A multiphase flow global-local physically consistent fusion framework

Abstract

The present application relates to the technical field of numerical simulation of heterogeneous porous medium multiphase flow and deep learning agent modeling, and provides a global-local physical consistent fusion framework for multiphase flow, which comprises a data preprocessing module, a multi-modal fusion module, an anisotropic formation coding module, a channel decoupling Transformer module and a physically embedded double-head decoding module arranged in sequence according to the data flow direction, and is matched with a model training and evaluation module; after the above structure is adopted, through the collaborative design of the multi-modal fusion module, the anisotropic formation coding module, the channel decoupling Transformer module, the physically embedded double-head decoding module and the model training and evaluation module, high-precision, high-physical-consistency and strong-generalization multiphase flow dynamic prediction is realized.

Description

Technical Field

[0001] This invention relates to the field of numerical simulation and deep learning proxy modeling technology for multiphase flow in heterogeneous porous media, specifically a global-local physical consistency fusion framework for multiphase flow. Background Technology

[0002] Accurate modeling of multiphase flow dynamics in heterogeneous porous media is a core fundamental problem in energy development, environmental geology, and water conservancy engineering. Taking oil and gas reservoir development as an example, underground reservoirs exhibit strong heterogeneity, multi-scale pore structures, and complex seepage patterns. Traditional high-fidelity numerical simulation methods, based on Darcy's law and mass balance equations, employ finite-difference numerical solutions. While these methods are physically rigorous and computationally accurate, computational costs increase exponentially with grid size, and a single simulation can take hours to days, making it difficult to support engineering requirements such as large-scale development scheme optimization, uncertainty analysis, and real-time dynamic control. In recent years, data-driven deep learning surrogate models have been widely studied and applied due to their millisecond-level inference speed. They achieve rapid prediction of multiphase flow dynamics by learning the input-output mapping relationship of numerical simulation samples. However, existing multiphase flow surrogate modeling techniques still face several technical bottlenecks, making it difficult to balance prediction accuracy, physical consistency, and engineering practicality.

[0003] The shortcomings of existing technologies are reflected in several dimensions: First, multi-source heterogeneous data suffer from dimensional mismatch. Existing methods often use direct concatenation or simple copy-and-paste to achieve dimensional alignment. Low-dimensional control signals are easily diluted by gradient noise from high-dimensional spatial features during backpropagation, resulting in low sensitivity of the surrogate model to injection-production regime adjustments and poor generalization ability. Second, there is a lack of ability to characterize the anisotropy of sedimentary strata. Traditional 3D convolutional networks generally use symmetric pooling strategies, which indiscriminately compress horizontal and vertical dimensions, destroying the vertical sequence structure of strata and losing key physical information such as differences in interlayer seepage resistance, thus failing to accurately characterize gravity differentiation. The problems include: 1) Vertical flow behaviors such as interlayer crossflow; 2) Lack of physical constraints on the global flow field and local wellpoint response. Existing multi-task proxy models usually treat the global flow field prediction and wellpoint production prediction as two independent regression branches, without considering their inherent physical relationship. This can easily lead to inaccurate predictions in the field but not in the wellpoint, and even produce prediction results that violate the mass balance equation; 3) Multi-task loss weights rely on manual parameter tuning. The loss magnitudes of the global flow field and wellpoint production are very different. Manually setting fixed weights requires repeated adjustments and is difficult to adapt to different training stages. This can easily lead to problems such as single-task-dominated training and gradient conflicts, which increases the difficulty of model training and limits generalization performance. Summary of the Invention

[0004] The purpose of this invention is to provide a global-local physical consistency fusion framework for multiphase flow to solve the problems of dimensional mismatch, loss of anisotropic features, and inconsistency in field-well prediction physics in the prior art.

[0005] A multiphase flow global-local physical consistency fusion framework includes a data preprocessing module, a multimodal fusion module, an anisotropic stratigraphic coding module, a channel decoupling Transformer module, and a physically embedded dual-head decoding module arranged sequentially according to the data flow direction, and is equipped with a model training and evaluation module.

[0006] The data preprocessing module is used to receive the original multiphase flow dataset, complete data normalization, dataset partitioning and adaptive zero-padding processing, and output the preprocessed multimodal data.

[0007] The input of the multimodal fusion module is connected to the output of the data preprocessing module. It is used to convert the low-dimensional control sequence and fluid property sequence into a three-dimensional control potential energy field, and perform attention-weighted fusion with the three-dimensional geological field features in the channel dimension to output the fused multimodal features.

[0008] The input of the anisotropic stratigraphic coding module is connected to the output of the multimodal fusion module. Through a deep locked pooling strategy, it expands the receptive field in the horizontal direction while maintaining the original resolution in the vertical direction, extracts multi-scale anisotropic geological spatial features, and outputs the encoded three-dimensional feature tensor.

[0009] The input of the channel decoupling Transformer module is connected to the output of the anisotropic stratum coding module. The encoded three-dimensional feature tensor is separated by channel to decouple different physical modal features. Three-dimensional position coding is introduced independently and a self-attention mechanism is executed to capture the long-range spatial dependence of the flow field and output the decoupled and enhanced bottleneck feature tensor.

[0010] The input end of the physically embedded dual-head decoding module is connected to the output end of the channel decoupled Transformer module, including a field prediction branch and a well prediction branch, which are used to synchronously output three-dimensional global flow field prediction data and local well point response prediction data that satisfy the physical conservation law. The local well point response prediction data is directly derived from the flow field state through the physically embedded wellbore integral operator.

[0011] The model training and evaluation module is used to complete model training using an adaptive loss function and to evaluate model performance based on the average relative error.

[0012] By adopting the above structure, through the collaborative design of the multimodal fusion module, the anisotropic stratum coding module, the channel decoupling Transformer module, the physically embedded dual-head decoding module, and the model training and evaluation module, high-precision, high-physical-consistency, and strong-generalization multiphase flow dynamic prediction is achieved.

[0013] Preferably, the multimodal fusion module includes a sequence-volume broadcast fusion unit and a channel attention unit, wherein the sequence-volume broadcast fusion unit includes a multilayer perceptron subunit and a tensor reshaping subunit;

[0014] The multilayer perceptron subunit has two built-in linear transformation layers and a GeLU activation function, which are used to map low-dimensional input sequences to a high-dimensional latent space;

[0015] The tensor reshaping subunit is used to convert high-dimensional latent vectors into a three-dimensional control potential energy field consistent with the dimension of the three-dimensional geological field feature space.

[0016] The channel attention unit is used to stitch together the three-dimensional geological field features and the three-dimensional control potential energy field in the channel dimension, and to adaptively recalibrate the fused feature map to enhance the feature weights of key physical variables.

[0017] Preferably, the anisotropic strata coding module is built on an improved VGG-3D architecture and incorporates a depth-locked pooling unit. During downsampling, it maintains the vertical depth dimension resolution unchanged and only performs dimensionality reduction on the horizontal length and width dimensions to preserve the vertical heterogeneity of the sedimentary strata.

[0018] Preferably, the channel decoupling Transformer module includes a spatial flattening subunit, a three-dimensional position encoding subunit, a single-channel Transformer processing subunit, and a feature splicing subunit;

[0019] The spatial flattening subunit is used to split the three-dimensional feature tensor into multiple single-channel features and convert them into a one-dimensional sequence.

[0020] The three-dimensional position coding subunit is used to add learnable three-dimensional position coding to a one-dimensional sequence;

[0021] The single-channel Transformer processing subunit independently performs global self-attention operations on each single-channel feature;

[0022] The feature splicing subunit splices all processed single-channel features along the channel dimension to reconstruct the complete bottleneck feature tensor.

[0023] Preferably, the field prediction branch restores the latent features to a full-resolution three-dimensional spatial grid through multi-level upsampling CNN (Convolutional Neural Networks) blocks, and outputs three-dimensional global flow field prediction data of pressure field, oil saturation field and water saturation field.

[0024] Preferably, the well prediction branch incorporates a wellbore integration operator unit, which is a large-kernel three-dimensional convolutional layer. The convolutional kernel covers the entire reservoir depth and the horizontal region around the well affected by the oil drainage radius, and is used to simulate the radial convergence and vertical accumulation process of fluid, outputting local well point production prediction data.

[0025] Preferably, the depth dimension of the convolution kernel of the wellbore integral operator unit is consistent with the longitudinal depth dimension of the reservoir, and the horizontal dimension of the convolution kernel is matched with the filling dimension to cover the horizontal area around the well affected by the oil drainage radius.

[0026] Preferably, the model training and evaluation module incorporates an adaptive loss calculation unit. This unit integrates field data loss, wellpoint local production loss, and a learnable regularization term. The total loss function dynamically balances the weights of the two losses using an adaptive weighting strategy based on homoscedastic uncertainty. The expression for the total loss function is:

[0027]

[0028] in, Due to the loss of on-site data, For localized production loss at well points, and The logarithm of the observation noise is a learnable parameter used to automatically adjust the weighting of the two losses. This is a learnable regularization term.

[0029] Preferably, the field data loss is used to quantify the deviation between the model-predicted distribution of the overall physical state and the true labels output by the high-fidelity numerical simulator. A global mean square error with a balancing coefficient is used, and the expression for the field data loss is as follows:

[0030]

[0031] in, The total number of valid grid cells participating in the on-site data loss calculation. , and These represent the pressure field, oil saturation field, and water saturation field predicted by the model, respectively, P and S. o and S w These represent the true values ​​for the pressure field, oil saturation field, and water saturation field, respectively. It is the error balance coefficient, used to reconcile the imbalance of error weights between the pressure field and the saturation field due to their different numerical magnitudes and physical meanings.

[0032] The local wellpoint production loss is used to measure the deviation between the model-predicted dynamic production indicators of single-well oil production, gas production, and water production and the actual wellpoint data output by the high-fidelity numerical simulator. Using relative mean square error, the expression for the local wellpoint production loss is:

[0033]

[0034] in, The wellpoint production value predicted by the model. This represents the true value of wellpoint production. This is a numerical smoothing term, with a value of This is used to avoid numerical anomalies where the denominator is zero when the actual output approaches zero. 100.0 is the loss amplification factor, used to increase the overall weight of local output loss at the well point.

[0035] The learnable regularization term is a set of trainable scalar parameters introduced to adaptively adjust the weight ratio of global field data loss and wellpoint local production loss. Regularization constraints prevent extreme weighting. The expression for the learnable logarithmic parameter of the observation noise is:

[0036]

[0037]

[0038] in, and These are the observation noise parameters for the field data task and the well point local task, respectively.

[0039] Preferably, the data preprocessing module has a built-in adaptive zero-fill unit, which is used to fill the irregular reservoir three-dimensional grid data and unify the dimensions of all input spatial attribute features.

[0040] After adopting the above technical solution, the beneficial effects of the present invention are:

[0041] 1. This invention provides a global-local physical consistency fusion framework for multiphase flow. Through a sequence-volume broadcast fusion mechanism set up by a multimodal fusion module, it uses learnable parametric projection to map one-dimensional control sequences and fluid property sequences into a three-dimensional control potential field. This field is then attention-weighted and fused with geological field features, replacing the traditional coarse method of simple copying and splicing. This mechanism is equivalent to injecting distributed source terms into the governing equations, enabling every grid point in the entire field to explicitly perceive global engineering control signals. This eliminates the semantic gap between heterogeneous data at the underlying level, avoids the dilution of low-dimensional control signals by spatial gradient noise, and significantly improves the model's dynamic response to injection-production regime adjustments, maintaining stable prediction accuracy under different development schemes.

[0042] 2. This invention provides a global-local physical consistency fusion framework for multiphase flow. The anisotropic stratum coding module adopts a depth-locked pooling strategy, compressing only the horizontal dimension to expand the receptive field during downsampling, while maintaining the original resolution of the vertical dimension throughout. This design embeds sedimentological prior knowledge, fully preserving the seepage resistance properties and vertical heterogeneity of each layer, avoiding the vertical information aliasing and loss caused by traditional symmetrical pooling. It can accurately characterize complex flow behaviors such as gravity differentiation, interlayer crossflow, and abrupt changes in vertical saturation, effectively suppressing numerical dissipation in flow field prediction and significantly improving the modeling accuracy of layered heterogeneous reservoirs.

[0043] 3. This invention provides a global-local physical consistency fusion framework for multiphase flow, employing a channel-decoupled Transformer architecture. It decomposes the features of different physical channels and performs self-attention operations independently, achieving decoupled modeling of different physical processes such as pressure conduction and saturation transport. On one hand, this design reduces the computational complexity of self-attention, significantly lowering the computational cost of long-range modeling of three-dimensional features. On the other hand, single-channel self-attention forces the model to learn the long-range spatial dependencies of different physical quantities. For example, the pressure channel captures pressure wave conduction in distant injection-production wells, and the saturation channel captures the water drive front propagation of high-permeability channels, effectively avoiding feature entanglement between different physical modes and significantly improving the physical interpretability and long-range process prediction accuracy of the model.

[0044] 4. This invention provides a global-local physically consistent fusion framework for multiphase flow. A physically embedded dual-head decoding module establishes a physical correlation between the field prediction branch and the well prediction branch through a wellbore integral operator. This operator employs a large-kernel three-dimensional convolution covering the entire reservoir depth, mathematically approximating the volume integral form of Darcy's law, and physically simulating the real process of fluid radially flowing into the wellbore from each layer and converging and accumulating vertically. This design transforms sparse production prediction into a volume integral constraint problem based on the flow field, forcing the global flow field and local wellpoint responses to satisfy the physical laws of seepage and the conservation of matter, fundamentally solving the industry pain point of inaccurate field and wellpoint prediction. Simultaneously, the wellbore integral operator acts as a physical bridge, accurately backpropagating production losses to the well perimeter flow field region, effectively alleviating gradient conflicts in multi-task learning, and achieving dual-optimal synergy between global field prediction and wellpoint prediction.

[0045] 5. This invention provides a global-local physically consistent fusion framework for multiphase flow. Based on homoscedastic uncertainty theory, it introduces a learnable regularization term to construct an adaptive loss function. Through backpropagation, it automatically optimizes the weight ratio of global field loss and local well loss, eliminating the need for repeated manual tuning of hyperparameters. The regularization term avoids task collapse caused by extreme weighting, ensuring effective training for both tasks. Simultaneously, this mechanism can adaptively match the inherent noise levels of different tasks, preventing overfitting of the model to high-noise data, resulting in a smoother training process, faster convergence, and improved generalization ability across different datasets. Attached Figure Description

[0046] Figure 1 This is a general framework diagram of the global-local physical consistency fusion framework for multiphase flow provided in the embodiments of the present invention;

[0047] Figure 2 This is a framework diagram of the data preprocessing module provided in an embodiment of the present invention;

[0048] Figure 3 This is a framework diagram of the multimodal fusion module provided in an embodiment of the present invention;

[0049] Figure 4 This is a framework diagram of the anisotropic stratigraphic coding module provided in an embodiment of the present invention;

[0050] Figure 5 This is a framework diagram of the channel decoupling Transformer module provided in an embodiment of the present invention;

[0051] Figure 6 This is a field prediction framework diagram of the physically embedded dual-head decoding module provided in this embodiment of the invention;

[0052] Figure 7 This is a well prediction framework diagram of the physically embedded dual-head decoding module provided in an embodiment of the present invention;

[0053] Figure 8 This is a schematic diagram of the dynamic weighting algorithm provided in an embodiment of the present invention;

[0054] Figure 9 This is a three-dimensional distribution thermogram of the mid-term pressure prediction results of the model evolution provided in this embodiment of the invention;

[0055] Figure 10 This is a three-dimensional distribution thermogram of the mid-stage pressure reference results of the model evolution provided in this embodiment of the invention;

[0056] Figure 11 This is a three-dimensional thermogram showing the relative pressure distribution during the mid-stage of model evolution, provided in this embodiment of the invention.

[0057] Figure 12This is a three-dimensional distribution heat map of the pressure prediction results in the later stage of model evolution provided in this embodiment of the invention;

[0058] Figure 13 This is a three-dimensional distribution thermogram of the pressure reference results in the later stage of model evolution provided in this embodiment of the invention;

[0059] Figure 14 This is a three-dimensional thermogram showing the relative pressure error during the later stages of model evolution, provided in this embodiment of the invention.

[0060] Figure 15 This is a three-dimensional distribution thermogram of the oil saturation prediction results during the mid-stage evolution of the model provided in this embodiment of the invention;

[0061] Figure 16 This is a three-dimensional distribution thermogram of the oil saturation reference result during the mid-stage evolution of the model provided in this embodiment of the invention;

[0062] Figure 17 This is a three-dimensional thermogram showing the relative error of oil saturation during the mid-evolution of the model provided in this embodiment of the invention.

[0063] Figure 18 This is a three-dimensional distribution thermogram of the oil saturation prediction results in the later stage of model evolution provided in this embodiment of the invention;

[0064] Figure 19 This is a three-dimensional distribution thermogram of the oil saturation reference result in the later stage of model evolution provided in this embodiment of the invention;

[0065] Figure 20 This is a three-dimensional thermogram showing the relative error of oil saturation in the later stage of model evolution provided in this embodiment of the invention;

[0066] Figure 21 This is a three-dimensional distribution thermogram of the water saturation prediction results during the mid-evolution of the model provided in this embodiment of the invention;

[0067] Figure 22 This is a three-dimensional distribution thermogram of the water saturation reference result during the mid-stage evolution of the model provided in this embodiment of the invention;

[0068] Figure 23 This is a three-dimensional thermogram showing the relative error of water saturation during the mid-evolution of the model, provided in this embodiment of the invention.

[0069] Figure 24 This is a three-dimensional distribution thermogram of the water saturation prediction results in the later stage of model evolution provided in this embodiment of the invention;

[0070] Figure 25 This is a three-dimensional distribution thermogram of the water saturation reference result in the later stage of model evolution provided in this embodiment of the invention;

[0071] Figure 26 This is a three-dimensional thermogram showing the relative error of water saturation in the later stage of model evolution provided in this embodiment of the invention;

[0072] Figure 27 This is a full-cycle comparison chart of the predicted solution and the reference solution for the gas production of Well4 provided in this embodiment of the invention;

[0073] Figure 28 This is a full-cycle comparison chart of the predicted solution and the reference solution for the gas production of Well5 provided in this embodiment of the invention;

[0074] Figure 29 This is a full-cycle comparison chart of the predicted solution and the reference solution for the gas production of Well8 provided in this embodiment of the invention;

[0075] Figure 30 This is a full-cycle comparison chart of the predicted solution and the reference solution for the gas production of Well9 provided in this embodiment of the invention;

[0076] Figure 31 This is a full-cycle comparison chart of the predicted solution and the reference solution for the gas production of Well10 provided in this embodiment of the invention;

[0077] Figure 32 This is a full-cycle comparison chart of the predicted solution and the reference solution for the gas production of Well11 provided in this embodiment of the invention;

[0078] Figure 33 This is a full-cycle comparison chart of the predicted solution and the reference solution for the oil production of Well4, provided in this embodiment of the invention;

[0079] Figure 34 This is a full-cycle comparison chart of the predicted solution and the reference solution for the oil production of Well5, a production well, provided in this embodiment of the invention.

[0080] Figure 35 This is a full-cycle comparison chart of the predicted solution and the reference solution for the oil production of Well8 production well provided in this embodiment of the invention;

[0081] Figure 36 This is a full-cycle comparison chart of the predicted solution and the reference solution for the oil production of Well9 provided in this embodiment of the invention;

[0082] Figure 37 This is a full-cycle comparison chart of the predicted solution and the reference solution for the oil production of Well10 provided in this embodiment of the invention;

[0083] Figure 38 This is a full-cycle comparison chart of the predicted solution and the reference solution for the oil production of Well11 provided in this embodiment of the invention;

[0084] Figure 39This is a full-cycle comparison chart of the predicted solution and the reference solution for the water production of Well4 provided in this embodiment of the invention;

[0085] Figure 40 This is a full-cycle comparison chart of the predicted solution and the reference solution for the water production of Well5 provided in this embodiment of the invention;

[0086] Figure 41 This is a full-cycle comparison chart of the predicted solution and the reference solution for the water production of Well8 provided in this embodiment of the invention;

[0087] Figure 42 This is a full-cycle comparison chart of the predicted solution and the reference solution for the water production of Well9 provided in this embodiment of the invention;

[0088] Figure 43 This is a full-cycle comparison chart of the predicted solution and the reference solution for the water production of Well10 provided in this embodiment of the invention;

[0089] Figure 44 This is a full-cycle comparison chart of the predicted solution and the reference solution for the water production of Well11 provided in this embodiment of the invention;

[0090] Figure 45 This is an average relative error curve diagram showing the gas production, oil production, and water production of all production wells (Well4, Well5, Well8, Well9, Well10, and Well11) provided in the embodiments of the present invention. Detailed Implementation

[0091] The invention will now be further described with reference to the accompanying drawings.

[0092] The orientations mentioned in this specification are based on the orientation of the invention during normal operation, and do not limit the orientation during storage and transportation. They only represent relative positional relationships and do not represent absolute positional relationships.

[0093] like Figure 1-8 As shown, a multiphase flow global-local physical consistency fusion framework is proposed. It adopts an asymmetric end-to-end deep learning topology and sets up a data preprocessing module, a multimodal fusion module, an anisotropic stratigraphic encoding module, a channel decoupling Transformer module, and a physically embedded dual-head decoding module in sequence according to the data flow direction. It is also equipped with a model training and evaluation module.

[0094] After the original multiphase flow dataset is normalized, partitioned, and adaptively zero-padding by the data preprocessing module, the preprocessed multimodal data is input into the multimodal fusion module. This module converts the low-dimensional control sequence and fluid property sequence into a three-dimensional control potential field, and performs attention-weighted fusion with the three-dimensional geological field features in the channel dimension. The fused multimodal features are then processed by the anisotropic stratigraphic encoding module to extract multi-scale anisotropic geological spatial features. The encoded three-dimensional feature tensor is input into the channel decoupling Transformer module to decouple different physical modal features by channel. Three-dimensional position encoding is independently introduced and a self-attention mechanism is executed to capture the long-range spatial dependence of the flow field. The decoupled and enhanced bottleneck feature tensor is output. Finally, the physically embedded dual-head decoding module synchronously outputs global flow field prediction and local well point production prediction. The model training and evaluation module completes end-to-end training based on an adaptive loss function and evaluates model performance through average relative error.

[0095] The present invention provides a global-local physical consistency fusion framework for multiphase flow, which achieves high-precision, high-physical-consistency, and strong-generalization multiphase flow dynamic prediction through the collaborative design of a multimodal fusion module, an anisotropic stratum coding module, a channel decoupling Transformer module, a physically embedded dual-head decoding module, and a model training and evaluation module.

[0096] like Figure 2 As shown, the data preprocessing module has built-in normalization unit, dataset partitioning unit and adaptive zero-padding unit. The input is the original multiphase flow dataset and the output is the preprocessed multimodal data.

[0097] The normalization unit is used to perform data normalization, that is, to perform Min-Max normalization on the original multiphase flow dataset, map each physical quantity to the [0,1] interval, eliminate the dimensional differences of different physical quantities and accelerate network convergence.

[0098] The dataset partitioning unit is used to partition the dataset. It uses a random sampling strategy to shuffle the samples and strictly divides them into training and validation sets in a 9:1 ratio, which are used for model training and generalization performance monitoring, respectively.

[0099] The data preprocessing module has a built-in adaptive zero-filling unit, which is used to fill irregular reservoir 3D mesh data, unify the dimension of all input spatial attribute features, and perform zero-value filling processing on boundary areas with insufficient spatial dimension to adapt to the tensor dimension requirements of subsequent convolution operations.

[0100] like Figure 3As shown, the input of the multimodal fusion module is connected to the output of the data preprocessing module. The multimodal fusion module includes a sequence-volume broadcast fusion unit and a channel attention unit, which are used to convert one-dimensional control sequences and fluid property sequences into three-dimensional control potential energy fields, and perform attention-weighted fusion with three-dimensional geological field features in the channel dimension to output fused multimodal features.

[0101] Specifically, the multimodal fusion module includes a sequence-volume broadcast fusion unit and a channel attention unit. The sequence-volume broadcast fusion unit includes a multilayer perceptron subunit and a tensor reshaping subunit. The multilayer perceptron (MLP) constructs the projection function. The multilayer perceptron subunit incorporates two linear transformation layers and the GeLU activation function to transform low-dimensional input sequences. Mapping to a high-dimensional latent space to capture its implicit nonlinear control characteristics, latent vectors The formula for calculation is:

[0102]

[0103] in, , , and For learnable parameters, This is the GeLU activation function. It is a two-layer MLP.

[0104] The tensor reshaping subunit operates through tensor reshaping. , latent vector Reconstructed into a three-dimensional control potential energy field consistent with the spatial dimensions of the three-dimensional geological field characteristics. This gives the control signal spatial distribution properties.

[0105] This mechanism is equivalent to injecting distributed source terms into the multiphase flow control equations, enabling every grid point in the field to explicitly perceive the global engineering control signal. This avoids the low-dimensional control signal being diluted by high-dimensional spatial gradient noise during backpropagation, thereby improving the model's response sensitivity to injection-production regime adjustments.

[0106] Channel attention units are used to integrate three-dimensional geological field features with three-dimensional control potential energy fields. The feature maps are concatenated along the channel dimension, and the fused feature maps are adaptively recalibrated to enhance the feature weights of key physical variables, ultimately outputting the fused multimodal features:

[0107]

[0108] in, It is a three-dimensional geological field feature. It is a three-dimensional control potential energy field.

[0109] like Figure 4 As shown, the input of the anisotropic stratigraphic encoding module is connected to the output of the multimodal fusion module. Through a deep-locked pooling strategy, the receptive field is expanded in the horizontal direction by reducing dimensionality, while maintaining the original resolution in the vertical direction. Multi-scale anisotropic geological spatial features are extracted to preserve the vertical heterogeneous physical properties of sedimentary strata and output the encoded three-dimensional feature tensor.

[0110] The anisotropic stratigraphic coding module is built on an improved VGG-3D architecture and incorporates depth-locked pooling units during the downsampling process. In the original text, the anisotropic kernel size is defined as... Its forward propagation calculation formula is:

[0111]

[0112] in, For the first Layer input features, For three-dimensional convolution operations, This is for deep-locked max pooling operations.

[0113] After multi-level downsampling, the vertical depth dimension resolution is kept constant, while only the horizontal length and width dimensions are reduced. By locking the vertical depth dimension resolution and reducing the horizontal length and width dimensions, the Darcy flow resistance distribution in the vertical direction is fully preserved. Multi-scale anisotropic geological spatial features are extracted, and the encoded three-dimensional feature tensor is output. This design incorporates sedimentological prior knowledge, fully preserves the seepage resistance properties and vertical heterogeneity of each layer, avoids the vertical information aliasing caused by traditional symmetrical pooling, and can accurately characterize vertical flow behaviors such as gravity differentiation and interlayer crossflow.

[0114] like Figure 5 As shown, the input of the channel decoupling Transformer module is connected to the output of the anisotropic stratum coding module. The encoded three-dimensional feature tensor is separated by channel to decouple different physical modal features. Three-dimensional position coding is introduced independently and a self-attention mechanism is executed to capture the long-range spatial dependence of the flow field and output the decoupled and enhanced bottleneck feature tensor.

[0115] The channel decoupling Transformer module is located at the bottleneck layer at the end of the encoder. It includes a spatial flattening subunit, a 3D position encoding subunit, a single-channel Transformer processing subunit, and a feature splicing subunit. The spatial flattening subunit is used to split the 3D feature tensor into multiple single-channel features and convert them into a one-dimensional sequence.

[0116] Specifically, assume that the bottleneck feature tensor output by the encoder via the anisotropic stratum coding module is... , The tensor is a real number field, meaning that all elements in the tensor are real values. The total number of channels. For depth, and For planar dimensions; split it into channels. A set of independent potential physical modes: , For the first The single physical modal features corresponding to each channel are the three-dimensional feature tensors. These three-dimensional feature tensors are then processed through spatial flattening sub-units. The features are split into multiple single-channel features and converted into one-dimensional sequences. The flattened original single-channel feature sequences are then obtained. , For the real number field, For depth, and For planar dimensions.

[0117] The 3D position coding subunit is used to add learnable 3D position coding to a 1D sequence. Three-dimensional location encoding can accurately encode the relative positional information of features within a specific sedimentary layer and horizontal well network, resulting in an input sequence with location information.

[0118]

[0119] in, For the first The input sequence after injecting position information into each channel This is the flattened single-channel original feature sequence. It is a learnable 3D positional encoding.

[0120] The single-channel Transformer processing subunit independently performs global self-attention operations on each single-channel feature, processing the input sequence with positional information. The data is fed into a single-channel Transformer block, where the query, key, and value matrices are calculated separately for each single-channel feature. Global self-attention operations are then performed within each single channel, calculated as follows:

[0121]

[0122] in, The first The query, key, and value matrix generated by each channel Scaling factor The Softmax activation function is used because All originating from the same channel, the self-attention mechanism is forced to focus on exploring the global spatial correlation within this specific physical mode (such as only for pressure features), thereby mathematically achieving decoupled modeling of different physical processes and effectively preventing disordered entanglement of pressure and saturation features.

[0123] Ultimately, all After each channel undergoes the aforementioned self-attention calculation independently, they are spliced ​​together along the channel dimension to reconstruct a complete bottleneck feature tensor enhanced by physical means, which is then sent to the decoder.

[0124] The channel decoupling Transformer module employs a "divide and conquer" physical model decoupling strategy, reducing computational complexity from that of standard VIT. Down to This significantly improves computational speed, enabling the pressure channel to capture pressure wave transmission between injection and production wells over long distances, and the saturation channel to capture the water drive front propagation of high-permeability channels, avoiding feature entanglement between different physical modes; thus enhancing the physical interpretability and prediction accuracy of the model.

[0125] like Figure 6 and Figure 7 As shown, the input of the physically embedded dual-head decoding module is connected to the output of the channel-decoupled Transformer module. It includes a field prediction branch and a well prediction branch, used to synchronously output three-dimensional global flow field prediction data and local wellpoint response prediction data that satisfy physical conservation laws. The local wellpoint response prediction data is directly derived from the flow field state through a physically embedded wellbore integration operator. The field prediction branch and the well prediction branch share the feature extraction results from the encoder and the bottleneck layer. Physical constraints between the field and well are established through the physically embedded wellbore integration operator, achieving consistent prediction of the global flow field and local wellpoint response.

[0126] The field prediction branch uses multi-level upsampling CNN (Convolutional Neural Networks) blocks to gradually restore deep latent features into a full-resolution three-dimensional spatial grid. The field prediction branch focuses on reconstructing the pressure field, oil saturation field, and water saturation field of the entire reservoir with high quality, providing a high-precision local flow environment for the well prediction branch, and outputting three-dimensional global flow field prediction data of pressure field, oil saturation field, and water saturation field.

[0127] The well prediction branch incorporates a wellbore integration operator unit, which is a large-kernel three-dimensional convolutional layer. The convolutional kernel covers the entire reservoir depth and the horizontal region around the well affected by the drainage radius. It is used to simulate the radial convergence and vertical accumulation process of fluid and output local well point production prediction data. The depth dimension of the convolutional kernel of the wellbore integration operator unit is consistent with the vertical depth dimension of the reservoir. The horizontal dimension of the convolutional kernel is matched with the fill size to cover the horizontal region around the well affected by the drainage radius.

[0128] The design of the wellbore integral operator in the well prediction branch fully utilizes the vertical resolution preserved by the previous encoder stage "depth locking". The core of the wellbore integral operator is a specially designed wellbore integral convolution kernel. :

[0129]

[0130] in, This is the output of the well prediction branch, i.e., the predicted well production value. The physically embedded wellbore integral convolution kernel is a learnable large-kernel 3D convolutional layer; it covers the entire reservoir thickness in the depth direction, i.e., the depth dimension is [missing information]. It covers the surrounding area within the well's oil supply radius in the horizontal direction; The decoder feature transformation function performs dimensionality adaptation and local feature enhancement on the bottleneck feature tensor. The first output of the Transformer module for channel decoupling Each channel enhances the feature and carries the decoupled single-physical-mode latent space information (such as pressure field and saturation field features).

[0131] Wellbore Integral Operator In mathematical form, it approximates Darcy's law in its volume integral form:

[0132]

[0133] in, This represents the actual total production of a single well in the reservoir, which is the final calculation result of the volume fraction and corresponds to the output of the well prediction branch. ; It is the seepage velocity vector of fluid in the porous medium of the reservoir, which is determined by the pressure gradient, permeability, and fluid viscosity through Darcy's law, and corresponds to the flow field information in the potential space characterized by pressure, oil saturation, and water saturation. This is the radial unit normal vector of the wellbore wall (usually pointing outwards is considered positive). This refers to the normal component of the seepage velocity, meaning that only the velocity component flowing vertically into the wellbore contributes to well production. The seepage influence area (oil supply or drainage area) of a single production well on the horizontal plane corresponds to the coverage area of ​​the convolution kernel in the horizontal direction. The total thickness range of the reservoir. This represents the total number of reservoir layers or the total depth, corresponding to the full coverage of the convolution kernel depth direction. Layer design.

[0134] The wellbore integral operator transforms sparse production prediction into a strongly physically constrained problem based on the volume integral of the local flow field. It forces the global flow field and the local well point response to satisfy the physical laws of seepage and the conservation of matter, thus solving the problem of "accurate field but inaccurate well" at its root. At the same time, it acts as a physical bridge to accurately backpropagate production loss to the flow field region around the well, alleviating gradient conflicts in multi-task learning and achieving dual-optimization synergy between global field and well point prediction.

[0135] In multi-task learning, the wellbore integral operator, as a shared latent space aggregation operation, acts as a "physical bridge" connecting continuous fields and sparse wells. By efficiently backpropagating the total loss function of production to the well periphery feature region of the latent space, it adjusts the gradient magnitude and direction of the two tasks, effectively alleviating the gradient competition and conflict commonly found in conventional fully connected regression, and truly achieving dual-optimal prediction for both fields and wells.

[0136] like Figure 8 As shown, the model training and evaluation module is used to complete model training with an adaptive loss function and to complete model performance evaluation based on the average relative error.

[0137] The model training and evaluation module incorporates an adaptive loss calculation unit. This unit integrates field data loss, wellpoint local production loss, and a learnable regularization term. The total loss function dynamically balances the weights of the two losses using an adaptive weighting strategy based on homoscedastic uncertainty, eliminating the need for manual parameter tuning. The expression for the total loss function is as follows:

[0138]

[0139] in, Due to the loss of on-site data, For localized production loss at well points, and The logarithm of the observation noise is a learnable parameter used to automatically adjust the weighting of the two losses. This is a learnable regularization term.

[0140] The field data loss uses a global mean square error with a balance coefficient to quantify the deviation between the model-predicted distribution of the overall physical state and the true labels output by the high-fidelity numerical simulator. The expression for the field data loss using the global mean square error with a balance coefficient is as follows:

[0141]

[0142] in, The total number of valid grid cells (excluding boundary fill areas) participating in the calculation of data loss in the field. , and These represent the pressure field, oil saturation field, and water saturation field predicted by the model, respectively, P and S. o and S w These represent the true values ​​for the pressure field, oil saturation field, and water saturation field, respectively. It is the error balance coefficient, used to reconcile the imbalance of error weights between the pressure field and the saturation field due to their different numerical magnitudes and physical meanings.

[0143] Local wellpoint production loss is used to measure the deviation between the model-predicted dynamic production indicators of single well oil, gas, and water production and the actual wellpoint data output by the high-fidelity numerical simulator. Using relative mean square error, the expression for local wellpoint production loss is:

[0144]

[0145] in, The wellpoint production value predicted by the model. This represents the true value of wellpoint production. This is a numerical smoothing term, with a value of This is used to avoid numerical anomalies where the denominator is zero when the actual output approaches zero. 100.0 is the loss amplification factor, used to increase the overall weight of the local output loss at the well point, so as to avoid the local output loss at the well point being overwhelmed by the loss of field data in the early stage of training.

[0146] The learnable regularization term is a set of trainable scalar parameters introduced to adaptively adjust the weight ratio of global field data loss and wellpoint local production loss. Regularization constraints prevent extreme weighting. The expression for the learnable logarithmic parameter of the observation noise is:

[0147]

[0148]

[0149] in, and These are the observation noise parameters for field data tasks and wellpoint local tasks, respectively. and These are trainable parameters that are optimized together with the network weights through backpropagation, rather than manually preset hyperparameters.

[0150] Model training was conducted on a CUDA-enabled GPU cluster using PyTorch's Distributed Data Parallelism (DDP) mode, with a fixed global random seed of 11 to ensure the high-precision prediction algorithm's strict reproducibility across multiple comparative experiments. The number of training epochs was set to 1000, and a combination of regularization training techniques was used, including checkpoints to preserve optimal weights, early stopping to monitor validation loss, and an early stopping strategy where training terminated if the validation set loss showed no improvement for 50 consecutive epochs.

[0151] The Adam optimizer is used for parameter iteration, and the learning rate is scheduled using a cosine annealing strategy. The initial maximum learning rate is adaptively calculated based on the following benchmark: It gradually decays to the minimum learning rate over the training cycle. The training process triggers a validation set evaluation every 5 epochs and saves the network weights every 10 epochs. At the same time, the system automatically captures and overwrites the optimal model snapshot with the minimum validation set loss function to ensure that the parameter set with the best generalization performance is extracted.

[0152] The model is divided into a front-end integrated mapping layer and channel attention mechanism for processing temporal data, a feature extraction module based on improved VGG-3D, a Transformer module for feature fusion, and a final output module. To ensure the reproducibility of the results, Table 1 summarizes the hyperparameter settings for all modules.

[0153] Table 1 Hyperparameter Settings

[0154]

[0155] Model performance is quantified and evaluated using Mean Relative Error (MRE). MRE eliminates the influence of dimensions and directly reflects the percentage of prediction error in the true value. The expression is:

[0156]

[0157] Where N is the total number of samples participating in the calculation; for field evaluation, it is the total number of effective grid nodes; for well evaluation, it is the total number of samples at well points and time steps. For the first The true value of each sample For the first The model prediction value for each sample. This is a numerical smoothing term. This indicator eliminates the influence of dimensions and intuitively reflects the percentage of prediction error relative to the true value.

[0158] The complete execution flow of a multiphase flow global-local physical consistency fusion framework provided by this invention is as follows:

[0159] 1. Collect the original multiphase flow dataset through the data preprocessing module, including static geological properties, initial flow field state, well network parameters, injection-production control sequence and fluid physical property parameters, input the data preprocessing module to complete normalization, dataset partitioning and adaptive zero-filling, and output the preprocessed multimodal data.

[0160] 2. The input end of the multimodal fusion module is connected to the output end of the data preprocessing module. The preprocessed low-dimensional control sequence and fluid property sequence are input into the multimodal fusion module, mapped into a three-dimensional control potential energy field through sequence-volume broadcast fusion, and then fused with the three-dimensional geological field features in the channel dimension through attention weighting, outputting fused multimodal features.

[0161] 3. The input end of the anisotropic stratigraphic coding module is connected to the output end of the multimodal fusion module. The multimodal features are fused and input into the anisotropic stratigraphic coding module. Multi-scale anisotropic spatial features are extracted through multi-level convolution with deep locking pooling. Through the deep locking pooling strategy, the receptive field is expanded in the horizontal direction while maintaining the original resolution in the vertical direction. Multi-scale anisotropic geological spatial features are extracted and the encoded three-dimensional feature tensor is output, which is also the encoded deep feature tensor.

[0162] 4. The input of the channel decoupling Transformer module is connected to the output of the anisotropic stratum coding module. The encoded 3D feature tensor is input into the channel decoupling Transformer module. The encoded 3D feature tensor is separated by channel to decouple different physical mode features. 3D position coding is introduced independently and a self-attention mechanism is executed to capture the long-range spatial dependence of the flow field. The decoupled and enhanced bottleneck feature tensor is output.

[0163] 5. The input end of the physically embedded dual-head decoding module is connected to the output end of the channel decoupled Transformer module, including the field prediction branch and the well prediction branch. The decoupled and enhanced bottleneck feature tensor is input into the physically embedded dual-head decoding module. The field prediction branch restores the latent features into a full-resolution three-dimensional spatial grid through a multi-level upsampling CNN block. The well prediction branch derives the well point yield based on the flow field features through the wellbore integral operator. The global flow field and local well point prediction results are output simultaneously.

[0164] 6. The model training and evaluation module is used to train the multiphase flow global-local physical consistency fusion framework provided in this embodiment of the invention. The total loss between the predicted value and the true value is calculated by the adaptive loss function, and the model performance is evaluated based on the average relative error. The network parameters are optimized by backpropagation. The prediction results are directly output during the inference stage and used for engineering scenarios such as development scheme optimization and dynamic control.

[0165] To verify the spatiotemporal evolution extrapolation capability of the global-local physically consistent fusion framework for multiphase flow provided in this invention embodiment in complex multiphase flow systems, cases were randomly selected from the test set from two dimensions: three-dimensional spatial fluid field reconstruction and single-well dynamic time series prediction. The prediction results of the model were visually displayed and compared in depth with real data from a high-precision numerical simulator. Experiments were conducted on a conceptual reservoir dataset, and the prediction results are as follows: Figure 9-45 As shown.

[0166] Figure 9-26 This study visually demonstrates the ability of the Global-Local Physically Consistent Fusion Framework (RISFM) model to dynamically reconstruct the spatial distribution and temporal evolution of multiphase fluid fields in a conceptual example. To comprehensively evaluate the physical fidelity of the model, two key time points were selected: the mid-evolutionary period (T=20 months) and the late-evolutionary period (T=50 months). The pressure field, oil saturation field, and water saturation field predicted by the RISFM model showed a high degree of agreement with the reference solution, and the corresponding spatial distribution of relative errors was presented.

[0167] In pressure field prediction, the maximum relative errors at T=20 and T=50 months were 3.77% and 4.08%, respectively. In oil saturation (So) prediction, the maximum relative errors at T=20 and T=50 months were 1.50% and 2.30%, respectively. For the more drastically evolving water saturation (Sw) field, the RISFM model effectively overcame the numerical dissipation phenomenon at facies abrupt changes, with maximum relative errors at T=20 and T=50 months of 6.91% and 8.12%, respectively. The relative errors were mainly concentrated at the leading edge of facies abrupt changes and near the well point, with errors in most areas below 1%, and approaching 0.1% in the main three-dimensional space (the dark blue area is absolutely dominant). Limited error fluctuations only existed in local grids with extremely drastic gradients, such as facies abrupt change surfaces or well bottoms (sparse light-colored highlighted areas in the figure), with a maximum relative error of less than 8.12%. The results demonstrate that the RISFM model can accurately capture the vertical gravity differentiation and planar water-driven propulsion patterns, effectively suppressing numerical dissipation.

[0168] Figure 27-45 This study comprehensively demonstrates the predictive performance of the RISFM model on the bottomhole flow rate dynamics of multiphase fluids in oil reservoirs over a long-term (five-year) production evolution process. For the full-cycle (five-year) gas, oil, and water production predictions of six producing wells, the prediction curves of this framework closely match the reference solution, showing no lag effect at the production step inflection point caused by injection-production system adjustments. The six producing wells include Well4, Well5, Well8, Well9, Well10, and Well11. The predicted solutions (red scatter dots) for gas production, oil production, and water production of each well were compared with the reference solution (green scatter dots) over the entire cycle. Figure 45The corresponding mean relative error (MRE) is summarized in the table.

[0169] The results show that, due to the influence of interlayer pressure loss, phase changes, and periodic adjustments to the operating regime (injection-production intervention), the production curve of the production well exhibits typical step-like fluctuations and highly nonlinear decay characteristics. (Observation) Figure 27-45 It is evident that, regardless of whether the gas production reaches levels as high as 70,000 m³ / day or the water production fluctuates dramatically, the RISFM model's prediction curves perfectly traverse the scatter points of each reference solution. Especially during annual injection-production system adjustments, at critical inflection points leading to drastic step changes in production, the model demonstrates a highly sensitive response with remarkable physical foresight. In time-series prediction tasks spanning up to five years, errors are prone to exponentially accumulating and amplifying with time steps. However, looking at the evolution trend of single production well curves, the RISFM model's prediction lines in the later stages of evolution still closely adhere to the baseline data. This proves that the "wellbore integral operator" embedded in RISFM successfully couples the macroscopic three-dimensional field evolution with local well production in a physically logical manner, fundamentally blocking the error drift path in the pure time-series autoregressive process.

[0170] The results show that the production prediction error of the RISFM model was strictly controlled to an extremely low level in all six representative wells: in the best-performing production wells Well9 and Well8, the MRE (Mean Reproduction Error) for the gas, oil, and water phases was below 3.5%, with Well9's three-phase error converging remarkably within a small range of 1.3% to 1.9%. Even in Well4, the production well with the greatest prediction difficulty and the most volatile early production, the maximum MRE for oil production was limited to around 8.5%, while the MREs for gas and water production were approximately 5.1% and 8.0%, respectively. Notably, the MRE curves for the gas (red line), oil (black line), and water (blue line) phases showed a high degree of consistency, which indirectly confirms that the model is not only accurate in single-phase prediction but also strictly follows the mass conservation and phase competition laws of reservoir multiphase fluid flow.

[0171] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A global-local physical consistency fusion framework for multiphase flow, characterized in that, It includes a data preprocessing module, a multimodal fusion module, an anisotropic stratigraphic coding module, a channel decoupling Transformer module, and a physically embedded dual-head decoding module arranged sequentially according to the data flow direction, and is equipped with a model training and evaluation module. The data preprocessing module is used to receive the original multiphase flow dataset, complete data normalization, dataset partitioning and adaptive zero-padding processing, and output the preprocessed multimodal data. The input of the multimodal fusion module is connected to the output of the data preprocessing module. It is used to convert the low-dimensional control sequence and fluid property sequence into a three-dimensional control potential energy field, and perform attention-weighted fusion with the three-dimensional geological field features in the channel dimension to output the fused multimodal features. The input of the anisotropic stratigraphic coding module is connected to the output of the multimodal fusion module. Through a deep locked pooling strategy, it expands the receptive field in the horizontal direction while maintaining the original resolution in the vertical direction, extracts multi-scale anisotropic geological spatial features, and outputs the encoded three-dimensional feature tensor. The input of the channel decoupling Transformer module is connected to the output of the anisotropic stratum coding module. The encoded three-dimensional feature tensor is separated by channel to decouple different physical modal features. Three-dimensional position coding is introduced independently and a self-attention mechanism is executed to capture the long-range spatial dependence of the flow field and output the decoupled and enhanced bottleneck feature tensor. The input end of the physically embedded dual-head decoding module is connected to the output end of the channel decoupled Transformer module, including a field prediction branch and a well prediction branch, which are used to synchronously output three-dimensional global flow field prediction data and local well point response prediction data that satisfy the physical conservation law. The local well point response prediction data is directly derived from the flow field state through the physically embedded wellbore integral operator. The field prediction branch restores latent features to a full-resolution three-dimensional spatial grid through multi-level upsampling CNN blocks, and outputs three-dimensional global flow field prediction data of pressure field, oil saturation field and water saturation field. The well prediction branch has a built-in wellbore integration operator unit, which is a large-kernel three-dimensional convolutional layer. The convolutional kernel covers the entire reservoir depth and the horizontal region around the well affected by the oil drainage radius. It is used to simulate the radial convergence and vertical accumulation process of fluid and output local well point production prediction data. The depth dimension of the convolutional kernel of the wellbore integration operator unit is consistent with the vertical depth dimension of the reservoir. The horizontal dimension of the convolutional kernel is matched with the fill size to cover the horizontal region around the well affected by the oil drainage radius. The model training and evaluation module is used to complete model training using an adaptive loss function and to complete model performance evaluation based on the average relative error.

2. The multiphase flow global-local physical consistency fusion framework according to claim 1, characterized in that, The multimodal fusion module includes a sequence-volume broadcast fusion unit and a channel attention unit. The sequence-volume broadcast fusion unit includes a multilayer perceptron subunit and a tensor reshaping subunit. The multilayer perceptron subunit has two built-in linear transformation layers and a GeLU activation function, which are used to map low-dimensional input sequences to a high-dimensional latent space. The tensor reshaping subunit is used to convert high-dimensional latent vectors into a three-dimensional control potential energy field consistent with the dimension of the three-dimensional geological field feature space. The channel attention unit is used to adaptively recalibrate the fused feature map to enhance the feature weights of key physical variables.

3. The global-local physical consistency fusion framework for multiphase flow according to claim 1, characterized in that, The anisotropic strata coding module is built on an improved VGG-3D architecture and incorporates a depth-locked pooling unit. During downsampling, it maintains the vertical depth dimension resolution unchanged and only performs dimensionality reduction on the horizontal length and width dimensions to preserve the vertical heterogeneity of the sedimentary strata.

4. The global-local physical consistency fusion framework for multiphase flow according to claim 1, characterized in that, The channel decoupling Transformer module includes a spatial flattening subunit, a three-dimensional position encoding subunit, a single-channel Transformer processing subunit, and a feature splicing subunit. The spatial flattening subunit is used to split the three-dimensional feature tensor into multiple single-channel features and convert them into a one-dimensional sequence. The three-dimensional position coding subunit is used to add learnable three-dimensional position coding to a one-dimensional sequence; The single-channel Transformer processing subunit independently performs global self-attention operations on each single-channel feature; The feature splicing subunit splices all processed single-channel features along the channel dimension to reconstruct the complete bottleneck feature tensor.

5. The global-local physical consistency fusion framework for multiphase flow according to claim 1, characterized in that, The model training and evaluation module incorporates an adaptive loss calculation unit. This unit integrates field data loss, wellpoint local production loss, and a learnable regularization term. The total loss function dynamically balances the weights of the two losses using an adaptive weighting strategy based on homoscedastic uncertainty. The expression for the total loss function is as follows: in, Due to the loss of on-site data, For localized production loss at well points, and The logarithm of the observation noise is a learnable parameter used to automatically adjust the weighting of the two losses. This is a learnable regularization term.

6. The multiphase flow global-local physical consistency fusion framework according to claim 5, characterized in that, The field data loss is used to quantify the deviation between the model-predicted distribution of the overall physical state and the true labels output by the high-fidelity numerical simulator. It employs a global mean square error with a balancing coefficient, and the expression for the field data loss is as follows: in, The total number of effective grid cells participating in the on-site data loss calculation. , and These represent the pressure field, oil saturation field, and water saturation field predicted by the model, respectively. , and These represent the true values ​​for the pressure field, oil saturation field, and water saturation field, respectively. It is the error balance coefficient, used to reconcile the imbalance of error weights between the pressure field and the saturation field due to their different numerical magnitudes and physical meanings. The local wellpoint production loss is used to measure the deviation between the model-predicted dynamic production indicators of single-well oil production, gas production, and water production and the actual wellpoint data output by the high-fidelity numerical simulator. Using relative mean square error, the expression for the local wellpoint production loss is: in, The wellpoint production value predicted by the model. This represents the true value of wellpoint production. This is a numerical smoothing term, with a value of This is used to avoid numerical anomalies where the denominator is zero when the actual output approaches zero. 100.0 is the loss amplification factor, used to increase the overall weight of local output loss at the well point. The learnable regularization term is a set of trainable scalar parameters introduced to adaptively adjust the weight ratio of global field data loss and wellpoint local production loss. Regularization constraints prevent extreme weighting. The expression for the learnable logarithmic parameter of the observation noise is: in, and These are the observation noise parameters for field data tasks and wellpoint local tasks, respectively.

7. The global-local physical consistency fusion framework for multiphase flow according to claim 1, characterized in that, The data preprocessing module has a built-in adaptive zero-fill unit, which is used to fill the irregular reservoir three-dimensional grid data and unify the dimensions of all input spatial attribute features.

Images