An AI-enabled 3D printing topology optimization and lightweight structure generation method

Abstract

This invention discloses an AI-enabled method for 3D printing topology optimization and lightweight structure generation, belonging to the field of 3D printing technology. The method includes the following steps: 1. Constructing an AI model; 2. Converting design requirements and process parameters into standardized splicing vectors and inputting them into the AI ​​model to generate a voxelized topology for 3D printing; 3. Optimizing based on performance verification, process adaptation verification, and lightweight verification to generate the final 3D printed topology-optimized and lightweight structure. This invention achieves an end-to-end process for topology optimization and lightweight structure generation through an AI model, shortening the optimization process that traditional methods require several hours to minutes, improving optimization efficiency and significantly reducing reliance on human experience. By integrating multi-objective optimization design with 3D printing process parameters and performance constraints, the generated structure can meet both strength and stiffness requirements and adapt to the printing process.

Description

Technical Field

[0001] This invention belongs to the field of 3D printing technology, specifically relating to an AI-enabled 3D printing topology optimization and lightweight structure generation method. Background Technology

[0002] Topology optimization and lightweight design are the core prerequisites for the application of 3D printing technology in high-end manufacturing. The core goal is to minimize weight by optimizing the spatial distribution of materials while ensuring that the structure meets the mechanical performance requirements such as strength, stiffness, and fatigue life. This reduces manufacturing costs and improves product performance, such as fuel efficiency in aerospace, implant compatibility in medical devices, and energy consumption in the automotive industry.

[0003] Current topology optimization methods, including traditional methods and variable density methods based on iterative solutions using finite element analysis, still have many technical limitations and are difficult to adapt to the development needs of 3D printing technology, as follows: 1. Traditional methods rely heavily on human experience. For example, during the optimization process, designers need to set initial parameters such as material density threshold, iteration step size, and constraint weight based on engineering experience. Moreover, the optimization results often have problems that are theoretically feasible but difficult to manufacture. For example, the complex pore structure generated by optimization may lead to problems such as too many overhang areas, difficulty in removing support structures, and insufficient molding accuracy during 3D printing. Designers need to repeatedly adjust structural parameters such as pore size and distribution density, and even recalculate the optimization. The whole process requires 1-3 people per day of work, which is difficult to meet the needs of efficient manufacturing. 2. Iterative solutions based on finite element analysis, such as the variable density method, require multiple finite element calculations to update the material distribution in order to approximate the optimal solution. For example, when using the variable density method for topology optimization of a simple aero-engine bracket structure, 200-300 iterations are required. Each iteration includes steps such as mesh generation, load application, and stress analysis, and the overall optimization process takes 3-8 hours. If dynamic loads or fatigue life constraints are involved, the calculation time can be extended to more than 12 hours, which seriously restricts the needs of mass production and rapid iterative design.

[0004] In view of this, an AI-enabled 3D printing topology optimization and lightweight structure generation method is designed to solve the above problems. Summary of the Invention

[0005] To address the problems mentioned in the background section, this invention provides an AI-enabled 3D printing topology optimization and lightweight structure generation method. This method solves the problems of long computation times in existing topology optimization techniques, which severely restrict mass production and rapid iterative design, and its heavy reliance on human experience.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for topology optimization and lightweight structure generation in 3D printing based on AI, comprising the following steps: 1. Construct an AI model, including the Transformer module and the U-Net module. The Transformer module extracts the correlation features between design requirements and process parameters, and the U-Net module outputs the material existence probability of the voxelized structure based on the correlation features to generate the voxelized topology structure for 3D printing. 2. Convert design requirements and process parameters into standardized splicing vectors and input them into the AI ​​model to generate a voxelized topology for 3D printing; Third, the generated 3D printed voxelized topology is optimized based on performance verification, process adaptation verification, and lightweight verification to generate the final 3D printed topology optimized and lightweight structure.

[0007] Furthermore, in step one, the Transformer module includes six encoder layers. Each encoder layer consists of a multi-head self-attention mechanism and a feedforward neural network. The multi-head self-attention mechanism includes eight attention heads, and the feedforward neural network includes a linear layer, a ReLU activation function, and a linear layer. The multi-head self-attention mechanism captures the correlation features between design requirements and process parameters through the eight attention heads. The feedforward neural network performs nonlinear transformation on the correlation features output by the multi-head self-attention mechanism to enhance its expressive power.

[0008] Further, in step one, the U-Net module includes four downsampling layers, four upsampling layers, and an output layer. The downsampling layers include convolutional layers and max pooling layers, while the upsampling layers include transposed convolutional layers, skip connection layers, and convolutional layers. The downsampling layers gradually reduce the size of the 3D feature map and increase the number of channels to extract global features of the structure. The upsampling layers use transposed convolution to gradually restore the size of the 3D feature map, and simultaneously fuse the local detail features of the downsampling layers through skip connections to generate a voxelized structure. The output layer uses a sigmoid activation function to output the probability of material presence in the voxelized structure within the range [0,1]. Voxels with a probability > 0.5 are determined to be material regions, and voxels with a probability ≤ 0.5 are determined to be pore regions. Finally, a voxelized topology structure that can be directly used for 3D printing is generated.

[0009] Furthermore, in step one, the loss function of the AI ​​model is a multi-objective loss function that includes a performance error term, a lightweight objective term, and a process adaptation term, and its expression is: Loss=α×Loss_perf+β×Loss_light+γ×Loss_process; Where α, β, and γ are the weights of each loss term; Loss_perf represents the performance error term; Loss_light represents the lightweight target term; and Loss_process represents the process adaptation term.

[0010] Furthermore, in step two, the step of converting the design requirements and process parameters into a standardized spliced ​​vector includes: normalizing, voxelizing, and cleaning the design requirements and process parameters; mapping the processed design requirements and process parameters into design requirement vectors and process parameter vectors respectively; and splicing them together to form a standardized spliced ​​vector.

[0011] Furthermore, in step three, the performance verification steps include: converting the generated voxelized topology into a three-dimensional model in STL format for finite element analysis, applying loads and boundary constraints according to design requirements, performing mechanical performance analysis, and the indicators include maximum stress, maximum deformation, natural frequency, and fatigue life. If any indicator exceeds the threshold, the error value of the indicator is fed back to the AI ​​model to trigger secondary optimization. The AI ​​model automatically increases the weight of the corresponding indicator in the loss function and regenerates the voxelized topology until all performance indicators meet the standards.

[0012] Furthermore, in step three, the process adaptation verification step includes: importing the qualified voxelized topology into the 3D printing process simulation, performing a feasibility analysis based on the process parameter library, and the indicators include the overhang area ratio, minimum wall thickness and pore size. If any indicator fails, the overhang area, wall thickness and pore size will be automatically optimized.

[0013] Furthermore, in step three, the lightweight verification steps include: after optimization, calculating the lightweight rate of the voxelized topology; if the lightweight rate is less than the target value, adjusting the weight of the lightweight target term in the loss function, regenerating the voxelized topology, until the lightweight rate reaches the target.

[0014] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention realizes an end-to-end process of topology optimization and lightweight structure generation through AI model, shortening the optimization process that traditional methods require several hours to minutes, improving optimization efficiency and greatly reducing reliance on human experience. Through multi-objective optimization design that integrates 3D printing process parameters and performance constraints, the generated structure can meet the mechanical performance requirements such as strength and stiffness and adapt to the printing process.

[0015] 2. This invention, by employing voxelization representation and AI models, can generate complex porous structures that are difficult to achieve using traditional optimization methods, significantly improving the weight reduction rate and outperforming existing methods. It is particularly suitable for high-end manufacturing fields that are sensitive to weight, and has self-iterative update capabilities. By continuously optimizing the AI ​​model through actual printing data, it continuously improves the accuracy and adaptability of structure generation as the application scenarios expand, solving the problem of poor generalization ability of traditional methods. The unified digital process enables seamless connection from design requirements to 3D printing models, reducing information loss caused by multi-software interaction and providing a standardized topology optimization solution for intelligent manufacturing. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] I. Building the dataset (a) Data Collection Collect multi-source data, including design parameters, 3D printing process parameters, and corresponding topology optimization results; Design parameters include the structure's geometric dimensions, load conditions, boundary constraints, and material mechanical properties. Geometric dimensions include length, width, height, and feature dimensions; Load conditions include static loads, dynamic loads, and fatigue loads. Static loads include uniformly distributed loads and concentrated loads. Dynamic loads include vibration frequencies and impact loads. Fatigue loads include the number of cycles and stress amplitude. Boundary constraints include fixed constraints, hinged constraints, and sliding constraints; Material mechanical properties include elastic modulus, Poisson's ratio, yield strength, and fatigue limit; 3D printing process parameters include material type, printing technology type, key process parameters, and process constraints; The material types include metallic materials Ti6Al4V, aluminum alloys AlSi10Mg, polymer materials ABS, polymer materials PLA, and photosensitive resins; Printing technologies include SLM, FDM, and SLA; Key process parameters include layer thickness, printing speed, and energy input parameters. Energy input parameters include laser power / scanning gap, extrusion rate / nozzle temperature, and ultraviolet light intensity / exposure time. Process constraints include maximum overhang angle, minimum wall thickness, and minimum pore size; The topology optimization results include the corresponding structural model, performance verification data, lightweighting indicators, and printing feasibility feedback. The structural model includes STL format files; Performance verification data includes the maximum stress, deformation, natural frequency, and fatigue life from finite element analysis; Lightweighting metrics include structural weight, material volume ratio, and lightweighting rate. The feasibility feedback for printing includes the proportion of supporting structures, molding accuracy, and printing success rate.

[0019] (ii) Dataset preprocessing The multi-source data in the constructed dataset undergoes normalization, voxelization, and data cleaning. Normalization: Continuous parameter data such as geometric dimensions, load size, and process parameters are normalized to the [0,1] interval, and discrete parameters such as material type, printing technology, and load type are uniquely encoded or embedded and transformed into numerical vectors to eliminate the influence of dimensional differences on model training. Voxelization: The structural model is converted into a voxelized mesh with a resolution of 256×256×256 or 512×512×512. Each voxel represents the presence of material using binary or probability values. Linear interpolation is used during voxelization to preserve structural details and avoid feature loss. Data cleaning and processing: Remove abnormal data whose performance parameters exceed physical common sense or whose process parameters do not match the structure; fill missing data with generative models such as interpolation or variational autoencoders (VAEs) to ensure the integrity and accuracy of the dataset.

[0020] (III) Dataset Construction The preprocessed data is organized into structured samples according to the relationship between "design requirements - process parameters - topology results - performance feedback", and a dataset containing at least 100,000 samples is constructed. The dataset was divided into a training set for learning model parameters, a validation set for hyperparameter tuning and overfitting suppression, and a test set for evaluating model generalization ability in a 7:2:1 ratio.

[0021] II. Building AI Models (I) Constructing an AI model with a dual-module architecture of "feature extraction-structure generation" The AI ​​model includes the Transformer module and the U-Net module; The Transformer module includes six encoder layers. Each encoder layer consists of a multi-head self-attention mechanism and a feedforward neural network. The multi-head self-attention mechanism includes eight attention heads, and the feedforward neural network includes a linear layer, a ReLU activation function, and a linear layer. The U-Net module consists of four downsampling layers, four upsampling layers, and an output layer. The downsampling layers include convolutional layers and max pooling layers, while the upsampling layers include transposed convolutional layers, skip connection layers, and convolutional layers. The Transformer module's multi-head self-attention mechanism uses eight attention heads to capture the intrinsic correlation features between design requirements and process parameters. The feedforward neural network performs nonlinear transformations on the intrinsic correlation features output by the multi-head self-attention mechanism to enhance the model's expressive power. The intrinsic correlation features are expanded into a 3D feature map through a broadcast mechanism. The U-Net module's downsampling layer uses 3×3×3 convolutional kernels and 2×2×2 max pooling to gradually reduce the size of the 3D feature map and increase the number of channels for extracting global features of the structure. The upsampling layer uses transposed convolution to gradually restore the size of the 3D feature map. At the same time, it fuses the local detail features of the downsampling layer through skip connections to avoid gradient vanishing and detail loss, generating a voxelized structure. The output layer uses the Sigmoid activation function to output the probability of material presence in the voxelized structure within the range [0,1]. Voxels with a probability > 0.5 are identified as material regions, and voxels with a probability ≤ 0.5 are identified as pore regions. Finally, a voxelized topology structure that can be directly used for 3D printing is generated.

[0022] (ii) Constructing the loss function The loss function is a multi-objective loss function that includes performance error term, lightweight target term and process adaptation term, and its expression is: Loss=α×Loss_perf+β×Loss_light+γ×Loss_process; Where α, β, and γ are the weights of each loss term, with values ​​of 0.4, 0.3, and 0.3 respectively; Loss_perf represents the performance error term; Loss_light represents the lightweight target term; and Loss_process represents the process adaptation term. The values ​​of α, β, and γ of 0.4, 0.3, and 0.3 are based on extensive experimental verification. These weight ratios can achieve the optimal balance between performance, lightweight design, and manufacturing process. If adjustments are needed for specific scenarios, they can be dynamically optimized using the validation set. Loss_perf: Employs the L2 loss function to calculate the difference between the predicted performance of the structure generated by the AI ​​model in terms of stress, deformation, etc., and the actual finite element analysis results, ensuring that the structure meets the mechanical performance requirements. Loss_light: Employs the cross-entropy loss function, using a preset lightweighting rate target as a benchmark, to calculate the difference between the material volume ratio of the generated structure and the target value, thereby promoting structural lightweighting; Loss_process: A penalty loss function is used to penalize regions in the generated structure where the overhang angle is <45° or the wall thickness is <minimum allowable process value, which do not meet the process constraints, thereby reducing the risk of process adaptation.

[0023] III. Training the AI ​​Model Pre-training on publicly available large-scale topology optimization datasets such as NASA's TopOpt dataset and MIT's 3D Printing Design Dataset enables the constructed AI model to master basic topology optimization rules and structure generation logic. The design requirements and process parameter data features in the structured samples of the constructed dataset are mapped into design requirement vectors and process parameter vectors and then concatenated to form a concatenated vector; The concatenated vectors are input into the pre-trained AI model, and the AI ​​model parameters are optimized using gradient descent to fine-tune the AI ​​model. During fine-tuning, the Adam optimizer is used with an initial learning rate of 1e-4, which decays to 1 / 10 of the original rate every 20 epochs. The training is iterated for 100 epochs until the validation set loss converges, i.e., the loss value is <0.01. An early stopping strategy is adopted, and training is stopped when the validation set loss does not decrease for 10 consecutive epochs to avoid model overfitting. After training is completed, the model parameters are stored in the AI ​​model for subsequent topology optimization and structure generation.

[0024] IV. Topology Optimization and Lightweight Structure Generation 3D Printing Initial Topology Generation: Receive new design requirements and 3D printing process parameters from the user and convert them into standardized splicing vectors, which are then input into the AI ​​model. The AI ​​model uses the Transformer module to analyze the inherent correlation features between the design requirements and process parameters, and uses the U-Net module to generate a voxelized initial topology. Finite element verification: The generated voxelized initial topology is converted into a 3D model in STL format for finite element analysis. Loads and boundary constraints are applied according to design requirements, and mechanical performance analysis is performed. The indicators include maximum stress, maximum deformation, natural frequency, and fatigue life. If any indicator exceeds the threshold, the error value of the indicator is fed back to the AI ​​model to trigger secondary optimization. The AI ​​model automatically increases the weight of the corresponding indicator in the loss function and regenerates the voxelized initial topology until all performance indicators meet the standards. If the performance still fails to meet the standards after secondary optimization, an intermediate STL model is output, and the designer is prompted to adjust the design requirements. At the same time, the sample is recorded in the dataset for subsequent model iteration updates. If all performance indicators meet the preset requirements, process adaptation and lightweight verification are performed. Maximum stress refers to stress that is less than or equal to the material's yield strength; The maximum deformation amount refers to the amount that must be less than or equal to the preset threshold. Natural frequency refers to the vibration frequency that should be avoided in the working environment; Fatigue life refers to the number of cycles required to be greater than or equal to the preset number. Process adaptation verification: The qualified initial voxelization topology is imported into 3D printing process simulation modules such as Autodesk Netfabb and Materialise Magics. Feasibility analysis is performed based on the process parameter library. The indicators include the proportion of overhang area, minimum wall thickness and pore size. If any indicator fails, local structure optimization is automatically performed. Overhang area ratio: The overhang angle is calculated based on the angle between the printing direction vector and the normal vector of the structural surface. Areas with an angle < 45° are identified as overhang areas that need support. The proportion of the overhang area to the total surface area of ​​the structure is calculated. Minimum wall thickness: By analyzing the 3D model slices, the wall thickness of each region of the structure is detected to determine whether it meets the minimum wall thickness requirements of the current printing technology. Pore ​​size: Detect the diameter and spacing of pores inside the structure to determine whether they meet the requirements for material flowability and post-processing; Overhang area optimization: Reduce the overhang angle by adding transition support structures or adjusting the pore direction; Wall thickness optimization: Local additive manufacturing is performed on areas with a wall thickness less than the minimum wall thickness to ensure that the wall thickness meets the standard, while preserving the lightweight effect as much as possible; Pore ​​optimization: Merge or enlarge pores smaller than the minimum pore size to avoid material blockage during printing; Lightweight Validation: After optimization, the lightweight ratio of the voxelized topology is calculated. If the lightweight ratio is less than the target value, the weight of the lightweight target term in the loss function is adjusted, and the initial voxelized topology is regenerated until the lightweight ratio reaches the target. If the lightweight ratio reaches the target, the final STL format 3D printing model is output, generating AI-enabled 3D printing topology optimization and lightweight structure. The lightweight ratio is expressed as: Lightweighting rate = (Original solid structure weight - Generated structure weight) / Original solid structure weight × 100% The weight of the original solid structure is calculated based on the same geometric outline and material as the generated structure. The target value can be dynamically adjusted according to design requirements. For example, it can be set to 50%-60% in the aerospace field and 30%-40% in the medical device field.

[0025] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for topology optimization and lightweight structure generation in 3D printing based on AI, characterized in that, Includes the following steps:

1. Construct an AI model, including the Transformer module and the U-Net module. The Transformer module extracts the correlation features between design requirements and process parameters, and the U-Net module outputs the material existence probability of the voxelized structure based on the correlation features to generate the voxelized topology structure for 3D printing.

2. Convert design requirements and process parameters into standardized splicing vectors and input them into the AI ​​model to generate a voxelized topology for 3D printing; Third, the generated 3D printed voxelized topology is optimized based on performance verification, process adaptation verification, and lightweight verification to generate the final 3D printed topology optimized and lightweight structure.

2. The AI-enabled 3D printing topology optimization and lightweight structure generation method according to claim 1, characterized in that: In step one, the Transformer module includes six encoder layers. Each encoder layer consists of a multi-head self-attention mechanism and a feedforward neural network. The multi-head self-attention mechanism includes eight attention heads, and the feedforward neural network includes a linear layer, a ReLU activation function, and a linear layer. The multi-head self-attention mechanism captures the correlation features between design requirements and process parameters through the eight attention heads. The feedforward neural network performs nonlinear transformation on the correlation features output by the multi-head self-attention mechanism to enhance its expressive power.

3. The AI-enabled 3D printing topology optimization and lightweight structure generation method according to claim 2, characterized in that: In step one, the U-Net module includes four downsampling layers, four upsampling layers, and an output layer. The downsampling layers include convolutional layers and max pooling layers, while the upsampling layers include transposed convolutional layers, skip connection layers, and convolutional layers. The downsampling layers gradually reduce the size of the 3D feature map and increase the number of channels to extract global features of the structure. The upsampling layers use transposed convolution to gradually restore the size of the 3D feature map and simultaneously fuse the local detail features of the downsampling layers through skip connections to generate a voxelized structure. The output layer uses the Sigmoid activation function to output the probability of material presence in the voxelized structure within the range [0,1]. Voxels with a probability > 0.5 are identified as material regions, and voxels with a probability ≤ 0.5 are identified as pore regions. Finally, a voxelized topology structure that can be directly used for 3D printing is generated.

4. The AI-enabled 3D printing topology optimization and lightweight structure generation method according to claim 3, characterized in that: In step one, the loss function of the AI ​​model is a multi-objective loss function that includes a performance error term, a lightweight objective term, and a process adaptation term, expressed as: Loss=α×Loss_perf+β×Loss_light+γ×Loss_process Where α, β, and γ are the weights of each loss term; Loss_perf represents the performance error term; Loss_light represents the lightweight target term; and Loss_process represents the process adaptation term.

5. The AI-enabled 3D printing topology optimization and lightweight structure generation method according to claim 4, characterized in that: In step two, the steps of converting the design requirements and process parameters into standardized spliced ​​vectors include: normalizing, voxelizing, and cleaning the design requirements and process parameters; mapping the processed design requirements and process parameters into design requirement vectors and process parameter vectors respectively; and splicing them together to form standardized spliced ​​vectors.

6. The AI-enabled 3D printing topology optimization and lightweight structure generation method according to claim 5, characterized in that: In step three, the performance verification steps include: converting the generated voxelized topology into a three-dimensional model in STL format for finite element analysis, applying loads and boundary constraints according to design requirements, and performing mechanical performance analysis. The indicators include maximum stress, maximum deformation, natural frequency, and fatigue life. If any indicator exceeds the threshold, the error value of that indicator is fed back to the AI ​​model to trigger secondary optimization. The AI ​​model automatically increases the weight of the corresponding indicator in the loss function and regenerates the voxelized topology until all performance indicators meet the standards.

7. The AI-enabled 3D printing topology optimization and lightweight structure generation method according to claim 6, characterized in that: In step three, the process adaptation verification steps include: importing the qualified voxelized topology into the 3D printing process simulation, performing a feasibility analysis based on the process parameter library, and the indicators include the overhang area ratio, minimum wall thickness and pore size. If any indicator fails, the overhang area, wall thickness and pore size will be automatically optimized.

8. The AI-enabled 3D printing topology optimization and lightweight structure generation method according to claim 7, characterized in that: In step three, the lightweight verification steps include: after optimization, calculating the lightweight rate of the voxelized topology; if the lightweight rate is less than the target value, adjusting the weight of the lightweight target term in the loss function and regenerating the voxelized topology until the lightweight rate reaches the target.

Images