Intelligent optimization design method for mold conformal runner based on deep learning

Abstract

The application relates to the technical field of mold design, and discloses a mold conformal runner intelligent optimization design method based on deep learning, which comprises the following steps: acquiring a mold three-dimensional model, extracting a glue position plane, a hollowed-out plane and thickness distribution data, obtaining processed point cloud data, inputting the processed point cloud data into a three-dimensional point cloud convolutional neural network to output a feature semantic graph, inputting the feature semantic graph and cooling parameters into a GNN model, performing mechanical constraint through Kangaroo2 in the GNN model to output a runner guide curve coordinate sequence, obtaining a runner layout based on the runner guide curve coordinate sequence, and analyzing temperature distribution and stress concentration areas in a simulation cooling process of the generated runner layout; and if the simulation result is substandard, dynamically adjusting runner parameters based on a PPO algorithm to output an optimized runner layout. The application significantly reduces the cost of manual intervention, comprehensively improves cooling performance, improves product qualification rate, and has self-adaptive capacity in complex scenes.

Description

Technical Field

[0001] This invention relates to the field of mold design technology, specifically to a deep learning-based intelligent optimization design method for mold conformal flow channels. Background Technology

[0002] Traditional shoe mold cooling channel design relies heavily on manual experience. When adapting to complex curved surfaces such as sole patterns and arch support structures, it is difficult to accurately conform to the curvature of the surface to achieve uniform cooling. Furthermore, the avoidance design of hollow areas such as ventilation holes and decorative grooves requires repeated manual adjustments, resulting in long design cycles, low cooling efficiency, and high demolding deformation rates. Although existing CAD tools have introduced parametric design, they lack the ability to intelligently analyze the geometric features of the mold, making it difficult to achieve the global optimal solution under complex working conditions. Summary of the Invention

[0003] The purpose of this invention is to solve the above problems by designing a deep learning-based intelligent optimization design method for mold conformal flow channels.

[0004] The first aspect of this invention provides a method for intelligent optimization design of conformal flow channels in molds based on deep learning, the method comprising the following steps:

[0005] The mold's 3D model is obtained, and the data of the glue surface, hollow surface, and thickness distribution are extracted to obtain the processed point cloud data. The processed point cloud data is then input into a 3D point cloud convolutional neural network based on ResNet-50 to output a feature semantic map.

[0006] The cooling parameters are obtained and the feature semantic map and cooling parameters are input into the GNN model. In the GNN model, mechanical constraints are applied through Kangaroo2 to output the flow channel guide curve coordinate sequence.

[0007] The flow channel layout is obtained based on the coordinate sequence of the flow channel guide curve. The generated flow channel layout is then used to analyze the temperature distribution and stress concentration areas during the simulation cooling process.

[0008] If the simulation results do not meet the standards, the optimized flow channel layout is output by dynamically adjusting the flow channel parameters based on the PPO algorithm, using cooling uniformity and flow channel pressure drop as reward functions.

[0009] Optionally, in a first implementation of the first aspect of the present invention, the step of acquiring a three-dimensional model of the mold, extracting data on the adhesive surface, the hollow surface, and the thickness distribution to obtain processed point cloud data, and inputting the processed point cloud data into a three-dimensional point cloud convolutional neural network based on ResNet-50 to output a feature semantic map includes:

[0010] The 3D model of the mold is converted into point cloud data. Data on the glue surface, hollow surface and thickness distribution are extracted. Outliers are removed by RANSAC algorithm and point cloud smoothing is performed by moving least squares method to obtain the processed point cloud data.

[0011] The preprocessed point cloud data is input into a 3D point cloud convolutional neural network based on ResNet-50, and the 3D convolutional layer is used to extract geometric features in the local neighborhood of each point.

[0012] By utilizing the residual block structure and stacking multiple convolutional layers, cross-scale semantic features are captured, and global max pooling is used to aggregate the local features extracted by each layer into a global feature vector.

[0013] Based on the curvature and thickness distribution of the adhesive surface, a heat accumulation probability map of the mold surface is output. The boundary of the hollow area is marked in binary form. The thickness change trend of each area of ​​the sole is calculated by the point cloud coordinate difference, and a feature semantic map is output in combination.

[0014] Optionally, in a second implementation of the first aspect of the present invention, the step of obtaining cooling parameters and inputting the feature semantic map and cooling parameters into a GNN model, and applying mechanical constraints through Kangaroo2 in the GNN model to output a flow channel guide curve coordinate sequence, includes:

[0015] The cooling parameters are converted into feature vectors, and after standardization, they are fused with the thermal sensitivity channel of the feature semantic map to generate temperature field constraints.

[0016] The geometric points in the feature semantic graph are transformed into graph nodes, the graph nodes are connected to obtain edges, and the edge weights are dynamically adjusted through an attention mechanism.

[0017] The GNN model transmits node information through graph convolution operations and outputs the flow channel direction probability distribution of each node. Monte Carlo sampling is used to extract 10-15 control points from the probability distribution. The control points are converted into smooth curves through cubic B-spline interpolation. The curve position is adjusted according to the glue surface offset distance and the hollow surface avoidance buffer to form the initial guide curve.

[0018] The initial guide curve is mechanically constrained by Kangaroo2, and the coordinate sequence of the flow channel guide curve is output.

[0019] Optionally, in a third implementation of the first aspect of the present invention, the step of applying mechanical constraints to the initial guide curve using Kangaroo2 and outputting the flow channel guide curve coordinate sequence includes:

[0020] Each control point of the initial guide curve is used as a particle in Kangaroo2 to determine the flow channel inlet point, and constraints are configured on the structural points, and rigid collision boundaries are set for the hollowed-out area.

[0021] Set the maximum number of iterations to 200 steps, calculate the resultant force on the current control point, and update the control point position;

[0022] When the maximum number of iterations is reached, equidistant sampling points on the optimized curve are extracted to form a flow channel guide curve coordinate sequence.

[0023] Optionally, in the fourth implementation of the first aspect of the present invention, the collision constraints are optimized in the first 50 steps of the iteration process, the curvature constraints are optimized in the middle 100 steps, and the anchor point constraints are adjusted in the last 50 steps.

[0024] Optionally, in the fifth implementation of the first aspect of the present invention, the step of dynamically adjusting the flow channel parameters and outputting the optimized flow channel layout based on the PPO algorithm, using cooling uniformity and flow channel pressure drop as reward functions, if the simulation results do not meet the standards, includes:

[0025] Identify the non-compliant areas, construct an optimization environment, and generate 50 candidate solutions for parameter adjustment based on the current optimization environment;

[0026] Each scheme is simulated and evaluated. Based on the latest simulation results, the cumulative reward of each scheme is calculated according to the reward function, and the flow channel parameters are adjusted accordingly.

[0027] After each iteration, the physical simulation is run again to verify the performance under the new parameters. When a set of parameters is found that can achieve the optimal balance between cooling uniformity and flow channel pressure drop, the iteration is stopped and the final optimized flow channel layout is output.

[0028] Optionally, in a sixth implementation of the first aspect of the present invention, the optimization environment includes a state space, an action space, and a reward function, wherein the state space includes at least geometric features, simulation results, and historical actions, and the action space includes at least continuous actions and discrete actions.

[0029] A second aspect of the present invention provides a deep learning-based intelligent optimization design system for conformal flow channels in molds, the system comprising:

[0030] The shoe mold feature analysis module is used to acquire the 3D model of the mold, extract the glue surface, hollow surface and thickness distribution data, obtain the processed point cloud data, and input the processed point cloud data into the 3D point cloud convolutional neural network based on ResNet-50 to output the feature semantic map.

[0031] The curve output module is used to obtain cooling parameters and input the feature semantic map and cooling parameters into the GNN model. In the GNN model, mechanical constraints are applied through Kangaroo2 to output the flow channel guide curve coordinate sequence.

[0032] The simulation module is used to obtain the flow channel layout based on the coordinate sequence of the flow channel guide curve, and to analyze the temperature distribution and stress concentration areas in the simulated cooling process of the generated flow channel layout.

[0033] The parameter optimization module is used to dynamically adjust the flow channel parameters and output the optimized flow channel layout based on the PPO algorithm, using cooling uniformity and flow channel pressure drop as reward functions if the simulation results do not meet the standards.

[0034] A third aspect of the present invention provides a deep learning-based intelligent optimization design device for mold conformal flow channels, the deep learning-based intelligent optimization design device for mold conformal flow channels includes a memory and at least one processor, the memory storing instructions; the at least one processor invokes the instructions in the memory to cause the deep learning-based intelligent optimization design device for mold conformal flow channels to perform the various steps of the deep learning-based intelligent optimization design method for mold conformal flow channels as described in any of the preceding claims.

[0035] A fourth aspect of the present invention provides a computer-readable storage medium storing instructions that, when executed by a processor, implement the steps of the deep learning-based intelligent optimization design method for mold conformal flow channels as described in any of the preceding claims.

[0036] The technical solution provided by this invention involves acquiring a three-dimensional model of the mold, extracting data on the glue surface, hollow surface, and thickness distribution to obtain processed point cloud data, and inputting the processed point cloud data into a three-dimensional point cloud convolutional neural network based on ResNet-50 to output a feature semantic map. Cooling parameters are acquired, and the feature semantic map and cooling parameters are input into a GNN model. In the GNN model, Kangaroo2 is used for mechanical constraints to output a flow channel guiding curve coordinate sequence. Based on the flow channel guiding curve coordinate sequence, the flow channel layout is obtained, and the generated flow channel layout is used to analyze the temperature distribution and stress concentration areas during the simulation cooling process. If the simulation results are unsatisfactory, cooling uniformity and flow channel pressure drop are used as reward functions, and the flow channel parameters are dynamically adjusted based on the PPO algorithm to output an optimized flow channel layout. This invention automatically analyzes the geometric features of the mold through a three-dimensional point cloud convolutional neural network, combines it with a GNN model to generate a flow channel layout, shortens the design cycle, supports real-time parameter adjustment and automatic flow channel reconstruction, significantly reduces manual intervention costs, comprehensively improves cooling performance, increases product qualification rate, and has adaptive capabilities for complex scenarios. Attached Figure Description

[0037] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0038] Figure 1 A schematic diagram of the first embodiment of the intelligent optimization design method for mold conformal flow channels based on deep learning provided in this invention;

[0039] Figure 2 A schematic diagram of the second embodiment of the intelligent optimization design method for mold conformal flow channels based on deep learning provided in this invention;

[0040] Figure 3 A schematic diagram of the structure of the intelligent optimization design device for mold conformal flow channels based on deep learning provided in an embodiment of the present invention. Detailed Implementation

[0041] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” or “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, apparatus, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0042] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 A schematic diagram of the first embodiment of the intelligent optimization design method for mold conformal flow channels based on deep learning provided by this invention. The method specifically includes the following steps:

[0043] Step 101: Obtain the 3D model of the mold, extract the glue surface, hollow surface and thickness distribution data to obtain the processed point cloud data, and input the processed point cloud data into the 3D point cloud convolutional neural network based on ResNet-50 to output the feature semantic map.

[0044] In this embodiment, the 3D model of the mold is converted into point cloud data, and the data of the glue surface, hollow surface, and thickness distribution are extracted. Outliers are removed by the RANSAC algorithm, and the point cloud is smoothed by the moving least squares method to obtain the processed point cloud data. The preprocessed point cloud data is input into a 3D point cloud convolutional neural network based on ResNet-50. Geometric features are extracted in the local neighborhood of each point using 3D convolutional layers. Using the residual block structure, multiple convolutional layers are stacked to capture cross-scale semantic features. Global max pooling is used to aggregate the local features extracted by each layer into a global feature vector. Based on the curvature and thickness distribution of the glue surface, the heat accumulation probability map of the mold surface is output. The boundary of the hollow area is marked in binary form. The thickness change trend of each area of ​​the sole is calculated by the point cloud coordinate difference, and the feature semantic map is output in a comprehensive manner.

[0045] In this embodiment, the RANSAC algorithm is used to remove outliers that may exist in the initial point cloud data: by randomly selecting a point set to fit a planar or curved surface model, calculating the distance between the remaining points and the model, and identifying points that exceed the threshold as outliers and removing them, the effective point set that conforms to the real geometric structure of the mold is retained. Subsequently, the moving least squares method is used to smooth the denoised point cloud: a local coordinate system is established with each point as the center, and a local polynomial surface is fitted by weighted least squares. The weight function is dynamically adjusted to retain edge features and curvature change regions. While removing high-frequency noise in the point cloud, key details such as the boundary of the glue surface and the outline of the hollow area are maintained, and finally, high-quality point cloud data with continuous curvature and uniform distribution is generated.

[0046] In this embodiment, the preprocessed point cloud data is converted into an input format that can be recognized by a three-dimensional convolutional neural network, which is usually a multi-dimensional array containing coordinates and normal vectors. The size of each point cloud sample is uniformly fixed with a fixed number of points, either 1024 or 2048. Insufficient parts are filled by random sampling, and excess parts are downsampled by sampling the farthest point. The coordinate values ​​are normalized, and all point cloud data are scaled to a unit cube space of [-1,1] to eliminate the influence of different mold sizes on the model.

[0047] Three-dimensional convolutional layers or point cloud-specific convolutional layers are used to extract geometric features in the local neighborhood of each point, including the point's curvature, local surface roughness, and neighborhood point density. The local neighborhood is determined by the K-nearest neighbor algorithm or ball query, with a radius of 0.05-0.1m. Batch normalization and ReLU activation function are used to enhance the nonlinear representation of features.

[0048] By utilizing a residual block structure and stacking multiple convolutional layers, the receptive field is gradually expanded to capture cross-scale semantic features. For example, the first residual block focuses on the local shape features of a single perforation; the second residual block combines multiple local features to identify the topological structure of complex perforated areas, such as the distribution pattern of ventilation holes; the third residual block integrates global geometric features to determine the curvature distribution trend of the sole surface, such as the high curvature region at the arch and the flat region at the toe; skip connections preserve the detailed features of the lower layers, avoiding the gradient vanishing problem of deep networks.

[0049] Global max pooling or attention pooling is used to aggregate the local features extracted from each layer into a global feature vector, which represents the overall geometric properties of the mold, such as the proportion of hollow areas, average curvature value, and wall thickness distribution range.

[0050] Heat Sensitivity Channel: Based on the curvature and thickness distribution of the glue surface, outputs a heat accumulation probability map of the mold surface. The red area represents the high heat area, such as the area with curvature ≥ 0.8, where runners should be arranged first; the blue area represents the low heat area, where the runner density can be reduced.

[0051] Hollow Mask Channel: The boundary of the hollow area is marked in binary form. The white area represents the hollow structure that needs to be avoided, such as vents and texture grooves, while the black area represents the solid area, which is directly used for subsequent Boolean operations to remove flow channel conflict parts.

[0052] Thickness gradient channel: The thickness change trend of each area of ​​the sole is calculated by the point cloud coordinate difference. The color depth indicates the thickness gradient. Darker areas indicate that the concave depth of the flow channel needs to be adjusted to balance cooling efficiency and structural strength.

[0053] The feature semantic map is aligned with the original point cloud data in terms of spatial coordinates. Each pixel corresponds to the actual physical position on the mold surface. The output resolution matches the point cloud sampling density. For example, when the pixel spacing is 0.5mm, the pixel accuracy of the semantic map is 0.5mm×0.5mm.

[0054] Step 102: Obtain cooling parameters and input the feature semantic map and cooling parameters into the GNN model. In the GNN model, apply mechanical constraints using Kangaroo2 to output the flow channel guide curve coordinate sequence.

[0055] In this embodiment, the cooling parameters can be extracted from the injection molding machine control system or historical process database, including coolant type and inlet temperature, injection pressure and cycle time, target cooling rate and demolding temperature threshold. The cooling parameters are converted into feature vectors, mapped to the [0,1] interval through standardization, and fused with the thermal sensitivity channel of the feature semantic map to generate temperature field constraints, such as higher coolant flow rate weights for high heat areas. Key geometric points in the feature semantic map are converted into graph nodes, such as curvature extrema, cutout boundary points, and thickness abrupt change points. Each node contains multi-dimensional features: geometric features, physical features, and process features. Nodes are connected only by a distance ≤5mm through spatial distance thresholds to construct local neighborhood relationships. Based on the similarity of thermal sensitivity and cutout mask features, nodes with similar semantics but far spatial distances are connected, such as cross-regional high heat points. Absolute coordinate information of nodes is injected using position encoding to preserve spatial geometric relationships. Edge weights are dynamically adjusted through an attention mechanism to highlight key areas, such as increasing the connection weight of nodes at the arch support structure.

[0056] Step 103: Obtain the flow channel layout based on the coordinate sequence of the flow channel guide curve, and analyze the temperature distribution and stress concentration areas during the simulation cooling process of the generated flow channel layout.

[0057] In this embodiment, based on the coordinate sequence of the flow channel guide curve, a complete flow channel layout is first generated through parametric modeling: according to the preset flow channel diameter and concave depth, a circular flow channel with a constant cross-section is generated along the guide curve, and a smooth transition design is adopted at the branch points to reduce flow resistance. The generated flow channel model is subjected to Boolean operation with the mold entity to ensure that there is no interference between the flow channel and the hollow area and that the minimum wall thickness requirement is met. Subsequently, the cooling process is simulated: a coupled computational fluid dynamics and heat conduction model is used, and boundary conditions such as coolant inlet temperature and flow velocity are set. The Navier-Stokes equations and heat conduction equations are discretized and solved through finite element mesh generation to simulate the coolant flow path and the temperature field changes of the mold. The key analysis focuses on temperature distribution: high-temperature areas are identified through temperature cloud maps, such as hot spots on the glue surface with curvature ≥0.8, and the temperature standard deviation is calculated to assess cooling uniformity; at the same time, structural mechanics simulation is performed, and injection pressure and thermal stress load are applied. Areas with stress concentration coefficient ≥3 are located through stress cloud maps, such as flow channel bends and abrupt changes in wall thickness. The simulation results are mapped to a feature semantic map, and areas that need optimization are marked, such as areas with temperature standard deviation >10℃ or maximum stress >80% of the material yield strength.

[0058] Step 104: If the simulation results do not meet the standards, the optimized flow channel layout is output by dynamically adjusting the flow channel parameters based on the PPO algorithm, using cooling uniformity and flow channel pressure drop as reward functions.

[0059] In this embodiment, substandard areas are identified, and an optimization environment is constructed. Based on the current optimization environment, 50 candidate schemes for parameter adjustment are generated. Each scheme is evaluated through simulation. Based on the latest simulation results, the cumulative reward of each scheme is calculated according to the reward function, and the flow channel parameters are adjusted. After each iteration, the physical simulation is run again to verify the performance under the new parameters. When a set of parameters that can achieve the optimal balance between cooling uniformity and flow channel pressure drop is found, the iteration is stopped, and the final optimized flow channel layout is output.

[0060] In this embodiment, the optimized environment includes a state space, an action space, and a reward function. The state space includes at least geometric features, simulation results, and historical actions, while the action space includes at least continuous actions and discrete actions.

[0061] In this embodiment, the cooling uniformity is evaluated based on the temperature distribution of the mold during the cooling process. Ideally, the temperature distribution should be as uniform as possible to avoid local overheating or insufficient cooling. The flow channel pressure drop takes into account the pressure loss of the fluid as it passes through the entire flow channel system. A lower pressure drop means a more efficient cooling effect and lower energy consumption.

[0062] In this embodiment, geometric features include the curvature distribution of the glue surface, the topology of the hollowed-out area, the mold wall thickness gradient, and feature semantic graph statistics; simulation results cover cooling uniformity indicators, flow channel pressure drop distribution, and structural strength parameters; historical action records the flow channel diameter adjustment, concave depth change value, and branch angle adjustment direction in the previous 3-5 rounds of optimization, which are normalized to form a standardized input vector; the action space adopts a continuous-discrete hybrid design, with continuous actions including flow channel diameter, concave depth, and coolant flow rate; discrete actions include adding or removing flow channel branches, switching inlet positions, and adjusting constraint priorities. All actions are associated with engineering constraint boundaries, such as diameter adjustment step size ≤ 1mm, branch angle... The degree is limited to 30°-150° to avoid generating infeasible solutions; the reward function is designed as a multi-objective weighted system, with cooling performance reward: +5 for every 1°C decrease in temperature standard deviation, and +20 for a uniformity index exceeding 0.9; flow channel efficiency reward: +3 for every 0.1MPa reduction in pressure drop, and +10 for a 5% reduction in total flow channel length; structural safety reward: +2 for every 0.5mm increase in minimum wall thickness, and +8 for a 10% reduction in the area of ​​stress concentration regions; constraint penalty: -15 for interference between flow channels and hollow areas, -10 for insufficient wall thickness, and -5 for parameter exceeding limits; the reward calculation integrates immediate and delayed feedback, balancing the effects of short-term adjustments and long-term optimization through discount factors.

[0063] In this embodiment, the flow channel inlet is the location of the cooling medium inlet and outlet, which is determined according to the position of the mold machine, and is usually located in the high-heat area of ​​the heel or toe; the mold glue surface is the curved surface (such as the working surface of the sole) where the shoe mold contacts the shoe material, which determines the reference surface for the flow channel generation; the mold shape is the overall geometric model of the shoe mold, including the outline and thickness distribution of the upper and sole; the mold hollow surface is the hollow area (ventilation hole, texture groove) that needs to be avoided, and conflicts are eliminated through Boolean operations; the concave depth is the safe distance between the flow channel and the glue surface that is automatically calculated based on the sole thickness.

[0064] In this embodiment, a 3D shoe mold model is imported, the glue surface, mold shape, and hollow surface are identified, and cooling parameters are set: injection pressure, cycle time, and runner wall thickness spacing. The glue surface is offset inward to generate a runner wall thickness layer. The offset amount formula is as follows:

[0065] Offsetdistance = k × material shrinkage rate × sole thickness

[0066] Where k is the safety factor (default 1.2~1.5);

[0067] Offset the cutout surface outwards by a safe distance, remove overlapping areas with the flow channel using Volume Boolean operations, manually draw reference lines along the sole outline in Rhino, and apply constraints using the Kangaroo2 plugin for Grasshopper: collision constraints prevent flow channels from self-intersecting and branch spacing from being too small, anchor point constraints fix the flow channel inlet (1,8) position, and curvature constraints control the flow channel direction to be consistent with the curvature of the shoe mold surface;

[0068] Initial flow channel parameters are generated based on input parameters (flow channel diameter, wall thickness, concave depth, etc.), the flow channel distribution density is adjusted, and the flow channel layout is calculated by combining the collision force parameters.

[0069] In this embodiment, by offsetting the adhesive surface and performing Boolean operations on the hollowed-out surface (such as pores) Volume, the flow channel can automatically avoid complex structures. Combined with Kangaroo2 collision constraints, the flow channels are prevented from intersecting themselves, allowing the flow channels to extend naturally along the curved surface, reducing manual intervention. Parameters such as flow channel diameter and concave depth can be modified in real time and trigger flow channel reconstruction.

[0070] Please see Figure 2 A schematic diagram of the second embodiment of the intelligent optimization design method for mold conformal flow channels based on deep learning provided in this invention. The method includes:

[0071] Step 201: Convert the cooling parameters into feature vectors, and after standardization, fuse them with the thermal sensitivity channel of the feature semantic map to generate temperature field constraints.

[0072] Step 202: Convert the geometric points in the feature semantic graph into graph nodes, connect the graph nodes to obtain edges, and dynamically adjust the edge weights through an attention mechanism;

[0073] Step 203: The GNN model transmits node information through graph convolution operations and outputs the flow channel direction probability distribution of each node. Monte Carlo sampling is used to extract 10-15 control points from the probability distribution. The control points are converted into smooth curves through cubic B-spline interpolation. The curve position is adjusted according to the glue surface offset distance and the hollow surface avoidance buffer to form the initial guide curve.

[0074] Step 204: Apply mechanical constraints to the initial guide curve using Kangaroo2 and output the flow channel guide curve coordinate sequence.

[0075] In this embodiment, each control point of the initial guide curve is treated as a particle in Kangaroo2 to determine the inlet point of the flow channel, and constraints are configured on the structural points. Rigid collision boundaries are set for the hollowed-out area. The maximum number of iterations is set to 200 steps. The resultant force on the current control point is calculated and the position of the control point is updated. When the maximum number of iterations is reached, equidistant sampling points on the optimized curve are extracted to form a coordinate sequence of the flow channel guide curve.

[0076] In this embodiment, the first 50 steps of the iteration process optimize the collision constraints, the middle 100 steps optimize the curvature constraints, and the last 50 steps adjust the anchor point constraints.

[0077] In this embodiment, particle properties are assigned mass attributes to control points on the curve, with the mass of key points increased by 2-3 times. A damping coefficient is set to control the convergence speed and avoid excessive oscillation. An attractive field is set above the adhesive surface to keep the flow channel curve close to the adhesive surface. A repulsive force field is set around the hollow area, with the intensity inversely proportional to the distance.

[0078] In this embodiment, constraints are applied in stages. The first 50 steps prioritize collision constraints, quickly eliminating channel self-intersection and ensuring a safe distance from the hollowed-out area by simulating the elastic repulsion between particles and the rigid rebound force of the hollowed-out boundary. The middle 100 steps focus on curvature constraint optimization, using virtual spring force to pull particles, making the curve curvature close to the curvature distribution of the glue surface. For example, the bending radius of the channel corresponding to the high curvature area of ​​the arch is automatically adjusted to 10-15mm. At the same time, the gravitational field causes the entire channel to shift towards the glue surface by 5-8mm. The last 50 steps fine-tune the anchor point constraints, fixing the inlet position and calibrating the outlet direction to ensure alignment with the mold machine interface. In each iteration, the resultant force on the particles, including collision force, curvature force, and anchor point tension, is calculated. The particle position is updated according to Newton's laws of motion until the maximum number of iterations is reached or the position change is less than 0.1mm, reaching a convergence threshold. After optimization, the final curve is sampled at 2mm intervals to extract a point sequence containing three-dimensional coordinates and curvature values, forming a high-precision channel guide curve coordinate sequence.

[0079] Figure 3 This is a schematic diagram of a deep learning-based intelligent optimization design device for mold conformal flow channels, provided in an embodiment of the present invention. This deep learning-based intelligent optimization design device 300 for mold conformal flow channels can vary significantly due to different configurations or performance. It may include one or more central processing units (CPUs) 310 (e.g., one or more processors) and a memory 320, and one or more storage media 330 (e.g., one or more mass storage devices) for storing application programs 333 or data 332. The memory 320 and storage media 330 can be temporary or persistent storage. The program stored in the storage media 330 may include one or more modules (not shown in the diagram), each module including a series of instruction operations on the deep learning-based intelligent optimization design device 300 for mold conformal flow channels. Furthermore, the processor 310 may be configured to communicate with the storage media 330 and execute the series of instruction operations in the storage media 330 on the deep learning-based intelligent optimization design device 300 for mold conformal flow channels to implement the method provided in the above embodiment.

[0080] The deep learning-based intelligent optimization design device 300 for mold conformal flow channels may also include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input / output interfaces 360, and / or one or more operating devices 331, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that... Figure 3The intelligent optimization design device structure for mold conformal flow channels based on deep learning shown does not constitute a limitation on the computer device provided by this invention. It may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0081] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when the instructions are executed on a computer, cause the computer to perform each step of the deep learning-based mold conformal flow channel intelligent optimization design method provided in the above embodiments.

[0082] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the above-described equipment or apparatus / unit can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0083] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0084] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A deep learning-based intelligent optimization design method for conformal flow channels in molds, characterized in that, The method includes the following steps: The mold's 3D model is obtained, and the data of the glue surface, hollow surface, and thickness distribution are extracted to obtain the processed point cloud data. The processed point cloud data is then input into a 3D point cloud convolutional neural network based on ResNet-50 to output a feature semantic map. The cooling parameters are obtained and the feature semantic map and cooling parameters are input into the GNN model. In the GNN model, mechanical constraints are applied through Kangaroo2 to output the flow channel guide curve coordinate sequence. The flow channel layout is obtained based on the coordinate sequence of the flow channel guide curve. The generated flow channel layout is then used to analyze the temperature distribution and stress concentration areas during the simulation cooling process. If the simulation results do not meet the standards, the optimized flow channel layout is output by dynamically adjusting the flow channel parameters based on the PPO algorithm, using cooling uniformity and flow channel pressure drop as reward functions.

2. The intelligent optimization design method for mold conformal flow channels based on deep learning as described in claim 1, characterized in that, The process involves acquiring a 3D model of the mold, extracting data on the adhesive surface, hollowed-out surface, and thickness distribution to obtain processed point cloud data, and then inputting the processed point cloud data into a ResNet-50-based 3D point cloud convolutional neural network to output a feature semantic map, including: The 3D model of the mold is converted into point cloud data. Data on the glue surface, hollow surface and thickness distribution are extracted. Outliers are removed by RANSAC algorithm and point cloud smoothing is performed by moving least squares method to obtain the processed point cloud data. The preprocessed point cloud data is input into a 3D point cloud convolutional neural network based on ResNet-50, and the 3D convolutional layer is used to extract geometric features in the local neighborhood of each point. By utilizing the residual block structure and stacking multiple convolutional layers, cross-scale semantic features are captured, and global max pooling is used to aggregate the local features extracted by each layer into a global feature vector. Based on the curvature and thickness distribution of the adhesive surface, a heat accumulation probability map of the mold surface is output. The boundary of the hollow area is marked in binary form. The thickness change trend of each area of ​​the sole is calculated by the point cloud coordinate difference, and a feature semantic map is output in combination.

3. The intelligent optimization design method for conformal flow channels of molds based on deep learning as described in claim 1, characterized in that, The process of acquiring cooling parameters and inputting the feature semantic map and cooling parameters into the GNN model, and applying mechanical constraints through Kangaroo2 in the GNN model to output the flow channel guide curve coordinate sequence, includes: The cooling parameters are converted into feature vectors, and after standardization, they are fused with the thermal sensitivity channel of the feature semantic map to generate temperature field constraints. The geometric points in the feature semantic graph are transformed into graph nodes, the graph nodes are connected to obtain edges, and the edge weights are dynamically adjusted through an attention mechanism. The GNN model transmits node information through graph convolution operations and outputs the flow channel direction probability distribution of each node. Monte Carlo sampling is used to extract 10-15 control points from the probability distribution. The control points are converted into smooth curves through cubic B-spline interpolation. The curve position is adjusted according to the glue surface offset distance and the hollow surface avoidance buffer to form the initial guide curve. The initial guide curve is mechanically constrained by Kangaroo2, and the coordinate sequence of the flow channel guide curve is output.

4. The intelligent optimization design method for conformal flow channels of molds based on deep learning as described in claim 3, characterized in that, The process of applying mechanical constraints to the initial guide curve using Kangaroo2 and outputting the flow channel guide curve coordinate sequence includes: Each control point of the initial guide curve is used as a particle in Kangaroo2 to determine the flow channel inlet point, and constraints are configured on the structural points, and rigid collision boundaries are set for the hollowed-out area. Set the maximum number of iterations to 200 steps, calculate the resultant force on the current control point, and update the control point position; When the maximum number of iterations is reached, equidistant sampling points on the optimized curve are extracted to form a flow channel guide curve coordinate sequence.

5. The intelligent optimization design method for mold conformal flow channels based on deep learning as described in claim 4, characterized in that, During the iteration process, the first 50 steps optimize collision constraints, the middle 100 steps optimize curvature constraints, and the last 50 steps adjust anchor point constraints.

6. The intelligent optimization design method for mold conformal flow channels based on deep learning as described in claim 1, characterized in that, If the simulation results do not meet the standards, the optimized flow channel layout is output by dynamically adjusting the flow channel parameters based on the PPO algorithm, using cooling uniformity and flow channel pressure drop as reward functions, including: Identify the non-compliant areas, construct an optimization environment, and generate 50 candidate solutions for parameter adjustment based on the current optimization environment; Each scheme is simulated and evaluated. Based on the latest simulation results, the cumulative reward of each scheme is calculated according to the reward function, and the flow channel parameters are adjusted accordingly. After each iteration, the physical simulation is run again to verify the performance under the new parameters. When a set of parameters is found that can achieve the optimal balance between cooling uniformity and flow channel pressure drop, the iteration is stopped and the final optimized flow channel layout is output.

7. The intelligent optimization design method for conformal flow channels of molds based on deep learning as described in claim 6, characterized in that, The optimization environment includes a state space, an action space, and a reward function. The state space includes at least geometric features, simulation results, and historical actions, while the action space includes at least continuous actions and discrete actions.

8. A system for implementing the intelligent optimization design method for mold conformal flow channels based on deep learning as described in claim 1, characterized in that, The system includes: The shoe mold feature analysis module is used to acquire the 3D model of the mold, extract the glue surface, hollow surface and thickness distribution data, obtain the processed point cloud data, and input the processed point cloud data into the 3D point cloud convolutional neural network based on ResNet-50 to output the feature semantic map. The curve output module is used to obtain cooling parameters and input the feature semantic map and cooling parameters into the GNN model. In the GNN model, mechanical constraints are applied through Kangaroo2 to output the flow channel guide curve coordinate sequence. The simulation module is used to obtain the flow channel layout based on the coordinate sequence of the flow channel guide curve, and to analyze the temperature distribution and stress concentration areas in the simulated cooling process of the generated flow channel layout. The parameter optimization module is used to dynamically adjust the flow channel parameters and output the optimized flow channel layout based on the PPO algorithm, using cooling uniformity and flow channel pressure drop as reward functions if the simulation results do not meet the standards.

9. A deep learning-based intelligent optimization design device for conformal flow channels in molds, characterized in that, The deep learning-based mold conformal flow channel intelligent optimization design device includes a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the deep learning-based mold conformal flow channel intelligent optimization design device to perform each step of the deep learning-based mold conformal flow channel intelligent optimization design method as described in any one of claims 1-7.

10. A computer-readable storage medium storing instructions thereon, characterized in that, When the instructions are executed by the processor, they implement each step of the intelligent optimization design method for mold conformal flow channels based on deep learning as described in any one of claims 1-7.

Images