LPBF breathable steel process parameter optimization method based on deep learning

Abstract

The invention belongs to the technical field of additive manufacturing process optimization and material performance regulation and control, and relates to a deep learning-based LPBF breathable steel process parameter optimization method, which comprises the following steps of: acquiring and preprocessing LPBF process data and simulated physical field data; constructing a multi-target prediction model based on deep learning; solving process parameters based on a multi-objective optimization algorithm; and performing closed-loop feedback iteration. And adopting Latin hypercube design to generate physical experiment parameters and simulation parameters. Simulation parameters are corrected through a small amount of experimental data, and the deviation between simulation and experimental data is smaller than or equal to 8%; the porosity, the aperture size and the ventilation coefficient of the ventilation steel are set according to requirements, the technological parameter constraint range is set according to LPBF equipment, a Pareto solution is obtained through an NSGA-III algorithm, and the solution with the highest score is selected as the optimal solution through weighted scoring. The optimal parameters are adopted for closed-loop verification, the prepared breathable steel sample is used for testing target performance, and the relative error of each index is smaller than or equal to + / -5%.

Description

Technical Field

[0001] This invention relates to the field of additive manufacturing process optimization and material performance control technology, and in particular to a deep learning-based method for optimizing process parameters of LPBF (Laser Powder Bed Fusion) permeable steel. This method integrates process parameters with multi-physics data and is especially suitable for precise parameter control in the preparation of customized permeable steel using LPBF (or SLM, Selective Laser Melting) process. Background Technology

[0002] As a special metal material containing a large number of interconnected or semi-interconnected pores, breathable steel has the characteristics of good air permeability, low density, and high specific strength. It is widely used in fields such as biomedicine, aerospace and injection mold venting systems. Its main performance indicators, such as porosity, pore size and air permeability coefficient, directly determine the application effect.

[0003] Currently, the main manufacturing processes for permeable steel include foaming agent forming, powder metallurgy, and spark plasma sintering. However, these traditional processes have significant limitations: they are difficult to manufacture parts with complex shapes, and the pore structure is difficult to control precisely, severely restricting the widespread application of permeable steel. LPBF (Layer-Plate Forming) technology, with its rapid melting and layer-by-layer printing characteristics, can achieve the fabrication of near-net-shape metal parts with complex structures and excellent mechanical properties. Furthermore, it can achieve connectivity and controllability of the pore structure by adjusting process parameters, providing a new pathway for the preparation of permeable steel.

[0004] However, the LPBF process for preparing permeable steel faces key technical challenges: 1. There is a strong nonlinear relationship between process parameters (laser power, scanning speed, etc.) and multiple physical fields and performance indicators. Traditional trial-and-error methods require hundreds of physical experiments, with each experiment costing over 10,000 yuan and optimization cycles lasting 3-6 months, and cannot cover the entire parameter space; 2. The influence of multiple physical fields on pore structure is difficult to quantify, leading to blind parameter optimization and often resulting in problems such as "porosity meets standards but permeability is insufficient" and poor synergy among multiple performance indicators; 3. Existing technologies have not formed a closed-loop solution of "data acquisition - model prediction - parameter optimization - feedback iteration," which cannot meet the customized needs of different scenarios.

[0005] Deep learning technology possesses powerful nonlinear mapping and multi-objective fitting capabilities, demonstrating its potential in additive manufacturing process optimization. However, current technologies have not yet deeply integrated deep learning with multi-objective optimization of LPBF process for preparing permeable steel, particularly lacking closed-loop optimization schemes that integrate multiphysics data and support real-time correction, thus failing to meet the customized, high-precision, and high-efficiency preparation requirements of permeable steel. Therefore, there is an urgent need to propose a deep learning-based multi-objective optimization method for LPBF process parameters to address the aforementioned technical challenges.

[0006] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of the present invention, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0007] To address the problems of low parameter optimization efficiency, high trial-and-error costs, poor synergy of multiple properties, and lack of customized closed-loop solutions in the existing LPBF process for preparing permeable steel, this invention provides a deep learning-based method for optimizing LPBF permeable steel process parameters. By integrating process parameters with multi-physics field data, a high-precision multi-objective prediction model is constructed, forming a closed-loop system of "data acquisition - model training - parameter optimization - feedback iteration," thereby reducing trial-and-error costs and improving the yield rate and generalization ability of parameter optimization.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A deep learning-based method for optimizing process parameters of LPBF permeable steel includes the following steps: Step S1: Acquisition and preprocessing of LPBF process data and simulation physics data; Step S2: Constructing a multi-objective prediction model based on deep learning; An improved multi-input multi-output neural network model consisting of convolution, attention, and residual connections is adopted. The input layer features include laser power, scanning speed, scanning spacing, layer thickness, peak temperature of the molten pool, cooling rate, peak thermal stress, and strain distribution. The feature extraction layer is set with 3×3, 5×5, and 3×3 convolution kernels, and the number of output channels is 64, 128, and 256 respectively. A multi-head attention layer is also adopted. Gradient vanishing is avoided by residual connections. The output features include porosity, pore size, and air permeability coefficient. The normalized predicted values ​​are output using the Sigmoid activation function. Initial hyperparameters were determined using a Bayesian optimization algorithm and adjusted through cross-validation. The AdamW optimizer was used with a weighted cross-entropy loss function. Overfitting was suppressed by L2 regularization. Training was stopped and the optimal model was saved when the relative errors of porosity, pore size, and air permeability prediction on the validation set were all ≤ ±5%. Step S3: Solving for process parameters based on multi-objective optimization algorithm; Based on the performance of LPBF equipment and material properties, process parameter boundaries are set; A custom target range is defined, and a judgment matrix is ​​constructed using the analytic hierarchy process. The target layer is "multi-performance optimality", and the criteria layer consists of porosity, pore size, and air permeability. The weight vector is determined through a consistency test. The process parameter space is sampled 1000 times based on Latin hypercube sampling to eliminate obviously unreasonable parameter combinations; the effective parameter combinations are input into the trained prediction model, the prediction performance index is output, and the Pareto front solution is obtained by optimizing it using the NSGA-Ⅲ algorithm. The weighted score of each Pareto solution is calculated based on the weight vector determined by AHP. The solution with the highest score is selected as the unique optimal combination of process parameters, and the specific parameter values ​​are output. Step S4: Closed-loop feedback iteration; Input the optimal process parameters into the LPBF equipment to conduct printing experiments and test the actual performance indicators. If the relative error between the actual performance and the predicted value is ≤ ±5%, the corresponding parameter combination is the final solution. If the error is > ±5%, the corresponding parameter combination data is added to the original dataset, the model is retrained, and step S3 is repeated until the error meets the requirements, thus forming a closed-loop optimization.

[0009] The following is a further defined technical solution of the present invention, wherein step S1 includes: Laser power, scanning speed, scanning spacing, and layer thickness, which are core parameters of the LPBF process, were selected as basic input parameters. A Latin hypercube design experimental scheme with uniform parameter space coverage and strong sample representativeness was adopted, and the parameter range was set. The thermal-stress coupling simulation software was used to generate multiphysics data in batches based on process parameters, including peak temperature of the molten pool, cooling rate, peak thermal stress, and strain distribution. The initial parameters of the simulation model were referenced from the thermal conductivity, specific heat capacity, and coefficient of thermal expansion of the material. Conduct calibration experiments to verify the validity of the simulation data; Data preprocessing includes removing outliers, standardizing input features, and normalizing output performance metrics. The SMOTE algorithm is used to expand small sample data, physical consistency is checked on the synthetic samples, redundant samples with similarity ≥95% to the original samples are removed, and random noise is added to the standardized input and output data to improve the model's generalization ability. The final labeled dataset is formed.

[0010] The following is a further technical solution of the present invention to verify the validity of the simulation data, including: selecting 20 typical process parameter combinations to conduct LPBF printing experiments, measuring porosity using the Archimedes drainage method, obtaining pore size through image analysis using a scanning optical microscope, and testing the gas permeability coefficient using a gas permeation test device.

[0011] The following is a further defined technical solution of the present invention: the 3σ criterion is used to remove abnormal data caused by simulation convergence failure and experimental equipment error; the input features are standardized by Z-score; and the output performance index is normalized by Min-Max and mapped to the [0,1] interval to eliminate the influence of dimensional differences on model training.

[0012] The following is a further defined technical solution of the present invention: the labeled dataset includes physical experimental data, calibrated simulation data, SMOTE synthetic data and noise enhancement data, and the data format is a one-to-one correspondence between input feature vectors and output performance vectors.

[0013] The following is a further defined technical solution of the present invention: the number of attention heads in the multi-head attention layer is 8, and the weight decay coefficient of the AdamW optimizer is 0.01.

[0014] The following is a further defined technical solution of the present invention: the population size of the NSGA-Ⅲ algorithm is 100, and the number of iterations is 200.

[0015] The following is a further defined technical solution of the present invention: the weighted score is proportional to the product of the performance prediction value and its corresponding weight.

[0016] Compared with the prior art, the present invention has the following technical effects: The physical experimental parameters and simulation parameters are generated using Latin hypercube design. The simulation parameters are corrected using a small amount of experimental data to ensure that the deviation between the simulation and experimental data is ≤8%. Outliers are removed, and a small sample dataset is synthesized using the SMOTE algorithm with added random noise to form a valid dataset.

[0017] For multi-input multi-output neural network models, the input layer dimension, convolutional kernel size, number of output channels, and multi-head attention layers are set, and gradient vanishing is avoided through residual connections. The ReLU activation function is used to introduce non-linearity, and the Sigmoid activation function outputs normalized predictions.

[0018] Based on the requirements, the porosity, pore size, and permeability coefficient of the permeable steel are set. The process parameter constraints are set according to the LPBF equipment. The Pareto solution is obtained using the NSGA-III algorithm, and the solution with the highest weighted score is selected as the optimal solution. Closed-loop verification is performed using the above optimal parameters. Permeable steel samples are prepared and the target performance is tested. If the relative error of each index is ≤±5%, the requirements are met.

[0019] The present invention will be further described below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic diagram of the closed-loop optimization process of the present invention; Figure 2 This is a schematic diagram of the improved neural network structure of the present invention. Detailed Implementation

[0022] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0023] like Figure 1 and 2 As shown, a deep learning-based method for optimizing process parameters of LPBF permeable steel is provided, including the following steps: Step 1: Acquisition and preprocessing of LPBF process data and simulation physics data; 1.1 Select the core parameters of the LPBF process as the basic input parameters and set the parameter range (laser power: 50-500W, scanning speed: 500-2000mm / s, scanning spacing: 50-150μm, layer thickness: 20-60μm). Adopt the Latin hypercube design experimental scheme with uniform parameter space coverage and strong sample representativeness.

[0024] 1.2. Using thermal-stress coupling simulation software (such as ANSYS, Workbench, ABAQUS, etc.), multi-physics field data such as peak temperature of molten pool, cooling rate, peak thermal stress, and strain distribution are generated in batches based on the above process parameters. The initial parameters of the simulation model are the basic properties of the reference material (such as Corrax stainless steel) such as thermal conductivity, specific heat capacity, and coefficient of thermal expansion.

[0025] 1.3. Conduct calibration experiments to verify the validity of simulation data. Twenty typical process parameter combinations were selected for LPBF printing experiments. Porosity was measured using the Archimedes displacement method. Pore size (statistical range 10-200 μm) was obtained through image analysis using a scanning optical microscope (magnification 50-200x). Gas permeability was tested using a gas permeation testing device (compliant with GB / T 1966-2005 standard). ).

[0026] 1.4 Data Preprocessing. The 3σ criterion is used to remove abnormal data caused by simulation convergence failures or experimental equipment errors (such as invalid samples with porosity <5%). Input features are standardized using Z-score (mean = 0, variance = 1), and output performance indicators (porosity, permeability coefficient, etc.) are normalized to the [0,1] interval using Min-Max normalization to eliminate the influence of dimensional differences on model training.

[0027] 1.5. The SMOTE algorithm (with 5 nearest neighbors, a 1:2 ratio of synthesized samples, and a boundary constraint threshold of 0.8) is used to expand the small sample data. Physical consistency checks are performed on the synthesized samples to ensure that process parameters are within the equipment's allowable range and that the multiphysics data conforms to thermodynamic laws. Redundant samples with ≥95% similarity to the original samples are removed. Random noise (noise amplitude ≤ 5% of the original data fluctuation range, determined based on the 3σ criterion) is added to the standardized input and output data to improve the model's generalization ability.

[0028] 1.6 The final result is a high-quality labeled dataset of ≥1500 sets, consisting of "physical experimental data (200 sets) + calibrated simulation data (1000 sets) + SMOTE synthetic data (200 sets) + noise enhancement data (100 sets)".

[0029] Step 2: Constructing a multi-objective prediction model based on deep learning; 2.1 An improved multi-input multi-output neural network model using convolution, attention, and residual connections is employed. Input layer features include laser power, scanning speed, scanning spacing, layer thickness, peak molten pool temperature, cooling rate, and peak thermal stress. Feature extraction layers use 3×3, 5×5, and 3×3 convolutional kernels, with output channels of 64, 128, and 256 respectively. A multi-head layer (8 attention heads) is also included. Residual connections prevent gradient vanishing. Output feature vectors for porosity, pore size, and permeability coefficient are generated, and normalized predicted values ​​are output using the Sigmoid activation function.

[0030] 2.2 Initial hyperparameters (learning rate = 0.001, batch size = 32, epoch = 200) were determined using a Bayesian optimization algorithm and adjusted through 5-fold cross-validation. The AdamW optimizer (weight decay coefficient = 0.01) was used, with a weighted cross-entropy loss function (porosity, pore size, and air permeability weights were initially 1, adjustable as needed). L2 regularization (λ = 0.0001) was applied to suppress overfitting. Training was stopped and the optimal model was saved when the relative errors of porosity, pore size, and air permeability predictions on the validation set were all ≤ ±5%.

[0031] Step 3: Solving for process parameters based on multi-objective optimization algorithm; 3.1 Based on the performance of the LPBF equipment and the material properties, set the process parameter boundaries (e.g., laser power ≤ 300W, scanning speed ≥ 500). wait).

[0032] 3.2. Define a target range (e.g., porosity 10%-40%, pore size 10-70μm, air permeability coefficient ≥0.05). The analytic hierarchy process (AHP) is used to construct a judgment matrix. The target layer is "multi-performance optimal", and the criteria layer is porosity, pore size, and air permeability. The weight vector is determined by the consistency test (CR<0.1).

[0033] 3.3. Based on Latin hypercube sampling, the process parameter space is sampled 1000 times to eliminate obviously unreasonable parameter combinations; the effective parameter combinations are input into the trained prediction model, the prediction performance index is output, and the NSGA-Ⅲ algorithm (population size = 100, number of iterations = 200) is used for optimization to obtain the Pareto front solution.

[0034] 3.4 Calculate the weighted score of each Pareto solution based on the weight vector determined by AHP (score = Σ(performance prediction value × corresponding weight)), select the solution with the highest score as the unique optimal combination of process parameters, and output the specific parameter values.

[0035] Step 4: Closed-loop feedback iteration; 4.1 Input the optimal process parameters into the LPBF equipment for printing experiments and test the actual performance indicators. If the relative error between the actual performance and the predicted value is ≤ ±5%, then this parameter combination is the final solution. If the error is > ±5%, supplement the original dataset with the set of "process parameters - actual multiphysics - actual performance" data, retrain the model, fine-tune the hyperparameters, and repeat step 3 until the error meets the requirements, forming a closed-loop optimization.

[0036] In summary, this invention addresses the issues of insufficient dataset validity and large discrepancies between simulation and experiment in small-sample scenarios by generating parameters through Latin hypercube design, correcting simulation parameters with a small number of experiments, and synthesizing small-sample data using the SMOTE algorithm to remove outliers. By optimizing the multi-input multi-output neural network architecture, rationally setting input dimensions and convolution kernel parameters, introducing a multi-head attention layer and residual connections, and combining ReLU and Sigmoid activation functions, it solves the problems of gradient vanishing and insufficient fitting accuracy for strong multi-input-multi-output coupling relationships inherent in traditional neural networks. Furthermore, by solving the Pareto optimal solution using the NSGA-III algorithm and weighted selection, combined with closed-loop verification (performance index relative error ≤ ±5%), it addresses the problems of inaccurate multi-objective optimization solutions and unreliable design results in traditional methods. Ultimately, it achieves efficient and accurate matching of permeable steel preparation parameters without relying on extensive trial-and-error experiments, significantly improving design efficiency and reliability.

[0037] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any person skilled in the art can make many possible variations and modifications to the technical solution of the present invention, or modify it into equivalent embodiments, without departing from the scope of the present invention's technical solution. Therefore, all equivalent changes made based on the shape, structure, and principle of the present invention without departing from the scope of the present invention's technical solution should be covered within the protection scope of the present invention.

Claims

1. A method for optimizing process parameters of LPBF permeable steel based on deep learning, characterized in that, Includes the following steps: Step S1: Acquisition and preprocessing of LPBF process data and simulation physics data; Step S2: Constructing a multi-objective prediction model based on deep learning; An improved multi-input multi-output neural network model consisting of convolution, attention, and residual connections is adopted. The input layer features include laser power, scanning speed, scanning spacing, layer thickness, peak temperature of the molten pool, cooling rate, peak thermal stress, and strain distribution. The feature extraction layer is set with 3×3, 5×5, and 3×3 convolution kernels, and the number of output channels is 64, 128, and 256 respectively. A multi-head attention layer is also adopted. Gradient vanishing is avoided by residual connections. The output features include porosity, pore size, and air permeability coefficient. The normalized predicted values ​​are output using the Sigmoid activation function. Initial hyperparameters were determined using a Bayesian optimization algorithm and adjusted through cross-validation. The AdamW optimizer was used with a weighted cross-entropy loss function. Overfitting was suppressed by L2 regularization. Training was stopped and the optimal model was saved when the relative errors of porosity, pore size, and air permeability prediction on the validation set were all ≤ ±5%. Step S3: Solving for process parameters based on multi-objective optimization algorithm; Based on the performance of LPBF equipment and material properties, process parameter boundaries are set; A custom target range is defined, and a judgment matrix is ​​constructed using the analytic hierarchy process. The target layer is "multi-performance optimality", and the criteria layer consists of porosity, pore size, and air permeability. The weight vector is determined through a consistency test. The process parameter space is sampled 1000 times based on Latin hypercube sampling to eliminate obviously unreasonable parameter combinations; the effective parameter combinations are input into the trained prediction model, the prediction performance index is output, and the Pareto front solution is obtained by optimizing it using the NSGA-Ⅲ algorithm. The weighted score of each Pareto solution is calculated based on the weight vector determined by AHP. The solution with the highest score is selected as the unique optimal combination of process parameters, and the specific parameter values ​​are output. Step S4: Closed-loop feedback iteration; Input the optimal process parameters into the LPBF equipment to conduct printing experiments and test the actual performance indicators. If the relative error between the actual performance and the predicted value is ≤ ±5%, the corresponding parameter combination is the final solution. If the error is > ±5%, the corresponding parameter combination data is added to the original dataset, the model is retrained, and step S3 is repeated until the error meets the requirements, thus forming a closed-loop optimization.

2. The method for optimizing process parameters of LPBF permeable steel based on deep learning as described in claim 1, characterized in that, Step S1 includes: Laser power, scanning speed, scanning spacing, and layer thickness, which are core parameters of the LPBF process, were selected as basic input parameters. A Latin hypercube design experimental scheme with uniform parameter space coverage and strong sample representativeness was adopted, and the parameter range was set. The thermal-stress coupling simulation software was used to generate multiphysics data in batches based on process parameters, including peak temperature of the molten pool, cooling rate, peak thermal stress, and strain distribution. The initial parameters of the simulation model were referenced from the thermal conductivity, specific heat capacity, and coefficient of thermal expansion of the material. Conduct calibration experiments to verify the validity of the simulation data; Data preprocessing includes removing outliers, standardizing input features, and normalizing output performance metrics. The SMOTE algorithm is used to expand small sample data, physical consistency is checked on the synthetic samples, redundant samples with similarity ≥95% to the original samples are removed, and random noise is added to the standardized input and output data to improve the model's generalization ability. The final labeled dataset is formed.

3. The method for optimizing process parameters of LPBF permeable steel based on deep learning as described in claim 2, characterized in that, The range of laser power, scanning speed, scanning spacing, and layer thickness parameters are set according to the LPBF equipment.

4. The method for optimizing process parameters of LPBF permeable steel based on deep learning as described in claim 2, characterized in that, Verification of the validity of simulation data includes: selecting 20 typical process parameter combinations for LPBF printing experiments, measuring porosity using the Archimedes drainage method, obtaining pore size through image analysis using a scanning optical microscope, and testing the gas permeability coefficient using a gas permeation test device.

5. The method for optimizing process parameters of LPBF permeable steel based on deep learning as described in claim 2, characterized in that, The 3σ criterion is used to remove abnormal data caused by simulation convergence failure and experimental equipment errors. The input features are standardized using Z-score, and the output performance index is normalized to the [0,1] interval using Min-Max normalization to eliminate the influence of dimensional differences on model training.

6. The method for optimizing process parameters of LPBF permeable steel based on deep learning as described in claim 2, characterized in that, The labeled dataset contains physical experimental data, calibrated simulation data, SMOTE synthetic data, and noise enhancement data. The data format is a one-to-one correspondence between input feature vectors and output performance vectors.

7. The method for optimizing process parameters of LPBF permeable steel based on deep learning as described in claim 1, characterized in that, The multi-head attention layer has 8 attention heads, and the weight decay coefficient of the AdamW optimizer is 0.

01.

8. The method for optimizing process parameters of LPBF permeable steel based on deep learning as described in claim 1, characterized in that, Users can customize the required porosity range, pore size range, and air permeability range.

9. The method for optimizing process parameters of LPBF permeable steel based on deep learning as described in claim 1, characterized in that, The NSGA-Ⅲ algorithm has a population size of 100 and an iteration count of 200.

10. The method for optimizing process parameters of LPBF permeable steel based on deep learning as described in claim 1, characterized in that, The weighted score is proportional to the product of the performance prediction value and its corresponding weight.

Images