Micro-channel heat exchanger additive manufacturing process prediction optimization method based on meta-learning
By employing an intelligent process optimization method based on meta-learning and reinforcement learning, the problems of complex process parameters and difficulty in transfer learning in the manufacturing of micro-channel heat exchangers are solved, achieving efficient and low-cost multi-size process optimization to meet the rapid response requirements of aerospace and other fields.
Patent Information
- Application Number
- CN202511330352.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-17
- Publication Date
- 2025-12-30
AI Technical Summary
Existing technologies for manufacturing microchannel heat exchangers suffer from problems such as complex process parameters and slight fluctuations leading to structural deviations and performance degradation. Traditional prediction methods cannot achieve transfer learning under new parameters, resulting in low process development efficiency and difficulty in meeting the demand for rapid production of multiple sizes.
An intelligent process optimization method based on meta-learning and reinforcement learning is adopted. The size parameters are automatically extracted through U-Net+SE network, a meta-learning process prediction model is constructed, and multi-objective optimization is achieved by combining the DDPG algorithm, forming a closed-loop control of "identification-prediction-optimization-verification".
It achieves high-precision recognition capability with small samples, significantly reduces experimental costs, has the ability to rapidly transfer across sizes, meets the rapid response requirements of aerospace and other fields for multi-specification parts, and improves manufacturing consistency and efficiency.
Smart Images

Figure CN121236458A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of special processing and artificial intelligence, in particular to a micro-channel heat exchanger additive manufacturing process prediction optimization method based on meta-learning. BACKGROUND
[0002] Micro-channel heat exchangers are widely used in aerospace, electronic device cooling, energy systems and other fields with extremely high requirements for heat exchange performance. In order to meet the needs of complex structure and high dimensional accuracy, special processing methods such as selective laser melting (SLM) have become a key means for manufacturing micro-channel heat exchangers. However, complex process parameters and small fluctuations are prone to cause structural deviation and performance decline, therefore, process-quality prediction and performance optimization are of great significance to improve processing consistency, reduce trial and error costs and speed up design iteration, and are the key support for realizing the engineering application of high-performance micro-channel heat exchangers.
[0003] Existing process prediction and optimization methods include process prediction methods and process optimization methods, but traditional prediction methods cannot realize transfer learning under new parameters, i.e. micro-channel structures have significant changes in width, height, spacing and wall thickness, and traditional process optimization requires a large number of physical samples for verification. For example, when using full factorial experimental design, only 5 parameters with 3 levels each require 243 experiments, which is a huge cost in terms of experimental time and resources. Deep learning methods in the prior art, such as CNN and MLP, require a large number of training data for each new size structure, usually ≥100 groups, and the model obtained by training is only applicable to a specific size. At the same time, the process optimization algorithm has a cold start problem, and traditional optimization algorithms such as genetic algorithm and particle swarm algorithm need to optimize from scratch for each new structure size, lacking knowledge transfer ability. For example, the parameters obtained by optimizing 200 channels are applied to 250 channels, the forming size error averages 79.7%, and iteration experiments need to be performed again. This "one-size-one-optimization" mode leads to low process development efficiency and is difficult to meet the demand for rapid production of multiple sizes. SUMMARY
[0004] The purpose of the present application is to provide a micro-channel heat exchanger additive manufacturing process prediction optimization method based on meta-learning, which utilizes meta-learning and reinforcement learning algorithms to convert the additive manufacturing process optimization problem of micro-channel heat exchangers into an intelligent process optimization decision-making process, and proposes a collaborative technology path of "automatic image recognition-small sample learning-intelligent closed-loop optimization" to solve the problems raised in the background art.
[0005] The specific technical solutions provided by the present application are as follows: a micro-channel heat exchanger additive manufacturing process prediction optimization method based on meta-learning, comprising the following steps: S1: Define the key geometric parameters of the micro-channel heat exchanger, determine the SLM process parameter space.
[0006] Preferably, the basic analysis unit of the micro-channel heat exchanger is composed of periodic repeating features, including cold and hot side channels and inter-wall structures, and the key geometric parameters of the micro-channel heat exchanger include: channel width, height, rib width of cold and hot side channels; and the thickness of the heat insulation / conduction wall arranged between the cold and hot channels.
[0007] S2: Adopt orthogonal table design to design SLM process experiment scheme containing multiple sizes.
[0008] Preferably, based on the SLM process, process optimization experiment design is carried out for the parameterized micro-channel heat exchanger structure, and laser power , scanning speed , layer thickness and scanning pitch are selected as key process variables, and combined with typical channel structure sizes, including but not limited to 100 , 200 , 300 , , parameter values, a multi-factor experiment scheme is constructed, and or L16(4^5) orthogonal table is used to arrange experiments to prepare micro-channel samples with different structure parameter combinations.
[0009] S3: Prepare experimental samples through SLM equipment, acquire morphology data using laser confocal and SEM, and perform automatic size parameter extraction based on U-Net+SE network.
[0010] Preferably, based on the above SLM process orthogonal experiment prepared samples, image acquisition, preprocessing and deep learning modeling work are carried out, an image dataset with structure label is constructed, and automatic size extraction and morphology parameter recognition are realized.
[0011] S4: Construct process-molding quality meta-task set, fuse simulation data training using MAML++ framework, and establish meta-learning process prediction model.
[0012] Preferably, the meta-learning process prediction model is: taking structure size parameters as task descriptors, learning cross-task mapping rules using MAML++ algorithm; establishing a meta-model through a small amount of basic experiments to realize cross-size prediction and support small sample online rapid adaptation.
[0013] Preferably, it specifically includes: S41: Build multi-scale tasks to decouple cross-size features and share mechanisms; S42: meta-learning training is performed by using the improved MAML++ framework; S43: cross-size generalization verification and adaptive adjustment are performed.
[0014] S5: the state and action spaces of the DDPG algorithm are established, a multi-objective reward function is constructed, and an Actor-Critic network is trained to output optimal process and forming quality parameter groups.
[0015] Preferably, based on the constructed meta-learning process prediction model, the target process optimization task is formalized as a multi-objective optimization problem, specifically: an optimization objective of minimizing forming size error, surface roughness and flatness is constructed; a three-level optimization architecture including a meta-model cross-size prediction layer, an intelligent process optimization strategy layer and an online updating layer is used for intelligent optimization of process parameters.
[0016] Preferably, the DDPG algorithm is used as the implementation path of the intelligent process optimization method, specifically: an implementation environment including a state space, an action space, a reward function, an Actor network as a policy network and a Critic network as a value network is constructed; a cross-size intelligent prediction optimization method based on the DDPG algorithm is trained; finally, the trained curve law is used to determine the strategy certainty to achieve optimization.
[0017] S6: a closed-loop verification system is constructed to verify the optimal process parameters output by steps S2-S5, and intelligent tuning and cyclic reinforcement learning are performed based on the verification results.
[0018] Preferably, the process optimization is converted into a multi-objective optimization problem, and the meta-optimizer is combined to efficiently search for optimal parameters in a virtual environment; the process is adjusted in real time through image recognition feedback to form a "recognition-prediction-optimization-verification" closed loop, ensuring forming precision and consistency.
[0019] Compared with the prior art, the present application has the following beneficial effects: (1) Small sample high-precision recognition capability: The semantic segmentation model based on meta-learning of the present application can adapt to different sizes of micro-channel structures with only a small number of labeled samples, realize micron-level precision boundary recognition (error < 2um), and has a robustness to complex topography such as channel deformation and fusion defects. The efficiency is improved by more than 15 times compared with traditional manual measurement.
[0020] (2) Experimental cost is greatly reduced: The present application cooperates the meta-learning prediction model with the virtual process optimization strategy to reduce the number of physical experiments by more than 80%. Taking a typical orthogonal experiment as an example, the number of samples required for optimization is reduced from 256 groups of traditional L16(4^5) to 5-10 groups, which significantly saves material and time cost.
[0021] (3) Multi-objective process optimization closed-loop control: The present application combines the real-time feedback of structural parameters of image recognition and the dynamic adjustment capability of meta-optimizer to realize the synchronous optimization of forming precision, surface quality and flatness, forming a "recognition-prediction-optimization" manufacturing quality closed-loop improvement chain.
[0022] (4) Cross-size rapid migration capability: The optimizer of the meta-learning framework of the present application only needs 1-2 fine-tuning to converge in a new structure size task, adapts to the process optimization needs of different channel configurations (such as square, irregular cross-section), shortens the migration training time by 90%, and meets the rapid response requirements of aerospace and other fields for multi-specification parts.
[0023] (5) The present application breaks through the limitations of traditional methods in strong sample dependence, long optimization period and insufficient generalization, and provides a standardized solution for the intelligent design and manufacturing of high-precision micro-channel heat exchangers. BRIEF DESCRIPTION OF DRAWINGS
[0024] Figure 1 is a cross-size intelligent prediction optimization method step flow chart of the micro-channel heat exchanger additive manufacturing process provided by the embodiment of the present application; Figure 2 is a micro-channel heat exchanger core structure parameter schematic diagram provided by the embodiment of the present application; Figure 3 is a micro-channel heat exchanger core structure parameter schematic diagram provided by the embodiment of the present application; Figure 4 is a meta-learning model training flow chart provided by the embodiment of the present application; Figure 5 is a comparison schematic diagram of the forming quality prediction accuracy of meta-learning provided by the embodiment of the present application; Figure 6 is a training curve schematic diagram of meta-reinforcement learning provided by the embodiment of the present application; Figure 7 is a best process parameter combination prediction schematic diagram of meta-reinforcement learning provided by the embodiment of the present application; Figure 8 is a suitability verification schematic diagram of the method of the present application provided by the embodiment of the present application. DETAILED DESCRIPTION
[0025] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art belong to the scope of protection of the present application.
[0026] Embodiment 1: AsFigures 1-8 The additive manufacturing process prediction and optimization method for microchannel heat exchangers based on meta-learning described in this embodiment includes: S1: Define the key geometric parameters of the microchannel heat exchanger and determine the SLM process parameter space.
[0027] In this embodiment, the structural parameterization of the microchannel heat exchanger is described as follows: The basic analytical unit of the microchannel heat exchanger is composed of periodic repeating features, including channels on both the hot and cold sides and the partition wall structure.
[0028] For example, in combination Figure 2 As shown, the cold side uses an equilateral square microchannel array, with the channel width defined as... Height is rib width is The hot side uses a high aspect ratio rectangular microchannel, with the channel width defined as... Height is rib width is A heat-insulating / heat-conducting wall is installed between the hot and cold aisles, with the following thickness parameters: and In this embodiment, , , , , , , , These geometric parameters constitute a complete system of structural variables, which can characterize various channel configurations such as symmetric / asymmetric, isomorphic / heterogeneous, including but not limited to circular, square and other parameterized irregular cross-sections.
[0029] S2: An experimental scheme for SLM process with typical dimensions is designed using an orthogonal array.
[0030] In this embodiment, based on the SLM process, an experimental design for process optimization was carried out for a parameterized microchannel heat exchanger structure. The laser power was selected for the experiment. Scanning speed Layer thickness and scanning spacing As a key process variable, and in conjunction with typical channel structure dimensions, including but not limited to 100 200 300 of , Parameter values were selected, a multi-factor experimental design was constructed, and the following methods were adopted. Alternatively, experiments can be arranged using an L16(4^5) orthogonal array to prepare microchannel samples with different combinations of structural parameters.
[0031] For example, scanning electron microscopy (SEM), metallurgical microscopy, laser confocal scanner, density meter, and other equipment are used to quantitatively characterize the dimensional accuracy and surface roughness of each sample. Key quality indicators such as value and flatness are used. Experimental data are used to establish the structural parameters of the microchannel, and the SLM process parameters and forming size deviation are analyzed. Surface roughness Flatness A database of mapping relationships between them.
[0032] S3: Experimental samples were prepared using SLM equipment, morphological data were acquired using laser confocal microscopy / SEM, and dimensions were automatically extracted based on U-Net+SE network.
[0033] In this embodiment, based on the sample prepared by the above-mentioned SLM process orthogonal experiment, image acquisition, preprocessing and deep learning modeling are carried out to construct an image dataset with structural labels, so as to realize automatic size extraction and shape parameter recognition.
[0034] For example, data acquisition and preprocessing are performed: Multiple high-resolution imaging techniques, such as SEM, metallurgical microscopy, and laser confocal scanner, are used to acquire images of typical cross-sections and surfaces of the sample. The resulting images have micrometer-level resolution, fully revealing details such as channel boundaries, rib structures, and surface melting traces. After image acquisition, image preprocessing operations are first performed, including but not limited to: image enhancement, such as contrast stretching and color normalization, to highlight channel edges; noise suppression using Gaussian filtering or bilateral filtering to reduce scanning artifacts; and edge smoothing and line fitting to improve the geometric stability of the structural contour. In the annotation stage, an intelligent annotation tool based on active learning is used. First, the algorithm generates initial contour suggestions for the channel region, then fine-tunes and confirms the boundary pixels, controlling the annotation error within a certain range. Within one pixel.
[0035] In this embodiment, a semantic segmentation network model based on the U-shaped convolutional neural network U-Net or the improved deep convolutional semantic segmentation network DeepLabV3+ is constructed. The model input is a pre-processed channel image, and the output is the category label (channel / non-channel) for each pixel, i.e., a binary segmentation mask. To improve the model's performance in microstructure edge recognition, an attention mechanism such as a compression-excitation module (SE module) is introduced into the network to enhance its responsiveness to channel boundary details. Training samples are derived from images corresponding to the aforementioned process experiments, and precise masks are obtained by combining manual annotations to construct a structural label set. During training: a weighted combination loss function of the Dice coefficient loss function and the binary cross-entropy loss function BCE is used; data augmentation techniques simulating rotation, scaling, and occlusion are applied to improve the model's adaptability to changes in structural size; and K-fold cross-validation is used to monitor the model's generalization ability under different channel structures. After training, the model can perform boundary recognition and segmentation on any given micro-channel image.
[0036] For example, geometric parameters are extracted. Using the mask image output by the semantic segmentation network, contour tracking and boundary fitting algorithms are used to extract the closed boundaries of the channel structure, and then the following key structural parameters are calculated: area. : Calculated based on the number of pixels within the channel area combined with resolution; Perimeter Calculated based on the path length of the boundary pixels; hydraulic diameter According to the formula Computation; the mask image output by the semantic segmentation network, such as Figure 3 As shown, the matching error between the channel structure boundary extracted by the contour tracking algorithm and the actual contour is ≤5. The area is automatically calculated based on this boundary. ,perimeter and hydraulic diameter Key parameters such as geometry parameters are determined, and a mapping relationship is established between these geometric parameters and the corresponding SLM process parameters.
[0037] S4: Construct a meta-task set of process-molding quality, use the MAML++ framework to integrate simulation data for training, and establish a meta-learning process prediction model.
[0038] In this embodiment, to achieve efficient prediction and generalization of the forming quality of channel structures in SLM additive manufacturing under small sample conditions (≤5 sets of experimental data / size), an improved MAML++ training method integrating multi-task prior knowledge is proposed. Compared with traditional single-task training and data-driven modeling machine learning methods, this method achieves implicit knowledge transfer and rapid adaptation across size tasks through a meta-learning framework. With only a very small number of experimental samples, it can accurately predict the forming size error of new-size microchannel structures (size variation range of ±60%) under different process parameters. Surface roughness Flatness The prediction efficiency is more than 3 times higher than that of traditional methods, and the experimental cost is reduced by 80%.
[0039] S41: Define the meta-task and perform cross-size generalization modeling; In this embodiment, a multi-scale task is constructed: by combining small-sample orthogonal experimental design, including... , The combination of process parameters under each set of dimensional parameters and its molding result constitutes a task. For example, task sample pairs are defined as follows: Input: A vector of process parameters normalized by Z-score. With structural parameters [ , ]; Output: the corresponding sample ,Right now: , ] .
[0040] Cross-size task distribution constraints: Experimental tasks across multiple structural sizes constitute a meta-training set, which must satisfy the following requirements: process parameter coverage ≥ 85%, ensuring sufficient exploration of the process space, and a structural size span ratio ≥ 1:3. It enhances the ability to generalize across dimensions, covering various sizes and process disturbances, and reflects the typical variability in the additive manufacturing process.
[0041] In this embodiment, a cross-size feature decoupling and sharing mechanism is implemented; a size-aware embedding module is introduced into the input layer to embed structural parameters [ , The feature vector is mapped to a low-dimensional feature vector and concatenated with process parameter features to enhance the model's sensitivity to size changes. A task-adaptive feature extractor is employed, which dynamically adjusts network weights through a gating mechanism to distinguish between process-related features and size-related features.
[0042] S42: Meta-learning training is performed using an improved MAML++ framework.
[0043] In this embodiment, combined with Figure 4 As shown, an improved MAML++ framework is adopted, using a 4-layer MLP base regression network, embedding improvements including task-specific batch normalization and cross-size residual connections. In task-specific batch normalization, an independent BN statistic is maintained for each size task to eliminate distribution differences between sizes; in cross-size residual connections, size-conditional residual paths are added between hidden layers to promote cross-size knowledge transfer.
[0044] For example, hierarchical optimization strategies based on inner-layer optimization and outer-layer optimization respectively: Inner layer optimization: Perform fast gradient updates using 3-5 sets of process parameters in a single task with a fixed size of 200 μm, for example; employ a hierarchical learning rate strategy, such as a feature extraction layer. Return to output layer This yields a task-specific model; at the same time, gradient clipping is introduced to prevent overfitting of small samples.
[0045] Outer layer optimization: Based on the loss fine-tuned across multiple tasks, implicit second-order optimization is used to accelerate the meta-parameter optimization. Updated. A task difficulty-aware weighting method is adopted, assigning higher weights to tasks with large size variations to enhance cross-size generalization capabilities.
[0046] For example, multi-source data fusion training includes finite element simulation data augmentation and the design of a confidence weighting mechanism; finite element simulation data augmentation mixes simulation data and experimental data in a 7:3 ratio and aligns feature distributions through domain adaptation loss; the design of a confidence weighting mechanism reduces the impact of errors in simulation data in critical size regions.
[0047] S43: Perform cross-size generalization verification and adaptive adjustment; In this embodiment, during the meta-testing phase, unseen dimensions are directly predicted, such as dimensions with ±60% variation, to evaluate the model's generalization performance under zero-sample conditions. First, single-sample online adaptation is performed. If a new set of experimental data for a specific size is provided, model adaptation is completed within 10 seconds through online meta-fine-tuning, supporting dynamic process adjustments. Then, the accuracy of cross-size generalization is verified. Figure 5 As shown, for a new structural size, for example Cold side passage For the hot-side rectangular channel, the following steps are implemented: Initial data is obtained through 2-3 sets of process parameter experiments; the bottom-level parameters of the network are frozen, and only the top-level regression layer is fine-tuned with a learning rate of 0.005; the model is updated using online incremental learning. This is the forgetting factor; it outputs the process parameters for any given size. Predicted value.
[0048] In this embodiment, a comparison between meta-learning and traditional methods yields the following results: (1) Comparison with the all-factor orthogonal experimental method: The process optimization method based on meta-learning proposed in this invention has significant differences from the traditional orthogonal experimental method in the additive manufacturing optimization of microchannel heat exchangers. The specific comparison is as follows: Traditional orthogonal experimental method: This method involves using process parameters... With structural dimensions Treated as parallel optimization variables, a full parameter space search is required. Orthogonal arrays and other experimental design methods require at least 18 experiments per fixed size group, and can only obtain a local optimum for that size. Different sizes require independent experiments, making it impossible to generalize across sizes; the number of experiments increases linearly with the number of sizes, resulting in high optimization costs.
[0049] This invention's meta-learning method treats structural dimension parameters as task descriptors, rather than optimization variables, and utilizes the MAML++ algorithm to learn the cross-task mapping rules of dimension-process-performance. A meta-model is established through a small number of basic experiments (18 runs), and new dimensions can be quickly adapted with only 2-3 validation experiments; this achieves cross-size prediction capabilities and avoids redundant experiments.
[0050] The experimental costs are compared in Table 1. Optimizing five different sizes of microchannel heat exchangers: the traditional method requires 90 experiments (18×5); the method of this invention requires only 33 experiments (18 basic experiments + 3×5 adaptation experiments), reducing the experimental cost by 63.3%, while also having cross-size prediction capabilities.
[0051] Table 1. Comparison of the number of experiments between the traditional orthogonal experimental method and the meta-learning method of this invention.
[0052] (2) Comparison with classic machine learning methods: such as SVR and random forest: Data requirements: require a large amount of labeled data, usually 15-20 sets of data per size; Feature engineering: relies on manually designed features, and it is difficult to automatically capture the non-linear relationship between size and process; Generalization performance: the prediction error on new sizes is usually >20%, and the model needs to be retrained; Adaptability: cannot achieve rapid adaptation to small samples, and a complete dataset is required for new sizes. (3) Comparison with deep learning methods: such as DNN deep learning methods: Data efficiency: requires at least 10-15 sets of data for each size to ensure basic accuracy; Training method: uses fixed datasets for end-to-end training, which cannot achieve task-oriented learning; Transfer ability: requires complete fine-tuning on new sizes, usually requiring 5-8 sets of data; Computational cost: large number of parameters, long training time, not suitable for online optimization scenarios. This invention's meta-learning method innovatively uses structural size parameters as task descriptors and leverages the MAML++ algorithm to learn cross-task mapping patterns; data efficiency: a meta-model is established through a small number of basic experiments (18), and only 2-3 validation experiments are needed for new sizes; generalization ability: cross-size prediction is achieved (error <9.3%), avoiding repeated experiments; adaptation mechanism: supports rapid online adaptation for small samples (completed within 1 day); cost advantage: the number of experiments follows a "constant term + low linear growth" (O(1) + O(n)) pattern; the experimental cost comparison is shown below. Optimization of 5 different sizes of microchannel heat exchangers: Traditional orthogonal method: 90 experiments (18×5); Classical ML method: 75-100 experiments (15-20 × 5); Deep learning method: 50-75 experiments (10-15×5); The method of this invention consists of 33 experiments (18 basic experiments + 3×5 adaptation experiments).
[0053] S5: Establish the state / action space of the DDPG algorithm, design a multi-objective reward function, and train the Actor-Critic network to output the optimal process and molding quality parameter set.
[0054] In this embodiment, in order to achieve the desired molding accuracy for a microchannel heat exchanger with a specific structural size (…), Design values), surface quality Flatness Under the premise of quickly obtaining the optimal combination of SLM additive manufacturing process parameters, this invention proposes a process parameter optimization method that combines a meta-learning cross-size prediction model with an intelligent optimization strategy. This method achieves rapid reverse optimization of process parameters by combining the cross-size generalization ability of the meta-learning model with the efficient search capability of the intelligent optimization algorithm.
[0055] For example, based on the constructed meta-learning process prediction model, the target process optimization task is formalized as a multi-objective optimization problem. The optimization objective is to minimize the molding size error. Surface roughness Flatness .
[0056] Its mathematical expression is: ; Among them, weight Furthermore, the sum of these values equals 1, and adjustments can be made based on actual project requirements.
[0057] In this embodiment, the present invention employs a three-level optimization architecture to achieve intelligent optimization of process parameters: Meta-model cross-size prediction layer: Loads a pre-trained MAML++ model, including task-specific batch normalization, and inputs process parameters. With structural dimensions [ , ], output The predicted values and their 95% confidence intervals. Cross-size generalization optimization: The model parameters are dynamically adjusted through a task-specific BN layer to adapt to different input size distributions, ensuring that the prediction accuracy remains stable within the size variation range (±60%).
[0058] The intelligent process optimization strategy layer offers three optional intelligent process optimization methods to adapt to different application scenarios. The gradient-guided finite-memory BFGS Newton algorithm is recommended for fast convergence scenarios. A Nesterov acceleration strategy is employed, combining the Hessian matrix approximation from the meta-model output. The meta-optimization search method is recommended for multi-task generalization scenarios. A two-layer LSTM supernetwork is used: the bottom LSTM generates the process parameter distribution, and the top LSTM predicts the optimization trajectory. Reinforcement learning is the recommended implementation method for complex constraint scenarios. An actor-critic dual-network DDPG architecture is used in conjunction with the meta-learning model, where the actor network outputs process parameter adjustments, and the critic network uses the meta-model as an environment simulator. The meta-model is used as a differentiable environment simulator; a multi-scale reward function is designed to balance short-term optimization objectives and long-term stability; and a process parameter mutation penalty mechanism is incorporated to avoid parameter oscillations.
[0059] Online update layer: Updates the meta-model parameters using new experimental data obtained during the optimization process through incremental learning. The update rules are as follows: ; Where λ=0.9 is the forgetting factor, and Dnew is the distribution of newly added data.
[0060] Dynamic adaptation mechanism: Through online updates, the metamodel can continuously adapt to the process characteristics of new sizes, improving long-term generalization ability.
[0061] For example, the implementation path of using the DDPG algorithm as a smart process optimization method includes: (1) Environment setup: State space (S): Current process parameters: [P, v, t, h] (normalized to [0, 1]); History window: the last 5 steps. Rate of change; Structural dimensional characteristics: (Scale normalization); Motion space (A): Continuous parameter adjustment amount: , , , ; Reward function design: ; in, For quality weight, This represents the penalty coefficient for changes in action.
[0062] Additional bonus: If all conditions are met and Reward +1.0.
[0063] (2) Network architecture design Actor Network (Policy Network): Input layer: State s (dimension 10); Hidden layer: 3 fully connected layers (256→128→64), ReLU activation; Output layer: Tanh activation, scaled to the action range; Critic Network (Value Network): Input layer: concatenation of state s and action a (dimension 14); Hidden layers: 4 fully connected layers (512→256→128→32), LeakyReLU (α=0.2); Output layer: linear output Q-value; 3) Training process Meta-model initialization: Load the pre-trained MAML++ model as an environment simulator.
[0064] Exploration phase: The Actor network explores in the multi-scale environment generated by the meta-model and collects optimized trajectories.
[0065] Strategy optimization: The Critic network evaluates the value of actions and guides the update of Actor network parameters.
[0066] Online adaptation: The meta-model is updated with new data increments obtained during the optimization process to achieve dynamic adaptation.
[0067] The detailed training process is shown in Table 2.
[0068]
[0069] 4) Training curve The convergence process of the DDPG algorithm described in this invention exhibits a three-stage characteristic, as follows: Figure 6 As shown, its typical training curves satisfy the following rules: Exploration phase (0-2000 rounds): The average reward value Rave increases rapidly, but fluctuates significantly, with a standard deviation σ ≥ 0.8. This is because the Actor network is dominated by random exploration, while the Critic network has not yet established an accurate Q-value estimate. Stable improvement phase (2000-4500 rounds): The reward curve shows a logarithmic growth trend, and the fluctuation of process parameters decreases, indicating that the value assessment tends to stabilize; Convergence Phase (>4500 rounds): Reward values plateau (fluctuation range ±2%), meeting convergence conditions. ; in, and They are the first round and the The reward value of a round indicates that the strategy's determinism has reached its optimal level.
[0070] 5) Comparison of optimization results: like Figure 7 As shown, the cross-size intelligent prediction optimization method based on the DDPG algorithm constructed in this invention exhibits high-efficiency convergence characteristics within a 10-second simulation cycle. The fully trained Actor network can output the optimal combination of process parameters that satisfies multi-objective constraints in just 0.4 seconds, which is two orders of magnitude faster than the convergence speed of traditional optimization algorithms (such as genetic algorithms).
[0071] like Figure 8 As shown, the optimization results of the cross-size intelligent prediction optimization method based on the DDPG algorithm on the meta-learning model are described. In a simulation cycle of 10s, the trained Actor network obtains the optimal molding quality parameters within 0.4s.
[0072] As shown in Table 3, the improvement effect of the optimized process on the microchannel forming quality was quantified: the dimensional accuracy Δd was improved most significantly (reduction of 40.5%); the surface roughness Ra was reduced by 41.3%; and the flatness ρ was improved by 30.4%.
[0073] Table 3 Molding Quality Indicators
[0074] S6: Construct a closed-loop verification system for intelligent optimization of process parameters.
[0075] In this embodiment, process optimization is transformed into a multi-objective optimization problem (minimization). , ,maximize The system combines a meta-optimizer to efficiently search for optimal parameters in a virtual environment; it also uses image recognition feedback to adjust the process in real time, forming a closed loop of "identification-prediction-optimization-verification" to ensure forming accuracy and consistency.
[0076] Example 2 The present invention can be implemented in ways including but not limited to the following types and specific implementation schemes: (1) Image recognition implementation method for size recognition Multimodal fusion-based approach: Combining optical microscopy and X-ray tomography data, feature-level fusion is used to improve the recognition accuracy of complex structures.
[0077] Lightweight network: It adopts a lightweight architecture such as Mobile Network Version 3 (MobileNetV3) to reduce computing resource consumption while maintaining accuracy, making it suitable for deployment in embedded devices.
[0078] Self-supervised learning: The feature extractor is pre-trained using unlabeled data, reducing the reliance on manual annotation.
[0079] (2) Meta-model implementation for small sample process prediction Based on graph neural networks: process parameters and structural features are constructed as graph data, and graph neural networks (GNNs) are used to capture the topological relationships between parameters.
[0080] Temporal prediction-based: The temporal dependence of process parameters is modeled using LSTM or Temporal Convolutional Network (TCN), which is suitable for the optimization of continuous manufacturing processes.
[0081] Ensemble learning: Combining the predictions of multiple base learners and improving generalization ability under small sample sizes through voting or weighted averaging.
[0082] (3) Implementation methods of process optimization Evolutionary strategies, such as the Covariance Matrix Adaptation Evolution Strategy (CMA-ES), perform a global search in the parameter space to avoid getting trapped in local optima.
[0083] Based on physical model assistance: a hybrid-driven optimization model is constructed by integrating finite element simulation data and experimental data.
[0084] Federated learning-based: Sharing model updates rather than raw data among distributed devices enables collaborative optimization across multiple devices.
[0085] The above implementation methods can be applied individually or in combination, and can be adapted to different manufacturing scenarios through modular design.
[0086] Example 3
[0087] In this embodiment, the invention may undergo several design changes or modifications during implementation to adapt to different application requirements or optimization goals: (1) Alternative implementation of meta-learning framework: Meta-learning methods based on memory networks, such as Meta-LSTM, can be used to replace the MAML++ framework to enhance the ability to memorize long-term process knowledge; probabilistic meta-learning methods, such as Gaussian process meta-learning, can be introduced to provide uncertainty assessment for process prediction.
[0088] (2) Alternative solutions for optimization algorithms: adopt a hybrid architecture of evolutionary algorithm and meta-learning, combining the advantages of global search and rapid adaptation; use distributed meta-learning methods to achieve multi-device collaborative optimization while protecting data privacy.
[0089] (3) Alternative design of image recognition module: The visual Transformer architecture is used to replace the convolutional neural network (CNN) to improve the cross-scale feature extraction capability; a self-supervised learning method is introduced to reduce the dependence on labeled data.
[0090] In this embodiment, the extended application scenario includes: (1) Multi-process and multi-material manufacturing applications: high-precision manufacturing applications such as micro-electrical discharge machining, chemical etching, and diffusion welding; extended model supports process optimization of multi-system materials such as metal-ceramic composite materials; added material coding vectors to realize knowledge transfer across material categories.
[0091] (2) Special structure manufacturing: adapt to the customized manufacturing needs of non-uniform microchannels such as gradient pore structure; support the optimization of process parameters for irregular cross-section channels.
[0092] In this embodiment, the optimization direction of the system architecture (1) Digital twin integration: Develop a closed-loop optimization system that combines virtual and real data; realize real-time interactive iteration between prediction models and physical experiments.
[0093] (2) Multi-fidelity optimization: Integrate simulation data and experimental data; establish a hierarchical optimization framework to improve the efficiency of initial optimization; Note: All variant solutions retain the core innovative features of "small sample learning + intelligent optimization" and support the needs of different industrial scenarios through modular design.
[0094] Table 4. Comparison of Chinese and English Professional Terms
[0095] This invention presents a cross-size few-shot learning system based on an improved MAML++ framework, achieving process quality prediction within a size variation range of ±60%. It innovatively organizes process experimental data from micro-channels of different sizes into independent tasks; employs second-order optimization and task-specific batch normalization to improve prediction stability; and after loading the pre-trained model, it rapidly fine-tunes and adapts it using 2-3 sets of new experimental data. The optimizer architecture includes an Actor-Critic network design and a multi-scale reward function; it first predicts the process impact through a meta-model, then uses reinforcement learning to output the optimal parameters; and it features an online update mechanism that feeds new experimental data back to the prediction model through incremental learning.
[0096] It should be noted that, in this invention, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.
[0097] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for predicting and optimizing the process of additive manufacturing of micro-channel heat exchangers based on meta-learning, characterized by: The method comprises the following steps: S1: defining key geometric parameters of the micro-channel heat exchanger, and determining an SLM process parameter space; S2: designing an SLM process experiment scheme containing multiple sizes by using an orthogonal table; S3: preparing an experimental sample by using an SLM device, acquiring morphology data by using a laser confocal and a SEM, and automatically extracting size parameters based on a U-Net+SE network; S4: constructing a process-molding quality meta-task set, training by using a MAML++ framework to fuse simulation data, and establishing a meta-learning process prediction model; The meta-learning process prediction model is: taking the structure size parameters as task descriptors, learning the cross-task mapping rule by using the MAML++ algorithm; establishing a meta-model by using a small amount of basic experiments, realizing cross-size prediction, and supporting small sample online rapid adaptation; S5: establishing a state and action space of a DDPG algorithm, constructing a multi-objective reward function, training an Actor-Critic network to output optimal process and molding quality parameters; S6: constructing a closed-loop verification system to verify the optimal process parameters output by steps S2-S5, and intelligently optimizing and cyclically reinforcing learning based on the verification result.
2. The meta-learning based micro-channel heat exchanger additive manufacturing process prediction optimization method of claim 1, wherein: The key geometric parameters of the micro-channel heat exchanger include: channel width, height and rib width of the cold and hot channels, and thickness of the heat insulation / heat conduction wall arranged between the cold and hot channels.
3. The meta-learning based micro-channel heat exchanger additive manufacturing process prediction optimization method of claim 1, wherein: The SLM process experiment scheme is specifically as follows: Based on the SLM process, laser power, scanning speed, layer thickness and scanning spacing are selected as key process variables, and a multi-factor experiment scheme is constructed in combination with typical channel structure sizes, including but not limited to Channel width parameter values of the cold and hot two sides channels An experiment is designed by using an orthogonal table to prepare micro-channel samples with different structure parameter combinations; A mapping relationship database between the micro-channel structure parameters, the SLM process parameters and the forming size deviation, the surface roughness and the flatness is established by using experimental data.
4. The meta-learning based micro-channel heat exchanger additive manufacturing process prediction optimization method of claim 1, wherein: The acquisition of the morphology data comprises: SEM, metallographic microscope and laser confocal scanner are used to collect images of typical cross sections and surfaces of the samples, and image preprocessing operations are performed on the collected images; a semantic segmentation network model based on a U-shaped convolutional neural network or an improved deep convolutional semantic segmentation network is constructed, the preprocessed channel images are trained, and a mask image is output; the closed boundary of the channel structure is extracted from the output mask image by using a contour tracking and boundary fitting algorithm, and key structure parameters including area, perimeter and hydraulic diameter are calculated; the key structure parameters are associated with the corresponding SLM process parameters to establish a mapping relationship.
5. The meta-learning based micro-channel heat exchanger additive manufacturing process prediction optimization method of claim 1, wherein: S4 specifically comprises: S41: building a multi-scale task, decoupling cross-size features and sharing mechanisms; S42: meta-learning training is performed by using an improved MAML++ framework; S43: cross-size generalization verification and self-adaptive adjustment are performed.
6. The meta-learning based micro-channel heat exchanger additive manufacturing process prediction optimization method of claim 5, wherein: The improved MAML++ framework is specifically as follows: An improved 4-layer MLP is used as a basic regression network, and a task-specific batch normalization and cross-size residual connection are embedded; a hierarchical optimization strategy including inner-layer optimization and outer-layer optimization is adopted; multi-source data fusion training is performed.
7. The meta-learning based micro-channel heat exchanger additive manufacturing process prediction optimization method of claim 5, wherein: The cross-size generalization verification comprises: initial data are obtained by using 2-3 groups of process parameter experiments; the network bottom layer parameters are frozen, and only the top regression layer is fine-tuned; the model is updated by using online incremental learning; prediction values of the forming size deviation, the surface roughness and the flatness under any process parameters of this size are output.
8. The meta-learning based micro-channel heat exchanger additive manufacturing process prediction optimization method of claim 1, wherein: The output optimal process and forming quality parameter group comprises: based on the constructed meta-learning process prediction model, the target process optimization task is formalized as a multi-objective optimization problem, specifically: The optimization target of minimizing the forming size error, surface roughness and flatness is constructed; A three-level optimization architecture including meta-model cross-size prediction layer, intelligent process optimization strategy layer and online updating layer is adopted for intelligent optimization of process parameters.
9. The meta-learning based micro-channel heat exchanger additive manufacturing process prediction optimization method of claim 1, wherein: The output optimal process and forming quality parameter group also comprises: taking DDPG algorithm as the implementation path of intelligent process optimization method, specifically: An implementation environment is constructed including state space, action space, reward function, Actor network as policy network and Critic network as value network; The cross-size intelligent prediction optimization method based on DDPG algorithm is trained; Finally, the trained curve law is used to determine the strategy certainty to achieve optimization.
10. The meta-learning based micro-channel heat exchanger additive manufacturing process prediction optimization method of claim 9, wherein: The training process is: Load the pre-trained MAML++ model as an environment simulator; Use the Actor network to explore in the multi-size environment generated by the meta-model, and collect the optimization trajectory; Use the Critic network to evaluate the action value and guide the Actor network parameter update; Incrementally update the meta-model with new data obtained during the optimization process to achieve dynamic adaptation.
Citation Information
Cited By
Multi-stream plate-fin heat exchanger parametric design and manufacturing method
CN121936289A
