum enhancement features with classical temporal features to obtain a fusion latent variable.
[0017] A three-dimensional deconvolutional decoder is used to perform stepwise upsampling and skip connections based on the fused latent variables, and output the predicted three-dimensional cloud field for the next two time intervals.
[0018] In one possible implementation, the input and preprocessing layer represent the original three-dimensional cloud field according to the following formula. :
[0019]
[0020] in, The original three-dimensional cloud field at time step 1, The original three-dimensional cloud field at time step 2, For time step The original three-dimensional cloud field at that time.
[0021] In one possible implementation, the spatiotemporal encoder includes: a convolutional skip connection module and a multi-scale Inception module;
[0022] The spatiotemporal encoder, based on the three-dimensional cloud field, performs three 2x downsampling operations by alternately stacking the convolutional skip connection module and the multi-scale Inception module, and outputs latent variables.
[0023] In one possible implementation, the topology-aware quantum entanglement layer (TEQL) maps the latent variables to phase angle vectors and encodes the phase angle vectors into the initial state of the qubits through a parameterized rotation gate.
[0024] Based on the initial state of the qubit, multi-level quantum evolution is performed to obtain quantum enhancement characteristics.
[0025] In one possible implementation, the topology-aware quantum entanglement layer (TEQL) linearly maps latent variables to phase angle vectors according to the following formula:
[0026]
[0027] in, For the Sigmoid function, This is the weight matrix. For bias terms, As a latent variable, This is the phase angle vector.
[0028] In one possible implementation, the topologically aware quantum entanglement layer (TEQL) maps the cloud vortices of the three-dimensional cloud field to a ring-shaped qubit chain, maps the fronts of the three-dimensional cloud field to a linear chain, and couples adjacent qubits through a controlled inverse gate (CNOT) to obtain quantum enhancement features.
[0029] The quantum enhancement features are obtained by performing joint measurements on the quantum enhancement features.
[0030] In one possible implementation, the topology-aware quantum entanglement layer (TEQL) represents the quantum enhancement feature according to the following formula. :
[0031]
[0032] in, It is a quantum state. For measurement operators.
[0033] In one possible implementation, the Dynamic Quantum-Classical Fusion Unit (DQCFU) utilizes LSTMCell to model the local dependencies of time series to obtain classical time series features;
[0034] The quantum enhancement features are concatenated with the hidden states of the LSTM model and input into a fully connected layer to generate four-way gated parameters.
[0035] The quantum enhancement features and classical temporal features are fused according to the following formula to obtain the fused latent variable. :
[0036]
[0037] in, and For four-way gating parameters, For time step The hidden state at that time For quantum enhancement features, It is a fully connected layer.
[0038] In one possible implementation, it further includes: a training module for calculating a loss based on the mean square error (MSE) between the predicted 3D cloud field and the true value and the structural similarity index (SSIM), and training the 3D cloud structure prediction system based on the loss.
[0039] In one possible implementation, the training module calculates the weighted loss according to the following formula. :
[0040]
[0041] in, and Scalar coefficients To predict the mean square error between the 3D cloud field and the true value, To predict the structural similarity index between the 3D cloud field and the true value.
[0042] The 3D cloud structure prediction system based on quantum-enhanced neural operators provided by this invention proposes for the first time a "topology-aware quantum entanglement layer" that maps the topology of a 3D cloud into a configurable qubit network to achieve non-local feature extraction; it introduces a "dynamic quantum-classical fusion unit" to achieve real-time bidirectional interaction between quantum state evolution and LSTM memory states; it constructs a "low-high dimensional hybrid interface" to solve the problem of high-dimensional cloud field compression and reconstruction, balancing accuracy and efficiency; it supports end-to-end training, is compatible with business mode outputs such as CMA-MESO, and has the potential for business deployment; and it outperforms all existing mainstream models in terms of accuracy, structure fidelity, and computational efficiency. Attached Figure Description
[0043] Figure 1 This is a schematic diagram of the structure of the three-dimensional cloud structure prediction system provided in an embodiment of the present invention;
[0044] Figure 2 This is a comparison chart of predictions between the three-dimensional cloud structure prediction system provided in the embodiments of the present invention and common three-dimensional structure prediction models;
[0045] Figure 3 This is a comparison chart of various indicators of the predicted three-dimensional cloud field provided by the three-dimensional cloud structure prediction system in the embodiments of the present invention. Detailed Implementation
[0046] The embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. The following detailed description of the embodiments and the accompanying drawings are used to illustrate the principles of the present invention by way of example, but should not be used to limit the scope of the present invention. That is, the present invention is not limited to the described preferred embodiments, and the scope of the present invention is defined by the claims.
[0047] In the description of this invention, it should be noted that, unless otherwise stated, "a plurality of" means two or more; the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance; those skilled in the art can understand the specific meaning of the above terms in this invention as appropriate.
[0048] To facilitate understanding of this invention, the proprietary terms involved will first be explained.
[0049] Multilayered clouds: Clouds with a layered structure at a single pixel observation point.
[0050] Moderate Resolution Imaging Spectroradiometer (MODIS) is a sensor technology.
[0051] Feng-Yun-4A Satellite, FY-4A, 4.2m × 2.0m × 3.8m (length × width × height);
[0052] The Advanced Geostationary Radiation Imager (AGRI) is a payload instrument on the FY-4A satellite.
[0053] Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observations (CALIPSO) is a satellite observation project using lidar (Lidar) technology.
[0054] The Cloud-Aerosol Lidar with Orthogonal Polarization (CALIOP) is a lidar device carried on the CALIPSO satellite.
[0055] Vertical feature mask (VFM) is a secondary product created from observation data of the CALIOP lidar system aboard the CALIPSO satellite.
[0056] This invention, Quantum-Enhanced Neural Operators for 3D Cloud Forecasting (QENO), is a quantum-classical hybrid intelligent prediction system for 3D cloud fields, designed for numerical weather prediction. It takes the 3 km resolution output of the China Meteorological Administration's CMA-MESO model as input. A lightweight CNN encoder first compresses the massive 3D cloud field into low-dimensional latent variables. Then, a "topologically aware quantum entanglement layer" maps these latent variables to a configurable qubit network, allowing complex structures such as vortices and fronts within the cloud to be explicitly expressed in the form of quantum entanglement, thus overcoming the limitation of traditional convolution, which can only perceive local conditions. The compressed quantum state undergoes parameterized rotation and entanglement gate evolution, and then a "dynamic quantum-classical fusion unit" modulates the memory gating of a classical LSTM in real time, achieving the fusion of quantum nonlocal features and classical temporal features. Finally, a deconvolutional decoder reconstructs high-fidelity 3D cloud fields for the next two time intervals.
[0057] Figure 1 A schematic diagram of the structure of the three-dimensional cloud structure prediction system provided in the embodiments of the present invention is shown below. Figure 1 As shown, the present invention provides a three-dimensional cloud structure prediction system based on quantum-enhanced neural operators, including: an input and preprocessing layer, a spatiotemporal encoder, a topology-aware quantum entanglement layer (TEQL), a dynamic quantum-classical fusion unit (DQCFU), and a three-dimensional deconvolution decoder.
[0058] The input and preprocessing layer is used to receive the original 3D cloud field and fold it into a 3D cloud field with a number of channels.
[0059] In one possible implementation, the original three-dimensional cloud field is represented by the following formula. :
[0060]
[0061] in, The original three-dimensional cloud field at time step 1, The original three-dimensional cloud field at time step 2, For time step The original three-dimensional cloud field at that time.
[0062] The three-dimensional cloud field can be represented by the following formula. :
[0063]
[0064] in, For time, For horizontal resolution, For vertical number of layers, For cloud structured data, For the original number of channels, For the number of channels after conversion, For width.
[0065] In one example, assuming a past T=5 time periods, horizontal resolution H=W=8, vertical layer number Z=4, and a single-element cloud field (such as cloud mask or cloud water content), the original input would be:
[0066]
[0067] In implementation, Z is folded to the channel dimension to obtain To facilitate 3D / 2.5D convolution processing; the values are first normalized to [0,1].
[0068] The spatiotemporal encoder is used to compress a three-dimensional cloud field into a low-dimensional latent variable space and output the latent variables.
[0069] In one possible implementation, the spatiotemporal encoder includes a convolutional skip connection module (ConvSC) and a multi-scale inception module (MSIM).
[0070] This module is responsible for compressing high-dimensional 3D cloud field input data into a low-dimensional latent variable space to reduce the computational overhead of subsequent quantum processing. It employs a ConvSC (Convolutional Skip Connection) structure, using 3×3×3 convolutions with learnable strides to achieve layer-by-layer downsampling, while skip connections prevent gradient vanishing. It integrates the multi-scale Inception module MSIM, using 3×3, 5×5, and 7×7 convolutional kernels in parallel to extract cloud structure features at different scales, effectively modeling everything from single-unit convection to large-scale fronts. The output is a low-dimensional latent variable, which serves as the input for quantum encoding.
[0071] The specific process is as follows: The spatiotemporal encoder is based on a three-dimensional cloud field. It uses the ConvSC convolutional skip connection module and the MSIM multi-scale inception module to be stacked alternately, and performs 3 times downsampling (total 8×) to output latent variables.
[0072]
[0073] In one example, we set B=1, H′=W′=1, and C′=32. Therefore, the encoded output at each time t is... .
[0074] The MSIM module consists of three parallel convolutional branch kernels of sizes 3×3, 5×5, and 7×7, which are then concatenated and compressed using a 1×1 kernel to ensure that both convective-scale details are extracted and large-scale frontal structures are covered.
[0075] Topology-aware quantum entanglement layer (TEQL) is used to map latent variables to a qubit network and output quantum-enhanced features;
[0076] In one possible implementation, latent variables are mapped to phase angle vectors via a parameterized rotating gate. , , The phase angle vector is encoded into the initial state of the qubit; multi-level quantum evolution is performed based on the initial state of the qubit to obtain the quantum enhancement feature.
[0077] In one possible implementation, the latent variables are calculated according to the following formula. Linear mapping to phase angle vector (Packaging for the number of qubits n):
[0078]
[0079] in, For the Sigmoid function, This is the weight matrix. For bias terms, As a latent variable, This is the phase angle vector.
[0080] Every angle Acting on a single-bit composite rotation (in implementation: For easier expression, simplified methods can be used. (Derivation explanation)
[0081] The phase angle vector is encoded into the initial state of the qubit according to the following formula:
[0082]
[0083] in, Indicates at a given time and index The quantum state is below and the quantum state is initially 0. Indicates surrounding The rotation angle of the axis is , Indicates surrounding Rotation angle of the axis , Indicates surrounding Rotation angle of the axis .
[0084] In one possible implementation, the quantum enhancement features are obtained by performing multi-layer quantum evolution based on the initial state of the qubits, including: mapping the cloud vortex of the three-dimensional cloud field to a ring qubit chain, mapping the front of the three-dimensional cloud field to a linear chain, and coupling adjacent qubits through a controlled inverse gate CNOT to obtain a quantum state; and performing joint measurement on the quantum state to obtain the quantum enhancement features.
[0085] This module maps compressed latent variables to a qubit network, leveraging the principles of quantum entanglement and superposition to enhance feature representation capabilities. It constructs a qubit entanglement graph based on cloud field structural characteristics. For example, cloud vortices in a three-dimensional cloud field are mapped as ring-shaped qubit chains, and fronts in the three-dimensional cloud field are mapped as linear chains. Adjacent qubits are coupled using controlled inverse gates (CNOTs), making quantum correlations isomorphic to the physical cloud field structure. After multi-layer quantum evolution, joint measurements of the quantum states yield quantum-enhanced features, explicitly containing nonlocal and cross-scale spatiotemporal dependencies.
[0086] For the "vortex" structure, an 8-bit ring entanglement graph is constructed; for the "front," a linear chain is constructed. Taking the ring as an example:
[0087]
[0088] The complete layer is
[0089]
[0090] Finally, to Pauli By performing a joint expectation measurement, the quantum enhancement feature is obtained:
[0091] In one possible implementation, the topology-aware quantum entanglement layer (TEQL) represents the quantum enhancement feature according to the following formula. :
[0092]
[0093] in, It is a quantum state. For measurement operators.
[0094] By mapping or concatenating the quantum states of multiple qubit groups according to their channels, quantum enhancement features are obtained. ,and Same dimension. The above TEQL's "topologically matched entangled network - joint measurement - quantum enhanced feature" process and formula (layered) and expect).
[0095] Dynamic quantum-classical fusion unit (DQCFU) is used to fuse quantum enhancement features with classical temporal features to obtain fusion latent variables.
[0096] In one possible implementation, this module achieves the fusion of quantum enhancement features and classical temporal features:
[0097] Classic approach: Modeling local dependencies in time series using LTMCell.
[0098] Quantum Branch: Quantum Enhancement Feature The hidden states of the LSTM are concatenated and input into the fully connected layer to generate four gating parameters (g0, g1, g2, g3).
[0099] Specifically, LTSMCell is used to model the local dependencies of time series to obtain classical time series features; quantum-enhanced features are then applied. The hidden states of the LSTM model are concatenated and input into the fully connected layer to generate four gating parameters.
[0100] The quantum enhancement features and classical temporal features are fused according to the following formula to obtain the fused latent variable. :
[0101]
[0102] in, and For four-way gating parameters, For time step The hidden state at that time For quantum enhancement features, It is a fully connected layer.
[0103] Through this gating mechanism, the quantum measurement results modulate the classical memory state in real time, ensuring that the model has both physical consistency and avoids noise accumulation caused by deep quantum circuits.
[0104] At each time t, first use LSTM to obtain ,in Extract low-dimensional features (or their linear transformations) of the encoder. (This is related to the quantum side.) Fusion:
[0105]
[0106] Will Divided into four sections and element-wise multiplication Fusion:
[0107]
[0108] The 3D deconvolutional decoder is used to perform stepwise upsampling and skip connections based on fused latent variables, and output the predicted 3D cloud field for the next two time intervals.
[0109] In one possible implementation, the loss is calculated based on the mean square error (MSE) between the predicted 3D cloud field and the true value, and the structural similarity (SSIM) is used to train the 3D cloud structure prediction system.
[0110] This module restores the fused latent variables into a full-resolution future 3D cloud. It employs a 3D Transposed Convolution network to achieve progressive upsampling and recovers detailed information through skip connections. The final output is the predicted 3D cloud field for the next two time intervals, which is compared with the true values. End-to-end training is performed based on the MSE + SSIM joint loss.
[0111] Use level 3 3D deconvolution (with jump connections) to Upsampled 8× stepwise to restore to H×W, outputting the 3D cloud field for the next T′=2 frames. Weighted loss used for training:
[0112] In one possible implementation, a training module is used to calculate the weighted loss according to the following formula. :
[0113]
[0114] in, and Scalar coefficients To predict the mean square error between the 3D cloud field and the true value, To predict the structural similarity index between the 3D cloud field and the true value.
[0115] To facilitate checking the end-to-end shape and numerical validity, a sample setting that can be manually verified is provided (this is not equivalent to production parameters and is for verification purposes only):
[0116] Let T=2, H=W=4, Z=2. =2.
[0117] The first layer of the Encoder, ConvSC, uses a kernel. (Equivalent to 2D average pooling), stride 2; MSIM retains only the 3×3 branch (weight unit convolution), thus:
[0118]
[0119] Get every moment (Take C′=8)
[0120] To facilitate manual calculation, the single-bit encoding is temporarily simplified to only... (Without prejudice to generality, the actual implementation is as follows) At this point, a single bit:
[0121] .
[0122] Without entanglement, n-bit components can be measured:
[0123]
[0124] After adding the ring-shaped CNOT, Become about The higher-order terms of the mixed nonlinear function (without a simple closed-form expression), but its first derivative can be rigorously calculated using the parameter shift rule as follows:
[0125] For any gate parameter θ (e.g. (the angle of rotation), expectation about the observable O:
[0126]
[0127] Satisfy parameter shift:
[0128]
[0129] Proof points: For a single-parameter unitary gate with generator Pauli... ,have Substitute f′(θ) and use Baker–Campbell–Hausdorff with Simplifying, we obtain the above formula. Applied to this invention: for quantum coding. Both TEQL stacking and CNOT satisfy the parameter shifting premise, so the entire TEQL can accurately backpropagate gradients end-to-end through this rule without the need for numerical difference approximation.
[0130] DQCFU: Take the hidden dimension d=8, let The output (check that its range is within (0,1)) is used to verify the fusion. and It is homogeneous and numerically bounded. The explicit gradient of DQCFU is calculated as follows:
[0131]
[0132] Gating parameters The calculation formula is as follows:
[0133]
[0134] This ensures that the chain gradient of quantum measurement → gating → LSTM is stable and computable.
[0135] Decoder: Three layers of deconvolution restore 1×1 to 4×4, output X^ and perform MSE / SSIM with the true value.
[0136] Boundary check: If the input is all zeros or all ones, When it degenerates to {0,π}, Z ∈{±1}\, network output and jump connections ensure numerical stability.
[0137] Additional engineering details:
[0138] Training configuration: 100 epochs, batch=8, lr=1e-3, single V100 card; the quantum part is simulated using TorchQuantum, supporting seamless migration to real quantum chips in the future.
[0139] Business coupling: Directly read the CMA-MESO 3 km / 3 h output; the entire system is embedded into the existing assimilation loop in the form of Python-API, and the cloud microphysics module is upgraded with zero modifications.
[0140] Through the synergy of the above four modules, QENO has for the first time put the concept of "quantum entanglement-cloud topological isomorphism" into practice in a 3D cloud forecasting task, significantly improving the ability to express non-local and cross-scale features, and verifying its advantages of high accuracy, scalability and deployability on real business data.
[0141] Figure 2 This is a comparison chart of the predictions of the 3D cloud structure prediction system provided in the embodiments of the present invention with those of common 3D structure prediction models. Through comparison with existing models, Figure 2 This clearly demonstrates the significant advantages of the present invention in terms of prediction accuracy and structural fidelity. Figure 2 The figure includes a comparison of the standard Ground Truth Channel 15 graph and several common models (such as SimVP, Earthformer, TAU, PredRNN_PlUS, ConvLSTM, SimVP_Plus, PhyDNet, MAU, etc.) with the QENO model in terms of prediction accuracy, detail capture capability, and structure fidelity. This figure visually demonstrates the superiority of the system proposed in capturing multi-scale spatiotemporal dependencies, nonlocal features, and cloud field structure fidelity.
[0142] The system of this invention was validated on a 64×64×42 grid dataset: The experiment used a 3D cloud field dataset generated by CMA-MESO, an operational regional numerical weather prediction (NWP) system with a spatial resolution of 3 km and a temporal resolution of 3 hours, covering 42 vertical levels and cropped at a 64×64 grid. All model training and inference were performed on an NVIDIA Tesla V100 GPU with CUDA acceleration, ensuring efficient computational processing. For the quantum component, the torchquantum framework was used to simulate parameterized quantum circuits, and their output was integrated with a classical spatiotemporal model. Each model was trained for 100 epochs with a batch size of 8 and an initial learning rate of 0.001, optimized using mean squared error (MSE) loss. In the experiments, the model was trained from the previous five frame sequences (T... The predictions for the next two time steps (T+1 and T+2) are generated in step 4 to T. For benchmarking, we compared QENO with eight state-of-the-art baseline models, including ConvLSTM, PhyD-Net, MAU, SimVP, SimVP Plus, Earthformer, TAU, and PredRNN Plus.
[0143] like Figure 3 As shown, among all models, QENO performed exceptionally well across all regression metrics. Particularly in terms of MSE (0.2038), MAE (0.2553), and RMSE (0.4515), its performance was more accurate than other competing models. QENO's SSIM value (0.6292) was also significantly higher than other models, indicating a significant advantage in maintaining structural similarity. Specifically, compared to SimVP (MSE 0.3940, RMSE 0.6277) and SimVP Plus (MSE 0.3477, RMSE 0.5897), QENO's MSE and RMSE were significantly lower, indicating higher accuracy in spatiotemporal structure prediction. This comprehensive evaluation quantifies the advantages of quantum-enhanced feature extraction in 3D cloud prediction. It provides a feasible, scalable, and fully autonomous and controllable intelligent upgrade path for cloud microphysics in domestic high-resolution numerical weather prediction systems.
[0144] The key protection points of this invention are as follows:
[0145] Topology-aware quantum entanglement layer (TEQL): maps the topology of a three-dimensional cloud into a quantum bit network structure;
[0146] Dynamic Quantum-Classical Fusion Unit (DQCFU): Enables real-time bidirectional interaction between quantum states and LSTM memory states;
[0147] Low-High Dimension Hybrid Interface (LHDHI): Enables compression, quantum encoding, and reconstruction of high-dimensional cloud fields;
[0148] An end-to-end trainable operational system, compatible with domestic numerical models such as CMA-MESO;
[0149] It is applicable to various meteorological scenarios such as 3D cloud field prediction, radar echo prediction, and cloud water content inversion.
[0150] Problems that existing technologies need to solve:
[0151] 1) Local convolution / attention mechanisms cannot model long-range spatial dependencies in cloud fields;
[0152] 2) Classical models cannot characterize nonlocal correlations in cloud fields that resemble quantum entanglement;
[0153] 3) Lack of end-to-end systems that deeply integrate quantum computing with classical spacetime modeling;
[0154] 4) Unable to achieve scalable, interpretable, and business-deployable predictive capabilities in high-dimensional cloud environments.
[0155] Compared with the prior art, the present invention has the following advantages:
[0156] 1) Improved accuracy: Outperforms existing models on all regression metrics;
[0157] 2) Structural fidelity: SSIM is significantly improved, with strong ability to reproduce details;
[0158] 3) Service Adaptation: Supports CMA-MESO output and has hourly and kilometer-level forecasting capabilities;
[0159] 4) Highly scalable: It can be extended to multi-element prediction such as radar echo, precipitation, and wind field;
[0160] 5) Domestic production: It does not rely on foreign algorithms and supports the deployment of domestically produced quantum chips.
[0161] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A three-dimensional cloud structure prediction system based on quantum-enhanced neural operators, characterized in that, include: An input and preprocessing layer is used to receive the original three-dimensional cloud field and convert the original three-dimensional cloud field into a three-dimensional cloud field with a number of channels; A spatiotemporal encoder is used to compress the three-dimensional cloud field into a low-dimensional latent variable space and output the latent variables; Topology-aware quantum entanglement layer (TEQL) is used to map the latent variables to a qubit network and output quantum-enhanced features; The Dynamic Quantum-Classical Fusion Unit (DQCFU) is used to fuse the quantum enhancement features with classical temporal features to obtain a fusion latent variable. A three-dimensional deconvolutional decoder is used to perform stepwise upsampling and skip connections based on the fused latent variables, and output the predicted three-dimensional cloud field for the next two time periods. The topology-aware quantum entanglement layer (TEQL) maps the latent variables to phase angle vectors and encodes the phase angle vectors into the initial state of the qubits through a parameterized rotation gate. Based on the initial state of the qubit, multi-level quantum evolution is performed to obtain quantum enhancement features; The process of performing multi-layer quantum evolution based on the initial state of the qubits to obtain quantum enhancement features includes: the topologically aware quantum entanglement layer (TEQL) maps the cloud vortices of the three-dimensional cloud field to a ring-shaped qubit chain, maps the fronts of the three-dimensional cloud field to a linear chain, and couples adjacent qubits through a controlled inverse gate (CNOT) to obtain quantum enhancement features.
2. The three-dimensional cloud structure prediction system according to claim 1, characterized in that, The input and preprocessing layers represent the original three-dimensional cloud field according to the following formula. : in, The original three-dimensional cloud field at time step 1, The original three-dimensional cloud field at time step 2, For time steps The original three-dimensional cloud field at that time.
3. The three-dimensional cloud structure prediction system according to claim 1, characterized in that, The spatiotemporal encoder includes: a convolutional skip connection module and a multi-scale Inception module; The spatiotemporal encoder, based on the three-dimensional cloud field, performs three 2x downsampling operations by alternately stacking the convolutional skip connection module and the multi-scale Inception module, and outputs latent variables.
4. The three-dimensional cloud structure prediction system according to claim 1, characterized in that, The topology-aware quantum entanglement layer (TEQL) linearly maps latent variables to phase angle vectors according to the following formula: in, For the Sigmoid function, This is the weight matrix. For bias terms, As a latent variable, This is the phase angle vector.
5. The three-dimensional cloud structure prediction system according to claim 1, characterized in that, The topology-aware quantum entanglement layer (TEQL) represents the quantum enhancement feature according to the following formula. : in, It is a quantum state. For measurement operators.
6. The three-dimensional cloud structure prediction system according to claim 1, characterized in that, The Dynamic Quantum-Classical Fusion Unit (DQCFU) utilizes LSTMCell to model the local dependencies of time series and obtain classical time series characteristics. The quantum enhancement features are concatenated with the hidden states of the LSTM model and input into a fully connected layer to generate four-way gated parameters. The quantum enhancement features and classical temporal features are fused according to the following formula to obtain the fused latent variable. : in, and For four-way gating parameters, For time steps The hidden state at that time For quantum enhancement features, It is a fully connected layer.
7. The three-dimensional cloud structure prediction system according to claim 1, characterized in that, Also includes: The training module is used to calculate the loss based on the mean square error (MSE) between the predicted 3D cloud field and the true value and the structural similarity index (SSIM), and to train the 3D cloud structure prediction system based on the loss.
8. The three-dimensional cloud structure prediction system according to claim 7, characterized in that, The training module calculates the weighted loss according to the following formula. : in, and Scalar coefficients To predict the mean square error between the 3D cloud field and the true value, To predict the structural similarity index between the 3D cloud field and the true value.