A Method and System for Predicting Multi-Response Transport Alignment and Physical Closure of Underwater Vehicles
By synchronously extracting multi-response data of underwater vehicles and constructing reversible transport mapping and discrete physical closed-loop operators, the problems of feature transfer and physical inconsistency in the prediction of hydrodynamic response of underwater vehicles are solved, and high-precision, physically consistent prediction results are achieved.
Patent Information
- Application Number
- CN202611310248.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-27
- Publication Date
- 2026-11-10
AI Technical Summary
Existing methods for rapid prediction of hydrodynamic response of underwater vehicles have significant shortcomings in feature transfer processing, cross-response consistency, and physical law protection. They cannot accurately obtain the wake field, drag time history, and hull surface load distribution, resulting in decreased reconstruction accuracy and physical inconsistencies.
By synchronously extracting multi-response data of underwater vehicles, a reversible transport mapping and a discrete physical closure operator are constructed to separate the physical feature position migration from the amplitude shape change. Prediction is then performed using a shared-private latent variable joint order reduction and a probabilistic surrogate model to ensure that the prediction results meet physical constraints.
It achieves high-precision, physically consistent underwater vehicle response prediction in a wide parameter space, reduces the number of modes, maintains local peak accuracy, and ensures the physical feasibility and consistency of the prediction results during the online prediction phase.
Smart Images

Figure CN123069898A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of shipbuilding and ocean engineering, underwater vehicle hydrodynamics and computational fluid dynamics numerical simulation, and specifically to a method and system for predicting multi-response transport alignment and physical closure of underwater vehicles. Background Technology
[0002] When an underwater vehicle navigates near the free surface, it simultaneously induces three important hydrodynamic responses: the free surface wake field (including wave morphology and height changes), the drag time history (including transient peak and steady-state values), and the hull surface load distribution (including pressure and wall shear stress). Accurately acquiring these three responses is of great significance for the vehicle's performance evaluation, attitude control, and stealth performance analysis.
[0003] Traditional engineering assessments typically rely on high-fidelity computational fluid dynamics (CFD) solutions for each specific condition. Each change in parameters such as speed, depth, configuration, or attitude necessitates remodeling, solving, and post-processing, with each calculation taking hours to days. In scenarios involving wide parameter space optimization or real-time decision support, the computational cost and time required by this approach are insufficient to meet engineering needs.
[0004] To overcome the aforementioned bottlenecks, rapid prediction methods based on reduced-order models and surrogate models have emerged in recent years. The typical technical approach involves generating several sample working conditions in the parameter space and performing CFD calculations. For each type of response, Proper Orthogonal Decomposition (POD), Principal Component Analysis (PCA), or Singular Value Decomposition (SVD) is performed to achieve dimensionality reduction. Then, Gaussian process regression or neural networks are used to establish the mapping relationship between working condition parameters and modal coefficients. Finally, the target response is reconstructed through modal superposition. This method has achieved some success in some simple flow problems, but existing solutions still have the following significant shortcomings in practical applications: First, the physical features shift with changing operating conditions, leading to distortion in the reduction of order (which can be simply understood as "features moving"). At shallow depths or high speeds, the wake peaks and troughs, the peak pressure on the hull surface, the flow separation line, and the extreme drag points all undergo significant spatial or temporal shifts with changes in parameters such as speed and depth. Existing methods directly apply reduction of order on fixed physical coordinates, failing to distinguish between "positional shifts" and "morphological changes." They mistakenly model feature movement as shape changes, resulting in a significant increase in the number of required modes, the smoothing of local peaks, and a decrease in reconstruction accuracy.
[0005] Secondly, independent predictions of various responses lead to inconsistencies in cross-response physics (simply put, "each telling its own story"). The wake field, drag history, and surface loads are governed by a unified flow state and physical laws; physically, drag should equal the integral of the hull surface pressure and wall shear stress along the wetted surface. However, existing methods treat various responses as independent prediction tasks, establishing separate surrogate models for each. The result is that, under the same operating conditions, the drag obtained from the independent predicted surface load integration differs from the independently predicted drag value, leading to a lack of unified physical basis for engineering decisions.
[0006] Third, the lack of physical constraints in surrogate models leads to predictions that violate fundamental physical laws. Purely data-driven surrogate models only pursue statistical fitting accuracy; even if their predicted mean is within the confidence interval, it may still violate fundamental physical conditions such as surface force integral relationships, conservation of momentum of the control volume, or the invertibility of coordinate mapping. Existing methods only output the prediction variance to characterize statistical uncertainty, failing to actively identify and eliminate such physical infeasibility.
[0007] In summary, existing methods for rapid prediction of hydrodynamic response of underwater vehicles have significant shortcomings in feature transfer processing, cross-response consistency, and physical law protection. There is an urgent need for a new prediction method that can uniformly address these three types of problems. Summary of the Invention
[0008] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for multi-response transport alignment and physical closure prediction of underwater vehicles. It separates the position migration and amplitude and shape changes of physical features, couples the wake field, drag history and surface load through shared latent variables, and directly writes the physical equation between drag and drag response obtained by integrating the distributed load into the online prediction process, thereby obtaining prediction results with clear feature positions, consistent across responses and satisfying physical constraints.
[0009] To address the aforementioned technical problems, this invention provides a method for predicting multi-response transport alignment and physical closure of underwater vehicles, comprising data acquisition and preprocessing, model building and training, and model application stages.
[0010] I. Data Acquisition and Preprocessing Stage: S1 Sample Generation and Calculation: This step determines the underwater vehicle's configuration, geometric scale, and operational parameter space. Multiple sample operational conditions are generated within this parameter space, and CFD calculations are performed on each. The operational parameters include at least speed and diving depth, and can be extended to angle of attack, heel angle, configuration parameters, and sea state parameters. This step provides high-fidelity training samples covering the parameter space for subsequent order reduction modeling and surrogate model training, ensuring the model has sufficient predictive power over a wide parameter range.
[0011] S2 Synchronous Extraction: From the CFD calculation results of each sample operating condition, complete physical field data of at least two types of heterogeneous hydrodynamic responses are extracted synchronously at a unified dimensionless time node or quasi-steady-state time window. Different responses under the same operating condition are assigned the same sample identifier and synchronization time index. Among them, the heterogeneous hydrodynamic response includes the surface wake field, drag history, and hull surface load. The surface wake field includes the surface wake height field or the wake control section velocity field; the drag history includes pressure drag, viscous drag, or total drag history; and the hull surface load includes the hull surface pressure distribution or wall shear stress distribution.
[0012] In this step, by unifying the time axis and synchronizing the time index, the time correspondence between different responses is ensured, which establishes a time consistency basis for subsequent shared latent variable modeling and physical closure constraints, and ensures that the wake, drag and surface load from the same moment can be correctly coupled in the shared latent variable space.
[0013] S3 Sample Physical Marker Extraction: Sample physical markers that move with the sample operating conditions are extracted from the complete physical field data of various heterogeneous hydrodynamic responses, resulting in surface wake field sample physical markers, drag history sample physical markers, and hull surface load sample physical markers. The surface wake field sample physical markers include at least one of the following: main wave crest ridge, main wave trough ridge, and centerline recovery point. The drag history sample physical markers include at least one of the following: initial response point, first extreme point, decay inflection point, and quasi-steady-state entry point. The hull surface load sample physical markers include at least one of the following: pressure peak trajectory, zero wall shear line or separation line, and bow and stern characteristic points.
[0014] In this step, key skeleton points / lines that best characterize the spatial location of physical features are selected from massive grid data, providing a small but sufficient number of anchor points for the subsequent construction of a reversible transport map. By accurately identifying the positional changes of these markers under different operating conditions, it becomes possible to separate the subsequent "positional migration" from the "morphological changes".
[0015] S4 Transport Alignment and Amplitude Separation: A reference operating condition is determined or a reference domain is generated from multiple samples, and reference physical markers are extracted from it. The reference operating condition is preferably a typical operating condition in the central region of the parameter space, and the extraction method of the reference physical markers is the same as that of the sample physical markers. Based on the correspondence between the sample physical markers of each sample and the reference physical markers, a reversible transport mapping from the reference domain to each sample domain is constructed; the reversible transport mapping satisfies the spatial reversibility condition. and time monotonicity condition This ensures that the mapping does not flatten or flip the mesh spatially, nor does it reverse the order of events temporally. In the formula, For mapping Jacobian matrix, This is the lower bound threshold of the Jacobian, which is a positive decimal. This indicates the local scaling factor of the time axis. Let be the lower bound threshold for time monotonicity, and be a positive decimal. The th... The sample at the th The transport mapping on the class response is obtained by minimizing the following objective function, specifically expressed as equation (1): (1), In the formula, the first term and These are the reference physical marker and the sample physical marker, respectively. Map the reference domain marker points The position obtained after deformation. This item indicates the mapping of the reference marker point. After the movement, the location should coincide as closely as possible with the actual sample marker. The smaller this value, the more accurately the mapping can "pry" the skeleton of the reference domain to the actual skeleton location of the sample.
[0016] Second item For mapping The gradient; This is the stiffness regularization coefficient, which controls the penalty intensity of this term; It is a unit moment. In this term, The closer This indicates that the mapping is closer to a pure translation. The smaller this term is, the more uniform the mapping deformation will be, and there will be no severe local stretching.
[0017] Third item For smoothing regularization, For mapping The Laplace operator describes the degree of local curvature of the map; This is the regularization coefficient, which controls the penalty strength of this term. This term prevents the mapping from becoming drastically distorted near a few key points, ensuring the continuity of the entire deformation field.
[0018] Using the aforementioned reversible transport mapping, the responses of each sample are pulled back to the reference domain to obtain the aligned response and transport parameters. The amplitude and bias of the aligned response are then separated, and the amplitude, bias, and transport parameters are stored separately. The aligned response is implemented according to equation (2). (2), In the formula, The pure morphological data of the i-th sample and the b-th type response are aligned and normalized on the reference domain coordinates; Reference domain coordinates; The original physical field data for the i-th sample and the b-th type of response; This is the response amplitude scaling factor; This is the bias value.
[0019] In this step, by matching the physical markers of the samples with those of the reference physical markers, the "spatial / temporal location migration of physical features" is completely separated from the "amplitude and shape changes of the response itself." The aligned response eliminates positional drift in a unified reference coordinate system, preventing positional migration from being mistaken for shape changes during subsequent order reduction, thus reducing the number of modes and maintaining local peak accuracy. The separate storage of amplitude, bias, and transport parameters allows for the prediction of position, size, and shape separately during subsequent modeling, reducing the complexity of each sub-problem.
[0020] S5 Discrete Physical Closure Operator Construction: On a unified reference surface grid and synchronous time nodes, a discrete physical closure operator is constructed between responses with physical equality relationships; the discrete physical closure operator is used to convert hull surface pressure and wall shear stress into integral drag, and to calculate the closure residual between the direct drag and the integral drag.
[0021] Specifically, the discrete physics closure operator obtains the integral drag by discretizing and summing the axial pressure and wall shear stress of each surface element on the same reference surface mesh; the closure residual is the difference between the direct drag and the integral drag. The hull surface is divided into... The axial pressure and wall shear stress of each surface element are discretized and summed on the same reference surface mesh to obtain the ... Integral resistance at each time point As shown in equation (3): (3), In the formula, For the first A unified dimensionless time node; For the first Each surface unit in Local pressure value at any given time; For far-field static pressure; This refers to the local surface pressure, which is the pressure difference relative to the far-field static pressure. For the first The external normal vector of each surface element is in The component of direction (direction of incoming flow / axial direction of hull); For the first Each surface unit in The wall shear stress at time t is The component of direction (direction of incoming flow / axial direction of hull); For the first The area of each surface unit.
[0022] The closed-loop residual between the direct resistance and the integral resistance can be calculated using equation (3). As shown in equation (4), (4), In the formula, This represents the total drag directly output in CFD calculations.
[0023] Under a unified discrete framework, a physical equation for "distributed load → concentrated force" is established. This equation serves as a physical closure penalty term in the joint order reduction stage and as an equation constraint for the constraint projection stage. By pre-calculating the closure residual for each sample, subsequent latent variable order reduction and surrogate model prediction can be performed within a physically closed framework, avoiding the physical contradiction of "integral resistance and direct resistance being inconsistent" from the data source.
[0024] Preferably, when the physical closure relationship also includes the control volume momentum relationship determined by the velocity and pressure of the wake control section, the difference between the surface load integral resistance, the control volume momentum deficit equivalent resistance, and the direct resistance together constitutes a multipath closure constraint.
[0025] II. Model Building and Training Phase: S6 Joint Order Reduction: The aligned response in step S4 is subjected to joint order reduction using shared and private latent variables. The overall flow state of the sample is encoded as a shared latent variable used by multiple types of responses, and the unique residual changes of each type of response are encoded as private latent variables corresponding to each response category, so that multiple types of responses under the same operating condition are coupled through shared latent variables.
[0026] Alignment of responses by response category Represented as equation (5): (5), In the formula, The mean of the reference domain; For the first A reduced-order basis for class responses; To share latent variables to the first The mapping matrix of class responses; Shared latent variables used by multiple types of responses; These are private latent variables. The shared latent variables and the private latent variables satisfy the orthogonality condition. This ensures that the shared and private parts do not overlap and that information is not redundant.
[0027] In this step, by sharing latent variables, the three types of responses share the same set of low-dimensional codes, explicitly modeling the physical fact of "the influence of the same flow state on different responses," thus avoiding redundant learning and parameter redundancy among multiple independent proxy models. Simultaneously, by using private latent variables, the unique details of each type of response are preserved, preventing the sharing process from "averaging out" the personalized characteristics of each response. Orthogonal constraints ensure that the two information sources are cleanly separated and do not interfere with each other.
[0028] S7 Physical Closure Embedding: Based on the discrete physical closure operator in step S5, a physical closure constraint is constructed in the latent space. This physical closure constraint requires that the drag obtained by the integral operator after reconstruction from the latent surface load variables is numerically equal to the directly predicted drag. Specifically, the latent variable representations from step S6 are substituted into the discrete integral operator in step S5 to construct the latent space closure matrix. ,make ,in To compile a large vector of shared latent variables, private latent variables, transport parameters, magnitude parameters, and closure residual coefficients, These are known boundary terms.
[0029] The physical closure constraint is added as an additional objective term to the joint order reduction objective function, so that the order reduction process can pursue reconstruction accuracy while taking into account physical consistency. The joint order reduction objective function is shown in equation (6).
[0030] The objective function consists of two terms: the first is the weighted reconstruction error, and the second is the physical closure error. These two terms are expressed as follows: To achieve a balance. For the reduced-order basis matrix of the b-th type response, the low-dimensional latent variables are mapped back to the original data space; The mapping matrix from the shared latent variables to the b-th type of response space; Let be the shared latent variable for the i-th sample, describing the overall flow state under this working condition; For the i-th sample and the b-th type of response, there is a private latent variable that describes the residual changes unique to this type of response; For the reconstruction error weighting coefficient of the b-th type response; The pure morphological data of the i-th sample and the b-th class response after alignment and standardization; The mean shape of the b-th type response across all training samples; The weighting coefficients are the weights of the physical closure constraints.
[0031] In this step, the physical equation established in S5 is "lifted" from the physical space to the latent space, giving the order reduction process "physical awareness" while compressing the data. Traditional methods do not consider physical constraints during order reduction, resulting in physically infeasible reconstruction results. This step embeds a physical closure penalty term into the order reduction objective, ensuring that the solution of the order reduction basis and latent variables is performed within the physically feasible domain from the beginning. Physical consistency is satisfied upon completion of the order reduction, eliminating the need for subsequent corrections.
[0032] S8 surrogate training: Using sample operating parameters as input and shared latent variables, private latent variables, transport parameters, magnitude, and bias as joint outputs, a probabilistic surrogate model is trained. Preferably, the probabilistic surrogate model is a multi-output Gaussian process, a sparse variational Gaussian process, a deep kernel multi-task Gaussian process, or a neural network with a probabilistic output head that can output joint predicted mean and joint covariance.
[0033] In this step, a probabilistic surrogate model is used instead of a deterministic model to obtain information on the uncertainty of the prediction, providing a quantitative basis for the subsequent covariance-weighted constrained projection. The joint covariance matrix fully characterizes the correlation structure between latent variables, enabling the constrained projection to distinguish between "deterministic directions" and "uncertain directions," thus achieving accurate correction.
[0034] III. The model application phase includes: S9 Constrained Projection: Input the parameters of the operating condition to be predicted into the probabilistic surrogate model to obtain the predicted values of the original latent variables. and joint covariance The predicted values of the original latent variables are calculated based on the joint covariance. Projecting onto the feasible set that satisfies the physical closure constraint yields the corrected latent variables. .
[0035] Specifically, the constrained projection problem is formulated as finding the point that is closest to the original predicted value in the covariance-weighted sense, under the premise of satisfying physical closure constraints, as shown in equation (7). (7), In the formula, Let be the vector of latent variables (optimization variables) to be solved. The optimal solution obtained by searching within the feasible region is . ; The original latent variable predicted mean (uncorrected point estimate) is the output of the probabilistic surrogate model. It is the inverse of the joint covariance matrix; It is a closed matrix in the latent space; Given boundary terms; Inequality constraints to ensure the invertibility of the mapping.
[0036] When the closed relation is a linear equation, the correction latent variable can be directly calculated as equation (8): (8), In the formula, The gain matrix that backpropagates the error in the closed residual space to the latent variable space determines how much correction each latent variable should bear. It is the inverse of the covariance matrix in the closed residual space, used to normalize the scale differences between different physical constraints; The physical closure deviation term of the original predicted value indicates the degree to which the current predicted value does not satisfy the physical closure equation—the larger the deviation, the larger the required correction.
[0037] The correction amount of the original latent variable prediction value is allocated according to the joint covariance, wherein the direction with larger joint covariance bears a larger correction amount and the direction with smaller joint covariance bears a smaller correction amount, so that the corrected latent variable retains the original prediction information of the probabilistic surrogate model to the greatest extent while satisfying the physical closure constraint.
[0038] After the surrogate model completes its predictions, the predicted values are "pulled" into the physical feasible region through covariance-weighted projection. Compared to adding physical constraints as regularization terms to the loss function (soft constraints), this step uses hard constraint projection to ensure that the prediction results satisfy the physical closure equation within a preset numerical tolerance.
[0039] S10 Inverse Reconstruction: Based on the correction latent variables, the aligned response is reconstructed in the reference domain, and restored to physical coordinates through the inverse mapping of the reversible transport mapping in S4, thus obtaining multiple heterogeneous hydrodynamic responses that mutually satisfy the physical closure relationship under the predicted operating condition. The final reconstruction of the class response is represented by equation (9). (9), In the formula, For the first Class response in real physical coordinates The final predicted value; For the first The bias of the response class (stored in S4) is used to recover the baseline subtracted during normalization; For the first The amplitude scaling factor of the class response (saved in S4) is used to restore the amplitude size that was reduced during normalization; No. The mean shape of the class response in the reference domain, i.e., the average skeleton after all training samples are aligned; For the first The reduced-order basis matrix of the class response (obtained by training in S6) expands the low-dimensional latent variables into a complete data form; To share latent variables to the first The mapping matrix of the class response space (obtained from training in S6); The corrected shared latent variables (from the constraint projection results of S9) describe the overall flow state of the working condition to be predicted. For the corrected first Private latent variables of the class response (from the constraint projection results of S9) describe the unique details of the class response; For the first The inverse mapping of the response reversible transport mapping (constructed in S4) converts physical coordinates... Map back to the reference domain coordinates ξ.
[0040] In this step, the corrected latent variables are restored layer by layer to physically understandable response data. This process is the complete inverse of S4 transport alignment and amplitude separation—reassembling the three sets of information (position, size, and shape) into a complete physical response while ensuring that the physical consistency imparted by the constraint projection of S9 is preserved in the final output.
[0041] Preferably, the method further includes a model reliability verification step, which constructs a reliability index using the prediction covariance, latent variable projection distance, physical closure residual, and minimum Jacobian of the reversible transport mapping of the probabilistic surrogate model. When the reliability index exceeds a threshold, the current probabilistic surrogate model is determined to be unreliable, and the operating condition to be predicted is added to a supplementary high-fidelity simulation sample set and the probabilistic surrogate model is updated.
[0042] Another aspect of the present invention discloses a multi-response transport alignment and physical closure prediction system for underwater vehicles, which utilizes the above-mentioned prediction method and includes: The synchronous response acquisition module is used to acquire at least two types of heterogeneous hydrodynamic responses synchronized under multiple sample operating conditions. The physical marker and transport alignment module is used to extract physical markers that move with the sample operating conditions for each type of heterogeneous hydrodynamic response, and construct a reversible transport mapping based on the correspondence between the physical markers and the reference physical markers, pull each sample response back to the reference domain and separate the amplitude and bias. The Physical Closure Operator module is used to construct discrete physical closure operators between responses that have physical equality relationships; The shared-private latent variable joint order reduction module is used to perform joint order reduction on the aligned response, decompose it into shared latent variables and private latent variables corresponding to each response category, and construct physical closure constraints in the latent space according to the discrete physical closure operator. The joint surrogate prediction module is used to train a probabilistic surrogate model with sample operating condition parameters as input and the shared latent variables, private latent variables, transport parameters, magnitude and bias as joint outputs. The constraint projection module is used to project the original latent variable prediction values and their joint covariance to a feasible set that satisfies the physical closure constraint, based on the original latent variable prediction values and their joint covariance of the working condition to be predicted. The inverse transport reconstruction module is used to reconstruct the aligned response in the reference domain and restore it to the physical coordinates through the inverse mapping of the reversible transport mapping, so as to obtain multiple heterogeneous hydrodynamic responses that satisfy the physical closure relationship under the predicted working condition.
[0043] The beneficial effects of this invention are as follows: 1. Effectively solves the problem of order reduction distortion caused by physical feature migration. This invention introduces physical marker-driven reversible transport alignment before order reduction, separating the spatial / temporal position migration of physical features from the amplitude and morphological changes of the response itself. The aligned response eliminates position drift in a unified reference coordinate system, preventing the subsequent order reduction from misinterpreting positional changes such as wave peak movement and pressure peak movement as morphological changes, thereby reducing the number of required modes and maintaining the accuracy of local high-gradient structures. Compared with the traditional fixed-coordinate POD method, both reconstruction accuracy and physical marker position accuracy are significantly improved under conditions of significant feature migration, such as shallow diving depth and high speed.
[0044] 2. Achieving physical consistency across responses. This invention achieves order reduction through a joint approach of shared and private latent variables, allowing the common variations of the same flow state in the wake field, drag history, and surface load to be expressed by the same low-dimensional variable. This avoids redundant parameters and contradictory prediction results among multiple independent surrogate models. Simultaneously, by constructing hard constraints in the latent space using discrete physical closure operators, and correcting the predicted values to the physically feasible region during the online prediction stage through covariance-weighted projection, it ensures that the final output drag and the surface load integral drag are equal within a preset tolerance, fundamentally solving the problem of physical inconsistencies caused by conflicting predictions.
[0045] 3. Balancing Physical Feasibility and Data Fidelity. This invention employs covariance-weighted constraint projection, utilizing the prediction uncertainty information provided by the probabilistic surrogate model to perform distributive correction on physical closure deviations—directions with higher uncertainty bear a larger correction amount, while directions with lower uncertainty bear a smaller correction amount. Compared to direct least-squares projection or soft constraint methods that only use it as a regularization term in the loss function, the projection results of this invention satisfy both physical closure constraints and retain, to the greatest extent possible, the patterns learned by the surrogate model from high-fidelity data, achieving an optimal balance between physical feasibility and data fidelity.
[0046] 4. Possesses adaptive evolution capability. This invention constructs a reliability index that comprehensively considers prediction covariance, latent variable projection distance, physical closure residuals, and minimum Jacobian transport mapping, enabling accurate differentiation between statistical extrapolation risk and physical infeasibility risk. When reliability is insufficient, it automatically triggers supplementary CFD sampling and model update mechanisms, allowing the system to continuously evolve during use and continuously improve long-term prediction accuracy. Attached Figure Description
[0047] Figure 1 This is an overall flowchart of the method of the present invention; Figure 2 A schematic diagram of the heterogeneous hydrodynamic response and synchronization relationship of an underwater vehicle near the free surface; Figure 3 A schematic diagram illustrating the extraction of physical markers for wake peaks, surface pressure peaks, and drag event points; Figure 4 A schematic diagram of reversible transport alignment driven by physical markers; Figure 5 A schematic diagram illustrating the physical closed loop relationship between the surface load of the hull and the resistance obtained through a discrete integration operator; Figure 6 A schematic diagram of the joint order reduction structure of shared-private latent variables; Figure 7 A schematic diagram of covariance weighted constraint projection; Figure 8 A schematic diagram of reverse transport reconfiguration and reliability assessment under the new operating conditions; Figure 9 This is a structural diagram of the system of the present invention. Detailed Implementation
[0048] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.
[0049] Example 1: The overall prediction process of the present invention is as follows Figure 1 As shown, the model comprises three stages: data acquisition and preprocessing, model building and training, and model application. The following section uses an underwater vehicle as an example to detail the specific implementation methods for each stage.
[0050] S1. Sample Generation and CFD Calculation: The underwater vehicle in this embodiment is a publicly disclosed standard submarine model, with a total length of 4.356m, a maximum hull diameter of 0.508m, and a length-to-slenderness ratio of 8.57. The model includes the hull, conning tower, and four rudders, as shown below. Figure 2 As shown.
[0051] Determine the operating parameter space: the speed V ranges from 3 to 10 knots, and the dimensionless diving depth ranges from 1.0 to 5.0. The operating parameters can also be extended to angle of attack (-5° to 5°), roll angle (-10° to 10°), etc.
[0052] Training and test cases are generated within the parameter space using maximin Latin Hypercube Sampling, Sobol sequences, or adaptive sampling. Latin Hypercube Sampling is a space-filling design method that can uniformly cover the entire parameter space with fewer sample points, avoiding sample point clustering or leaving large uncovered areas. The maximin criterion (maximizing minimum distance) further ensures that the minimum distance between sample points is maximized, making the sample points as uniformly distributed as possible in the parameter space. 60-100 training sample cases and 15-20 test sample cases are generated. The test cases are not involved in the subsequent establishment of reference labels, transport mappings, reduced basis models, and surrogate models.
[0053] High-fidelity near-free surface two-phase flow CFD calculations were performed on each sample operating condition. A solver based on the Reynolds-averaged Navier-Stokes equations (RANS) was used, and the k-ω SST model was selected as the turbulence model. The free surface was captured using a volume of fluid (VOF) two-phase flow model. The computational domain size was set as follows: 3L upstream from the bow, 5L downstream from the stern, 2L on each side and bottom from the hull centerline, with the top being the free surface. Boundary conditions were set as follows: velocity inlet, pressure outlet, VOF two-phase flow model for capturing the free surface, and no-slip walls on the hull surface. A hybrid structured and unstructured mesh was used, with localized refinement applied to the hull boundary layer and the vicinity of the free surface to ensure accurate measurement of the near-free surface. + The number of grid cells is approximately 30-100, with a total grid size of about 8-15 million. CFD calculations for each sample condition yield the free surface height field, pressure drag, viscous drag, total drag history, and the pressure distribution and wall shear stress distribution on the hull surface. All samples are divided into training and test sets according to the operating conditions. Test conditions are not involved in the establishment of reference flags, transport mappings, reduced-order bases, and surrogate models.
[0054] S2, Synchronous Extraction like Figure 2As shown, when an underwater vehicle navigates near the free surface, it simultaneously induces three types of heterogeneous hydrodynamic responses: the free surface wake field (including wave system morphology and height changes), the hull surface load distribution (including pressure and wall shear stress), and the drag time history (including transient peak and steady-state values). At a unified dimensionless time node (e.g., 0, 0.1, 0.2, ..., 1.0), complete physical field data for these three types of heterogeneous hydrodynamic responses are synchronously extracted from the CFD calculation results of each sample condition: the free surface height field (full-field grid data at each time node), the drag history (values of pressure drag, viscous drag, and total drag at each time node), and the hull surface pressure distribution and wall shear stress distribution (values of each surface grid cell at each time node). The three types of responses under the same condition are assigned the same sample identifier and synchronous time index.
[0055] S3. Extraction of physical markers from samples: like Figure 3 As shown, physical markers that move with the sample conditions are extracted from the complete physical field data of various responses.
[0056] For the wake field, the main wave crests and troughs are extracted from the gradient of the free surface height field and the eigenvalues of the Hessian matrix. Specifically, the wave crests correspond to the lines connecting local maxima in the height field, and the wave troughs correspond to the lines connecting local minima. The centerline recovery point is the position where the free surface height is first restored to the still water surface along the submarine's centerline direction.
[0057] For the resistance process, from the resistance time series Identifying extreme points by the zeros of the first derivative ( ), determined by the sliding variance threshold Determine the quasi-steady-state entry time, where The initial response point is determined based on the steady-state variance statistics of the training samples. The initial response point is when the resistance first exceeds [a certain threshold]. The moment ( (Take 1% of the steady-state resistance), the inflection point of the attenuation is the rate of decrease in resistance. The moment when the value first approaches zero from a negative value.
[0058] For the hull surface load, it is determined by the axial pressure distribution on the hull surface. Extreme point tracking yields the pressure peak trajectory (tracked along the time axis). The spatial variation of the maximum value is due to the wall shear stress. The zero-crossing point is used to identify the separation line position, and the bow and stern feature points are fixed reference points at the bow and stern of the boat.
[0059] S4, Transport Alignment and Amplitude Separation Select a typical operating condition (e.g., speed V=6 knots, diving depth h / D=3.0) in the central region of the parameter space as the reference operating condition, and extract its physical characteristics as reference physical characteristics. If the operating condition distribution of the training samples is not uniform, a weighted average of the physical characteristics of all training samples can be taken to generate a comprehensive reference domain.
[0060] like Figure 4 As shown, a reversible transport mapping from the reference domain to each sample domain is constructed based on the correspondence between the physical identifiers of each sample and the physical identifiers of the reference. The sample at the th The transport mapping on the class response is obtained by minimizing the following objective function, specifically expressed as equation (1): (1).
[0061] This embodiment uses a radial basis function (RBF) mapping with positive Jacobian constraints to achieve spatial transport alignment. The optimization problem in equation (1) is solved numerically by expanding the radial basis function, specifically using a thin-plate spline kernel to construct the mapping in the following form. (Referring to physical markers...) Source point, sample physical marker point For the target point, the mapping is constructed as shown in equation (10). (10) In the formula, For thin plate spline kernel functions, This refers to the number of control points (i.e., the number of physical markers). , , The coefficients are undetermined. This is achieved by solving a system of linear equations and applying the spatial invertibility condition. Obtained by constraints, The value is set to 0.01. The transport mapping of the time series employs piecewise cubic Hermite interpolation with monotonic constraints to ensure the time monotonicity condition. , Take 0.01.
[0062] Further explanation of equation (10): ξj is the coordinate of the j-th control point in the reference domain; thin plate spline kernel (r) = r²logr is applicable to two-dimensional parametric coordinates. To ensure the uniqueness of the mapping solution, the coefficients also satisfy Σjαj=0 and ΣjαjξjT=0. For three-dimensional physical coordinates, a multiharmonic spline kernel adapted to the spatial dimension is used. The example values for the spatial invertibility threshold δ and the temporal monotonicity threshold δt can both be 0.01, and are determined by combining the validation set error, mesh quality, and time node interval, respectively.
[0063] The reversible transport mapping is used to pull the responses of each sample back to the reference domain, yielding the aligned response and transport parameters. The aligned response is then further separated into amplitude and bias according to equation (2): (2). Among them, for the water surface wake height field, the amplitude Take the average height difference between the crests and troughs, and offset. The average height across the entire field (i.e., the height of the still water surface) is taken; for the drag history, the amplitude is the difference between the peak drag value and the steady-state drag value, and the offset is the steady-state drag value; for the surface pressure distribution, the amplitude is the range of the pressure coefficients, and the offset is the far-field pressure. Save the amplitude, bias, and transport parameters separately.
[0064] S5. Construction of Discrete Physical Closed Operators Discrete physical closed-loop operators are constructed on a unified reference surface mesh (using the same hull surface mesh topology as the reference operating condition, with the mesh number determined through mesh independence verification; in this embodiment, this can be 50,000-100,000 surface elements) and synchronous time nodes, such as... Figure 5 As shown. The surface of the hull is divided into... The tiny facet element, in the... The integral resistance is calculated as shown in equation (3) at each time point: (3), In the formula, For the first A unified dimensionless time node; For the first Each surface unit in Local pressure value at any given time; For far-field static pressure; This refers to the local surface pressure, which is the pressure difference relative to the far-field static pressure. For the first The external normal vector of each surface element is in The component of direction (direction of incoming flow / axial direction of hull); For the first Each surface unit in The wall shear stress at time t is The component of direction (direction of incoming flow / axial direction of hull); For the first The area of each surface unit.
[0065] The closed-loop residual between the direct resistance and the integral resistance can be calculated using equation (3). As shown in equation (4), (4).
[0066] This closed residual serves as one of the output targets of the subsequent surrogate model. Through this discrete integral relationship, a clear physical equation is established between the hull surface loads (pressure and wall shear stress) and the drag, ensuring the physical consistency between the two in the subsequent order reduction and prediction processes.
[0067] S6, Joint Order Reduction Perform a joint reduction of shared and private latent variables on the aligned response obtained from S4, the structure of which is as follows: Figure 6 As shown. The aligned response representation for each response category is shown in equation (5). (5). Among them, shared latent variables The dimensions are determined jointly based on the total reconstruction error and physical closure error on the validation set. In this embodiment, a grid search is performed by reserving a validation set, and the dimensions of the shared latent variables are 3-8 (adjusted according to the number of training samples and parameter space complexity), while the dimensions of the private latent variables are... The dimension can be 1-3. Reduced-order basis. The shared mapping matrix is obtained by performing eigenvalue decomposition on the covariance matrix of various response alignment data (i.e., the constrained POD method). and shared latent variables By simultaneously optimizing reconstruction error and orthogonality constraints The orthogonality constraint ensures that the information described by the shared latent variables and the private latent variables does not overlap or become redundant, resulting in a clean and clear decomposition of reduced order.
[0068] like Figure 6 As shown, the three types of responses are coupled through sharing latent variables, while each retains its own private latent variables to describe its unique patterns of change.
[0069] S7, Physically Closed Embedded Substituting the latent variables in S6 into the discrete integral operator in S5, we construct a closed latent space matrix. Specifically, It is assembled from three parts: (1) a surface integral operator (which converts surface pressure and shear stress into integral drag); and (2) a branched reduced-order basis. (Expand latent variables into physical quantities); (3) Time resampling operator (unify the time nodes of each sample to the reference time axis). The overall objective function for joint order reduction is shown in equation (6):
[0070] Among them, the weighting coefficient Adjustments are made based on the magnitude of various responses (e.g., the magnitude of the wake height field is...). m, the magnitude of the resistance is N needs to be determined by the weighting coefficients. To make the reconstruction errors of the two types of responses similar in magnitude, by taking... , (The specific amount will be adjusted based on the actual data volume). The value of is determined through a grid search on the validation set, making the physical closure error and reconstruction error numerically comparable. Typically, is taken as . - Magnitude.
[0071] S8, Agent Training Using sample operating parameters (speed, dimensionless depth, and selectable angle of attack, etc.) as input, and sharing latent variables... Private latent variables in each branch Transport parameters (spatial / temporal offset of each physical marker point), amplitude and bias For joint output, a probabilistic surrogate model is trained. This embodiment employs a multi-output Gaussian process, and the covariance function uses an Automatic Relevance Determination (ARD) squared exponential kernel. As shown in equation (11): (11), In the formula, This is a vector of operating condition parameters; The variance of the signal; For the first The feature length scale of each operating condition parameter (learned from the training data through maximum likelihood estimation). For noise variance; This refers to the Kronecker delta function. The ARD kernel can automatically identify the degree of influence of various operating parameters on the output—if the characteristic length scale of a certain parameter... A relatively large value indicates that the parameter has a relatively small impact on the prediction results.
[0072] Equation (11) gives the input correlation kernel of the multi-output Gaussian process; for the multi-output case, the noiseless input kernel in Equation (11) is combined with the output correlation matrix to form the cross covariance between different outputs, and noise terms are set according to the outputs to obtain the joint prediction covariance Σ.
[0073] The kernel function described above defines how the similarity between any two points in the operating condition parameter space is mapped to the correlation in the output latent variable space. The hyperparameters of the kernel function ( , , The hyperparameters are automatically learned from the training data by maximizing the log-marginal likelihood. Once determined, for any operating condition parameter to be predicted, the Gaussian process model directly provides the joint prediction mean using the standard conditional distribution formula. and joint covariance The analytical expression requires no additional model training or iterative optimization. Therefore, the probabilistic surrogate model in this embodiment is an end-to-end prediction model—taking operating condition parameters as input, it directly outputs the complete probability distribution of latent variables.
[0074] Training a multi-output Gaussian process is achieved by maximizing the log-marginal likelihood, with the output being the joint prediction mean. and joint covariance When the number of training samples is large (e.g., more than 200), a sparse variational Gaussian process can be used to reduce computational complexity, with 30-50 induced points.
[0075] In this embodiment, the training of the surrogate model is determined to be convergent based on the following criteria: For the hyperparameters of a Gaussian process ( , , This is optimized by maximizing the logarithmic marginal likelihood. When the relative increment of the logarithmic marginal likelihood in a series of consecutive iterations is lower than a preset threshold (e.g., 10), the optimization is achieved. -3 When the hyperparameter training is converged, it is determined that the training has converged.
[0076] It should be noted that the training in this invention includes joint order reduction parameters ( , , , The optimization of the Gaussian process hyperparameters involves two parts: the reduced basis and latent variables, which are trained sequentially. First, the reduced basis and latent variables are obtained through the joint reduced objective function in S6 and S7. Then, the hyperparameters are obtained through training the surrogate model in S8. The surrogate training in S8 is based on the reduced parameters from S6-S7 and does not require iterative iteration between the two stages. Once the reduced parameters are determined, the optimization of the Gaussian process hyperparameters is performed independently; the two optimizations are not performed alternately.
[0077] Once the above convergence conditions are met, the probabilistic surrogate model is considered to have been trained and can proceed to the next stage of model application.
[0078] S9, Constrained Projection Input the parameters of the operating conditions to be predicted (such as speed V=8 knots, diving depth h / D=2.0) into the trained probabilistic surrogate model to obtain the predicted mean of the original latent variables. and joint covariance .
[0079] Based on the joint covariance, the predicted values of the original latent variables are projected onto a feasible set that satisfies the physical closure constraints, such as... Figure 7 As shown. The constrained projection problem is given by equation (7): (7), Since the closed-loop relationship is a linear equation, the correction latent variable can be directly calculated as equation (8): (8).
[0080] This projection assigns a larger closure correction to the direction of the latent variable with higher uncertainty, and a smaller correction to the direction with lower uncertainty. Figure 7 As shown. If the linear projection still violates... The mapping invertibility condition (e.g., the predicted transport parameters cause the Jacobian in a local region to fall below a threshold) then maintains... At the same time, additional corrections are made through quadratic programming with inequality constraints, so that the final solution satisfies both types of constraints.
[0081] S10, Reverse Reconstruction like Figure 8 As shown, the aligned response is reconstructed in the reference domain based on the corrected latent variables, and then restored to the physical coordinates through the inverse mapping of the reversible transport map, as shown in equation (9): (9). Finally, the free surface wake field, drag history and hull surface load distribution under the predicted working condition are obtained, and the three satisfy the physical closed relationship between each other.
[0082] Verification and Comparison: Four sets of control experiments were set up using the same training samples, validation samples, and reserved test conditions to verify the effectiveness of each core component of the invention: Option 1: Fixed-coordinate independent POD order reduction + Gaussian process regression surrogate model (representing existing technology); Option 2: Based on Option 1, only transport alignment is added (excluding physical closure constraints); Option 3: Based on Option 1, only add physical closure constraint projection (excluding transport alignment); Option 4: The complete technical solution of this invention is transport alignment + shared-private latent variable joint order reduction + physical closure constraint projection.
[0083] Evaluation metrics include various response relative errors, physical marker position errors, resistance integral closure errors, minimum Jacobian, latent variable projection distance, and single-condition prediction time.
[0084] The above indicators were calculated on the same hold-out test set, and the mean, standard deviation, and maximum value of each indicator were recorded to evaluate the impact of different schemes on the ability to maintain mobility characteristics, cross-response consistency, and prediction efficiency. The actual comparative values were based on the results of each scheme under the same test conditions; no quantitative conclusions on superiority or inferiority were drawn before the test results were obtained.
[0085] Model adaptive update mechanism It should be noted that although the probabilistic proxy model described in S8 has been trained, the model of this invention is not limited to "training once and using permanently". The operating parameter space of an underwater vehicle is usually a high-dimensional continuous space (speed, diving depth, angle of attack, heel angle, configuration parameters, sea state parameters, etc.). The number of initial training samples is limited by the computational cost of CFD, usually 60 to 100, which can only cover a limited area in the parameter space, leaving large unexplored areas. When the operating condition to be predicted falls in these uncovered areas, the model inevitably faces the risk of statistical extrapolation.
[0086] To this end, the present invention introduces a reliability-driven adaptive update mechanism in the model application stage. Its core logic is: the model continuously monitors the reliability of its own predictions during use. When the reliability is insufficient, it actively requests the addition of high-fidelity samples and incorporates the new samples into the training set for incremental updates, so that the coverage of the model gradually expands and the prediction accuracy continuously improves as it is used.
[0087] The reliability index R is obtained by normalizing and weighting the predictive covariance, latent variable projected distance, physical closure residual, and minimum Jacobian of the invertible transport mapping of the probabilistic surrogate model. When the reliability index R exceeds the threshold R... th If the reliability of the current probabilistic proxy model is deemed insufficient, the working condition to be predicted is added to the supplementary high-fidelity simulation sample set and the probabilistic proxy model is updated.
[0088] It should be noted that after new samples are added, an incremental update strategy is used to update the model, rather than retraining it from scratch, in order to reduce update costs. 1. Reduced order basis Keep them unchanged, only encode the latent variables of the new samples. , Add to the set of latent variables; 2. The hyperparameters of the Gaussian process surrogate model are initialized with the current value and optimized using a small number of new samples (usually converged in 5-10 gradient steps), without having to maximize the log marginal likelihood from scratch. 3. If the cumulative number of new samples reaches a certain amount (e.g., exceeding 20% of the initial training sample count), a complete model retraining will be triggered to ensure the global optimality of the model structure. 4. After the update is completed, the updated model replaces the original model for subsequent predictions. When this situation occurs again, the model will already have knowledge of the area, significantly improving prediction reliability.
[0089] Through the above mechanism, this invention forms a complete closed loop of "prediction-reliability assessment-sampling-update-re-prediction". As the model is used more times, the parameter space it covers gradually expands, and its prediction accuracy and reliability continuously improve, ultimately achieving effective coverage of the entire target parameter space with limited CFD computational cost. This adaptive evolutionary capability is not possessed by traditional "train once, use forever" surrogate models.
[0090] Extended Applicability Notes This implementation is not limited to the DARPA SUBOFF AFF-8 configuration, but is also applicable to other underwater vehicles, including but not limited to: submersibles, autonomous underwater vehicles (AUVs), torpedoes, unmanned underwater vehicles (UUVs), and other near-free surface vehicles.
[0091] When the sample also includes wake control section velocity and pressure data, a control volume momentum operator can be further constructed to make the surface load integral drag, wake momentum deficit equivalent drag and direct drag form a multi-path physical closure, further enhancing the physical consistency of the prediction.
[0092] Example 2: In another embodiment, the reversible transport mapping uses Large Deformation Diffeomorphic Metric Mapping (LDDMM) instead of radial basis function mapping. This method solves for the time-varying velocity field. Ordinary differential equations A smooth, invertible mapping from the reference domain to the sample domain is generated. This mapping is obtained by minimizing the weighted sum of the regularization energy of the velocity field and the matching error of the physical markers. The regularization energy term ensures the smoothness and invertibility of the mapping, while the matching error term ensures accurate alignment between the physical markers in the reference domain and the physical markers in the sample domain. This method theoretically guarantees that the Jacobian determinant of the mapping is strictly greater than zero, meaning the mapping is always invertible and will not produce local flips or flattening. It is particularly suitable for scenarios where the distribution of physical markers is sparse or the deformation between samples is large.
[0093] Example 3: In another embodiment, when the sample data also includes the velocity and pressure distribution at the wake control section, the physical closure relationship is extended to multi-path closure. Specifically, at the downstream wake control section of the hull (e.g., a cross-section 0.5L from the stern), the equivalent drag is obtained through momentum integration. This equivalent drag, along with the drag obtained from surface load integration in step S4 and the drag directly output by the CFD, together constitute the multi-path physical closure constraint. When constructing the closure matrix in the latent space, both the surface integration path and the control volume momentum integration path are incorporated, ensuring that the prediction results simultaneously satisfy the consistency verification of the two physical paths. Multi-path closure can form a "three-way redundancy verification" mechanism—if one of the three drag sources deviates significantly from the other two, the model will automatically identify and coordinate the correction during constraint projection, thereby further improving the robustness and reliability of the prediction results.
[0094] Example 4: The present invention also provides an electronic device, including a processor and a memory, wherein the memory stores a computer program, and the computer program, when executed by the processor, implements the underwater vehicle multi-response transport alignment and physical closure prediction method described in any of the above embodiments. The electronic device may be a general-purpose computer, workstation, server, embedded system, or other programmable data processing device.
[0095] Example 5: The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the method described in any of the above embodiments. The storage medium may be ROM, RAM, hard disk, optical disk, flash memory, or other non-transitory computer-readable storage media of any form.
[0096] This invention also provides a multi-response transport alignment and physical closure prediction system for underwater vehicles, such as... Figure 9 As shown, it includes the following functional modules: The modules are connected sequentially according to the data processing direction: the synchronous response acquisition module outputs the synchronous hydrodynamic field to the physical marker and transport alignment module, and outputs the load data therein to the physical closure operator module; the aligned response, transport parameters, amplitude, and bias output by the physical marker and transport alignment module, as well as the discrete physical closure operator and closure residual output by the physical closure operator module, are jointly input into the shared-private latent variable joint order reduction module; the order reduction basis and latent variables output by this module are used to train the joint surrogate prediction module; the original latent variable prediction values and joint covariance output by the joint surrogate prediction module are input into the constraint projection module, and the obtained corrected latent variables, together with the transport parameters, amplitude, and bias, are input into the inverse transport reconstruction module to output multiple types of hydrodynamic responses.
[0097] The synchronous response acquisition module is used to acquire at least two types of heterogeneous hydrodynamic responses synchronized under multiple sample operating conditions. The physical marker and transport alignment module is used to extract physical markers that move with the sample operating conditions for each type of heterogeneous hydrodynamic response, and construct a reversible transport mapping based on the correspondence between the physical markers and the reference physical markers, pull each sample response back to the reference domain and separate the amplitude and bias. The Physical Closure Operator module is used to construct discrete physical closure operators between responses that have physical equality relationships; The shared-private latent variable joint order reduction module is used to perform joint order reduction on the aligned response, decompose it into shared latent variables and private latent variables corresponding to each response category, and construct physical closure constraints in the latent space according to the discrete physical closure operator. The joint surrogate prediction module is used to train a probabilistic surrogate model with sample operating condition parameters as input and the shared latent variables, private latent variables, transport parameters, magnitude and bias as joint outputs. The constraint projection module is used to project the original latent variable prediction values and their joint covariance to a feasible set that satisfies the physical closure constraint, based on the original latent variable prediction values and their joint covariance of the working condition to be predicted. The inverse transport reconstruction module is used to reconstruct the aligned response in the reference domain and restore it to the physical coordinates through the inverse mapping of the reversible transport mapping, so as to obtain multiple heterogeneous hydrodynamic responses that satisfy the physical closure relationship under the predicted working condition.
[0098] Those skilled in the art should understand that the functions of the above modules correspond to Figure 1 The corresponding steps in the method flow shown can be implemented by software, hardware, or a combination of both, and the specific deployment method does not constitute a limitation on the present invention.
[0099] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A method for predicting multi-response transport alignment and physical closure of underwater vehicles, characterized in that: This includes the data acquisition and preprocessing stage, the model building and training stage, and the model application stage. The data acquisition and preprocessing stage includes: S1 Sample Generation and Calculation: Determine the underwater vehicle configuration, geometric scale, and operating condition parameter space; generate multiple sample operating conditions within the parameter space; and perform CFD calculations on each sample operating condition. S2 Synchronous Extraction: From the CFD calculation results of each sample working condition, at least two types of complete physical field data of heterogeneous hydrodynamic responses are extracted synchronously at a unified dimensionless time node or quasi-steady-state time window, and different responses under the same working condition are assigned the same sample identifier and synchronous time index. S3 Sample Physical Marker Extraction: Extract sample physical markers that move with the sample operating conditions from the complete physical field data to obtain water surface wake field sample physical markers, resistance history sample physical markers, and hull surface load sample physical markers. S4 Transport Alignment and Amplitude Separation: Determine a reference operating condition or generate a reference domain from multiple samples, and extract reference physical markers from it; construct a reversible transport mapping from the reference domain to each sample domain based on the correspondence between the sample physical markers of each sample and the reference physical markers; use the reversible transport mapping to pull the responses of each sample back to the reference domain to obtain the aligned response and transport parameters, and separate the amplitude and bias of the aligned response, saving the amplitude, bias and transport parameters separately; S5 Discrete Physical Closure Operator Construction: On a unified reference surface mesh and synchronous time nodes, a discrete physical closure operator is constructed between responses with physical equality relationships; the discrete physical closure operator is used to convert hull surface pressure and wall shear stress into integral drag, and to calculate the closure residual between direct drag and the integral drag; The model building and training phase includes: S6 Joint Order Reduction: The aligned response in step S4 is subjected to joint order reduction of shared and private latent variables. The overall flow state of the sample is encoded as a shared latent variable used by multiple types of responses, and the unique residual changes of each type of response are encoded as private latent variables corresponding to each response category, so that multiple types of responses under the same working condition are coupled through shared latent variables. S7 Physical Closure Embedding: Based on the discrete physical closure operator in step S5, a physical closure constraint is constructed in the latent space. The physical closure constraint requires that the drag obtained by the integral operator after reconstruction from the latent variables of the surface load is numerically equal to the drag directly predicted. The physical closure constraint is written as an additional objective term into the objective function of the joint order reduction, so that the order reduction process pursues reconstruction accuracy while taking into account physical consistency. S8 surrogate training: Using sample operating parameters as input, and shared latent variables, private latent variables, transport parameters, magnitude and bias as joint output, train a probabilistic surrogate model; The model application phase includes: S9 Constraint Projection: Input the parameters of the working condition to be predicted into the trained probabilistic surrogate model to obtain the original latent variable prediction values and joint covariance. Based on the joint covariance, project the original latent variable prediction values onto a feasible set that satisfies the physical closure constraint to obtain the corrected latent variables. S10 Inverse Reconstruction: Based on the correction latent variable, the aligned response is reconstructed in the reference domain, and the physical coordinates are restored through the inverse mapping of the reversible transport mapping in S4, to obtain the various heterogeneous hydrodynamic responses that mutually satisfy the physical closure relationship under the predicted operating condition.
2. The method for multi-response transport alignment and physical closure prediction of underwater vehicles according to claim 1, characterized in that: In step S2, the heterogeneous hydrodynamic response includes the surface wake field, the drag history, and the hull surface load. The surface wake field includes the surface wake height field or the wake control section velocity field. The drag history includes pressure drag, viscous drag, or total drag history. The hull surface load includes the hull surface pressure distribution or the wall shear stress distribution.
3. The method for multi-response transport alignment and physical closure prediction of underwater vehicles according to claim 1, characterized in that: In step S3, the physical markers of the wake field sample include at least one of the main wave crest ridge, the main wave trough ridge, and the centerline recovery point; the physical markers of the drag history sample include at least one of the initial response point, the first extreme point, the attenuation inflection point, and the quasi-steady-state entry point; and the physical markers of the hull surface load sample include at least one of the pressure peak trajectory, the zero wall shear line or the separation line, and the bow and stern characteristic points.
4. The method for multi-response transport alignment and physical closure prediction of underwater vehicles according to claim 1, characterized in that: In step S4, the reversible transport mapping satisfies the conditions of spatial reversibility and temporal monotonicity to ensure that the mapping will not flatten or flip the grid in space, nor will it reverse the order of events in time.
5. The method for multi-response transport alignment and physical closure prediction of underwater vehicles according to claim 1, characterized in that: In step S5, the discrete physical closure operator obtains the integral resistance by discretizing and summing the axial pressure and wall shear stress of each surface element on the same reference surface grid; the closure residual is the difference between the direct resistance and the integral resistance.
6. The method for multi-response transport alignment and physical closure prediction of underwater vehicles according to claim 1, characterized in that: In step S7, when the physical closure constraint also includes the momentum relationship of the control volume determined by the velocity and pressure of the wake control section, the difference between the surface load integral resistance, the equivalent resistance of the control volume momentum deficit, and the direct resistance together constitutes the closure constraint.
7. The method for multi-response transport alignment and physical closure prediction of underwater vehicles according to claim 1, characterized in that: In step S8, the probabilistic proxy model is a multi-output Gaussian process, a sparse variational Gaussian process, a deep kernel multi-task Gaussian process, or a neural network with a probabilistic output head that can output the joint prediction mean and joint covariance.
8. The method for multi-response transport alignment and physical closure prediction of underwater vehicles according to claim 1, characterized in that: In step S9, during the constrained projection, the correction amount of the original latent variable prediction value is allocated according to the joint covariance. The direction with a larger joint covariance bears a larger correction amount, and the direction with a smaller joint covariance bears a smaller correction amount, so that the corrected latent variable retains the original prediction information of the probabilistic surrogate model to the greatest extent while satisfying the physical closure constraint.
9. The method for multi-response transport alignment and physical closure prediction of underwater vehicles according to claim 1, characterized in that: It also includes a model reliability verification step, which constructs a reliability index using the prediction covariance, latent variable projection distance, physical closure residual, and minimum Jacobian of the reversible transport mapping of the probabilistic surrogate model. When the reliability index exceeds the threshold, it is determined that the current probabilistic surrogate model is not reliable enough, and the working condition to be predicted is added to the supplementary high-fidelity simulation sample set and the probabilistic surrogate model is updated.
10. A multi-response transport alignment and physical closure prediction system for underwater vehicles, which utilizes the prediction method described in any one of claims 1-9, characterized in that: include: The synchronous response acquisition module is used to acquire at least two types of heterogeneous hydrodynamic responses synchronized under multiple sample operating conditions. The physical marker and transport alignment module is used to extract physical markers that move with the sample operating conditions for each type of heterogeneous hydrodynamic response, and construct a reversible transport mapping based on the correspondence between the physical markers and the reference physical markers, pull each sample response back to the reference domain and separate the amplitude and bias. The Physical Closure Operator module is used to construct discrete physical closure operators between responses that have physical equality relationships. The shared-private latent variable joint order reduction module is used to perform joint order reduction on the aligned response, decompose it into shared latent variables and private latent variables corresponding to each response category, and construct physical closure constraints in the latent space according to the discrete physical closure operator. The joint surrogate prediction module is used to train a probabilistic surrogate model with sample operating condition parameters as input and the shared latent variables, private latent variables, transport parameters, magnitude and bias as joint outputs. The constraint projection module is used to project the original latent variable prediction values and their joint covariance to a feasible set that satisfies the physical closure constraint, based on the original latent variable prediction values and their joint covariance of the working condition to be predicted. The inverse transport reconstruction module is used to reconstruct the aligned response in the reference domain and restore it to the physical coordinates through the inverse mapping of the reversible transport mapping, so as to obtain multiple heterogeneous hydrodynamic responses that satisfy the physical closure relationship under the predicted working condition.
