Physical perception driven artificial intelligence regulation and control method for additive manufacturing light field and DLP printing light field regulation and control method
By using a physical perception-driven artificial intelligence method to control the light field, an adaptive light pattern with uniform energy distribution is generated, which solves the warping deformation problem caused by thermal stress in DLP printing and improves printing accuracy and edge resolution.
Patent Information
- Application Number
- CN202511042587.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-28
- Publication Date
- 2025-11-07
AI Technical Summary
In DLP 3D printing technology, the rapid photopolymerization reaction of photosensitive resin slurry under exposure leads to significant exothermic effects and thermal stress problems, resulting in structural warping and deformation and attenuation of edge resolution.
By deeply analyzing the geometric features of the target structure and the interaction between light and slurry, and using a physical perception-driven artificial intelligence control method for additive manufacturing light field, an adaptive light pattern with uniform energy distribution is generated to suppress warping deformation caused by thermal effects and optimize energy distribution.
It significantly improves the vertical structural accuracy and edge resolution of DLP printing, reduces warping deformation caused by thermal stress, and achieves practical accuracy at the micron level.
Smart Images
Figure CN120912422A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of additive manufacturing, and more particularly relates to a physical perception driven additive manufacturing light field artificial intelligence regulation method and a DLP printing light field regulation method. BACKGROUND
[0002] Digital Light Processing (DLP) three-dimensional printing technology generates a dynamic light field pattern with the help of a spatial light modulator, and through surface projection exposure, a local photopolymerization reaction of a photosensitive resin slurry is induced, and a three-dimensional structure is formed by layer-by-layer stacking. Its theoretical forming precision can reach the micron level, but the actual printing quality is restricted by the complex coupling of light physical effects and thermal chemical responses.
[0003] The core challenge of DLP technology is that when the photosensitive resin slurry undergoes rapid photopolymerization reaction under surface projection exposure, it is accompanied by a significant heat release effect - a large amount of light energy is absorbed and converted into heat energy in a short time per unit volume of slurry. Due to the inherent differences in the thermal physical properties of the resin matrix, fillers and cured network (such as thermal expansion coefficient, thermal conductivity), and the spatial non-uniformity of the Gaussian distribution of light field energy, a sharp temperature gradient field will be formed instantaneously in the cured layer. This non-uniform thermal expansion will induce a complex local thermal stress field.
[0004] In addition, due to the spatial difference in thermal shrinkage rate and the non-uniformity of curing shrinkage, the generated thermal stress cannot be completely or uniformly released. When the accumulated residual stress exceeds the structural stiffness bearing limit, it will drive the cured layer and the overall structure to undergo unexpected macroscopic warping deformation, which is specifically manifested as edge curling of thin plates, inclination of high aspect ratio columns and significant deformation of cantilever structures.
[0005] The traditional DLP printing technology adopts a binary pattern exposure mode (target area full bright, non-target area full dark), which is simple to operate, but ignores the scattering characteristics of Gaussian energy, introduces a low-pass filtering effect in the optical transmission chain, and causes serious resolution attenuation of the edge of the formed structure.
[0006] In view of the above bottleneck, a breakthrough intelligent light field regulation strategy is urgently needed. SUMMARY
[0007] The main purpose of the present application is to provide a physical mechanism driven intelligent light field regulation strategy: by deeply analyzing the geometric characteristics of the target structure and the physical mechanism of the light-paste interaction, the optimized surface projection light pattern coupled with the paste curing threshold information and the diffraction effect law is generated adaptively; by synchronously regulating the overall shape profile and the internal pixel gray value of the projection pattern, the uniform distribution of light field energy is realized, so as to suppress the warping deformation caused by thermal effect, eliminate the unintended structure bending, optimize the energy distribution uniformity, avoid local heat concentration, fundamentally improve the vertical structure precision of DLP printing, and break through the inherent printing structure limitation of traditional methods.
[0008] In order to achieve the above purpose, the present application adopts the following technical scheme:
[0009] A physical perception driven additive manufacturing light field artificial intelligence regulation method, comprising the following steps:
[0010] Collecting standardized surface projection light pattern-light field distribution map pair data, and constructing a surface projection light field feature multi-gradient distribution data set;
[0011] Fusing a residual separable convolutional network in a UNet network to construct a surface projection light multi-scale feature extraction module, using the surface projection light feature extraction module to learn the surface projection light pattern and the light field distribution map in the surface projection light field feature multi-gradient distribution data set, preliminarily encoding the surface projection light information to obtain an encoded feature map, and learning the mapping relationship between the surface projection light pattern and the light field distribution map;
[0012] Constructing a recurrent optimization module based on a convolution gate recurrent unit (ConvGRU) sequence optimization structure, and inputting the encoded feature map into the recurrent optimization module for recurrent optimization;
[0013] Constructing a light energy diffraction physical model, and perceiving a constraint loss function to explicitly constrain the implicit features of the light field distribution, accelerate the network convergence efficiency, and make the surface projection light pattern more consistent with the energy diffraction and absorption law in the actual light curing process.
[0014] Further, the recurrent optimization module learns the dynamic evolution process of the light field distribution of the surface projection light pattern in the surface projection light field feature multi-gradient distribution data set in the time sequence iteration process; wherein, the recurrent optimization module takes the light field distribution map obtained by the surface projection light pattern in the current round as the implicit feature, combines the energy threshold constraint in the light curing process, dynamically predicts the correction amount of the light pattern, adjusts the intensity distribution of the pattern in the recurrent optimization process, so that the light field energy distribution of the surface projection light is more uniform and stable.
[0015] Further, the calculation process of ConvGRU is as shown in the formula:
[0016] z t =σ(Wz x Concat(F t , h t-1 ), h0 = E T ;
[0017] r t = s(W r x Concat(F t , h t-1 ) ) ;
[0018]
[0019]
[0020] where t represents the serial number of the current node, E T represents the energy threshold, z t represents the update gate, r t represents the reset gate, F t is the feature image of the current node, represents the update momentum of the hidden state, h t represents the hidden state; W z , W r , W h represent the weights of the convolution kernel; represents matrix multiplication, and Concat represents the channel concatenation operation.
[0021] Further, the light energy diffraction physical model perceives a constraint loss function; a loss function is constructed by combining Gaussian distribution and Beer-Lambert law, and the calculation process is as follows:
[0022] L Total = l1L CE + l2L LF ;
[0023]
[0024] where L CE is the cross-entropy loss function of the recurrent neural network, L LF is the face projection light field distribution loss function, E T is the slurry solidification threshold, m is a constant representing the light propagation attenuation coefficient, H and W represent the height and width of the face projection light image, respectively.
[0025] A DLP printing light field regulation method is applied to the above-mentioned physical perception driven additive manufacturing light field artificial intelligence regulation method, and the regulation method comprises the following steps:
[0026] S100, target solidification structure design:
[0027] According to the printing requirement design target three-dimensional structure, a series of single layer solidification area binary images are extracted by a slicing algorithm as initial input data for network optimization;
[0028] S200, network training and optimization
[0029] The single layer solidification area binary image obtained in step S100 is encoded by a face projection light feature extraction module to extract light field information, iteratively adjusted by a ConvGRU recurrent optimization module, combined with a physical prior constraint loss function to calculate errors and back propagation, and the network parameters are optimized;
[0030] S300, output the optimized face projection light pattern, and the face projection light pattern comprises a gray gradient distribution.
[0031] Compared with the prior art, the present application has the following beneficial effects:
[0032] Through the artificial intelligence light field regulation method driven by physical perception, combined with the DLP three-dimensional printing technology, the following significant effects can be achieved:
[0033] (1) Dynamically optimize the gray gradient distribution of the face projection light pattern, effectively suppress the local energy concentration, and significantly reduce or even eliminate the warping deformation phenomenon caused by the thermal effect;
[0034] (2) Self-adaptive regulation of light pattern shape and pixel gray value, realizing uniform distribution of light field energy, and improving the structure consistency of the solidification layer;
[0035] (3) Breakthrough the precision limit of traditional binary pattern exposure, significantly improve the size precision and edge resolution of DLP printing, and realize the micron-level theoretical precision in practical application;
[0036] (4) Reduce the complexity of artificial design of light field distribution, realize adaptive optimization of light pattern through intelligent algorithm, and improve the vertical structure precision of DLP printing. BRIEF DESCRIPTION OF DRAWINGS
[0037] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiment or prior art description will be briefly introduced below. Obviously, the drawings in the following description are only embodiments of the present application, and those skilled in the art can obtain other drawings according to the provided drawings without creative labor.
[0038] Figure 1 Optimization of light field model deep learning network flowchart;
[0039] Figure 2 Face projection light multi-scale feature extraction module flowchart;
[0040] Figure 3 A schematic diagram of a convolution gate recurrent unit (ConvGRU). DETAILED DESCRIPTION
[0041] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.
[0042] Embodiment 1
[0043] Reference Figures 1-3 A physical perception driven additive manufacturing light field artificial intelligence regulation method comprises the following steps:
[0044] Collecting standardized face projection light pattern-light field distribution map pair data to construct a face projection light field feature multi-gradient distribution data set;
[0045] Fusing a residual separable convolutional network in a UNet network to construct a face projection light multi-scale feature extraction module, using the face projection light feature extraction module to learn the face projection light pattern and the light field distribution map in the face projection light field feature multi-gradient distribution data set, preliminarily encoding the face projection light information to obtain an encoded feature map, and learning the mapping relationship between the face projection light pattern and the light field distribution map;
[0046] Constructing a recurrent optimization module based on a sequence optimization structure of a convolution gate recurrent unit (ConvGRU), and inputting the encoded feature map into the recurrent optimization module for recurrent optimization;
[0047] Constructing a light energy diffraction physical model, and perceiving a constraint loss function to explicitly constrain the implicit features of the light field distribution, accelerate the network convergence efficiency, and make the face projection light pattern more consistent with the energy diffraction diffusion and absorption rules in the actual light curing process.
[0048] The physical perception driven additive manufacturing light field artificial intelligence regulation method provided in the embodiment learns the light field energy distribution of the face projection light image, dynamically regulates the intensity distribution of the face projection light pattern, and thus realizes the task of generating an energy uniform face projection light pattern. The overall flowchart of the network is shown in Figure 1 .
[0049] In computer vision tasks, a recurrent neural network (RNN) framework is used for sequence feature information extraction to realize optimization of a computer vision task target in combination with sequence change feature information.
[0050] In the embodiment, the face projection light multi-scale feature extraction module fuses a residual separable convolutional network in the UNet network, and the module structures of the shallow layer and the deep layer are different. The UNet network can effectively retain the texture information and edge feature information of low-dimensional information and extract the overall regional features of high-dimensional information. The residual separable convolutional network combines a plurality of jump connections and convolutional modules, improves the feature expression capability by decoupling the multi-scale spatial correlation and channel correlation, and the network structure can maximize the feature extraction capability of the separable convolution as the feature image changes. The combination of the UNet and the residual separable convolutional network can fuse the shallow features and the deep features, effectively solve the problem of different semantic consistency at different levels, and improve the spatial, intensity and edge information extraction capability of the face projection light. The specific structure is as shown in Figure 2 .
[0051] In the embodiment, the cycle optimization module learns the dynamic evolution process of the light field distribution of the face projection light pattern in the time sequence iteration process in the multi-gradient distribution data set of the face projection light field features. The cycle optimization module takes the light field distribution map obtained by the face projection light pattern in the current round as the implicit feature, combines the energy threshold constraint in the light curing process, dynamically predicts the correction amount of the light pattern, adjusts the intensity distribution of the pattern in the cycle optimization process, so that the light field energy distribution of the face projection light is more uniform and stable.
[0052] To generate a face projection light pattern that makes the target light field energy more uniform, the boundary more clear, and the intensity more controllable, the cycle optimization module realizes the gradual evolution of the face projection pattern in the iteration process, and realizes the dynamic adjustment of the pixel intensity distribution. The module learns the mapping relationship between the face projection light pattern and the light field distribution map, combines the current pattern state and the historical update information, and optimizes the output pattern in multiple iterations. In the cycle optimization process, a sequence optimization structure based on a convolution gate recurrent unit (ConvGRU) is used to learn the dynamic evolution process of the light field distribution of the face projection light pattern in the time sequence iteration process, as shown in Figure 3 .
[0053] The cycle optimization module takes the light field distribution map obtained by the face projection light pattern in the current round as the implicit feature, combines the energy threshold constraint in the light curing process, dynamically predicts the correction amount of the light pattern, adjusts the intensity distribution of the pattern in the cycle optimization process, so that the light field energy distribution of the face projection light is more uniform and stable.
[0054] In the embodiment, the calculation process of the ConvGRU is as shown in the formula:
[0055] z t =σ(W z ×Concat(F t ,h t-1 )),h0=E T ;
[0056] r t = σ(W r × Concat(F t , h t-1 ));
[0057]
[0058]
[0059] where t represents the serial number of the current node, E T represents the energy threshold, z t represents the update gate, r t represents the reset gate, F t is the feature image of the current node, h represents the update momentum of the hidden state, h t represents the hidden state; W z , W r , and W h represent the weights of the convolution kernel; · represents matrix multiplication, and Concat represents the channel concatenation operation.
[0060] The diffraction effect law of light energy in the slurry is introduced as a physical perception constraint prior, and in the embodiment, the physical prior constraint calculation distribution loss is a light energy diffraction physical model perception constraint loss function, which is used to explicitly constrain the implicit features of the light field distribution, so as to improve the model convergence efficiency and make the face projection light pattern more consistent with the energy diffraction diffusion and absorption law in the actual light curing process.
[0061] where the loss function is constructed by combining the Gaussian distribution and the Beer-Lambert law, and the calculation process is as follows:
[0062] L Total = λ1L CE + λ2L LF ;
[0063]
[0064] where L CE is the cross-entropy loss function of the recurrent neural network, L LF is the face projection light field distribution loss function, E T is the slurry curing threshold, μ is a constant representing the light propagation attenuation coefficient, and H and W represent the height and width of the face projection light image, respectively.
[0065] The physical perception driven additive manufacturing light field artificial intelligence regulation method provided by the embodiment breaks through the limitation of traditional pure data driven method lacking of physical constraints.
[0066] The dynamic evolution process of the face projection light field optimization is learned by using the recurrent neural network based on ConvGRU, the current state and historical information are combined to realize iterative optimization, the convergence efficiency and interpretability of the model are improved, and the problem that the traditional static optimization method cannot adapt to dynamic light field changes is solved.
[0067] Embodiment 2
[0068] A DLP printing light field regulation method is provided for the application of the physical perception driven additive manufacturing light field artificial intelligence regulation method provided in embodiment 1, and the regulation method comprises the following steps:
[0069] S100, target solidification structure design:
[0070] The target three-dimensional structure is designed according to the printing requirement, a series of single-layer solidification area binary images (the target area is 1 and the non-target area is 0) are extracted by a slicing algorithm, and the single-layer solidification area binary images are used as initial input data for network optimization;
[0071] S200, network training and optimization
[0072] The single-layer solidification area binary image obtained in step S100 is encoded by a face projection light feature extraction module to obtain light field information, the light field information is iteratively adjusted by a ConvGRU recurrent optimization module, an error is calculated by combining a physical prior constraint loss function and is back propagated, and network parameters are optimized;
[0073] S300, output the optimized face projection light pattern, and the face projection light pattern comprises a gray gradient distribution.
[0074] The technical solutions of the present application are fully described above, it should be noted that the specific embodiments of the present application are not limited by the above description, all technical solutions formed by the person skilled in the art according to the spirit and essence of the present application in structure, method or function, etc. Using equivalent transformation or equivalent transformation, all technical solutions fall within the protection scope of the present application.
Claims
1. A physical perception driven additive manufacturing light field artificial intelligence regulation method, characterized in that, The method comprises the following steps: Collecting standardized face projection light pattern-light field distribution pair data to construct a face projection light field feature multi-gradient distribution data set; Fusing a residual separable convolutional network in a UNet network to construct a face projection light multi-scale feature extraction module, using the face projection light feature extraction module to learn the face projection light pattern and the light field distribution in the face projection light field feature multi-gradient distribution data set, preliminarily encoding the face projection light information to obtain an encoded feature map, and learning the mapping relationship between the face projection light pattern and the light field distribution; Constructing a recurrent optimization module based on a sequence optimization structure of a convolution gate recurrent unit (ConvGRU), and inputting the encoded feature map into the recurrent optimization module for recurrent optimization; Constructing a light energy diffraction physical model and a constraint loss function to explicitly constrain the implicit features of the light field distribution, accelerate the network convergence efficiency, and make the face projection light pattern more consistent with the energy diffraction diffusion and absorption rules in the actual light curing process.
2. The physically perceptually driven additive manufacturing light field artificial intelligence regulation method according to claim 1, wherein, The recurrent optimization module learns the dynamic evolution process of the light field distribution of the face projection light pattern in the time sequence iteration process in the face projection light field feature multi-gradient distribution data set; wherein the recurrent optimization module takes the light field distribution map obtained from the face projection light pattern of the current round as the implicit feature, dynamically predicts the correction amount of the light pattern by combining the energy threshold constraint in the light curing process, adjusts the intensity distribution of the pattern in the recurrent optimization process, so that the light field energy distribution of the face projection light is more uniform and stable.
3. The physically aware, drive additive manufacturing light field artificial intelligence regulatory method of claim 1, wherein, The calculation process of the ConvGRU is shown in the formula: z t = σ(W z × Concat(F t ,h t-1 )), h0= E T ; r t = σ(W r × Concat(F t , h t-1 )); where t represents the serial number of the current node, E T represents the energy threshold, z t represents the update gate, r t represents the reset gate, F t is the feature image of the current node, represents the update momentum of the hidden state, h t represents the hidden state; W z , W r , W h represents the weight of the convolution kernel; · represents matrix multiplication, and Concat represents the channel concatenation operation.
4. The physically aware, drive additive manufacturing light field artificial intelligence regulatory method of claim 1, wherein, The light energy diffraction physical model perceives the constraint loss function; a loss function is constructed by combining Gaussian distribution and Beer-Lambert law, and the calculation process is as follows: L Total = λ1L CE + λ2L LF ; where L CE is the cross-entropy loss function of the recurrent neural network, L LF is the face projection light field distribution loss function, E T is the slurry solidification threshold, μ is a constant representing the light propagation attenuation coefficient, and H and W represent the height and width of the face projection light image, respectively.
5. A method of DLP printing light field steering, characterized in that, The application of the physical perception driven additive manufacturing light field artificial intelligence regulation method according to any one of claims 1 to 4, wherein the regulation method comprises the following steps: S100, target curing structure design: Design a target three-dimensional structure according to the printing requirements, extract a series of single-layer curing area binary images through a slicing algorithm, and use the binary images as the initial input data for network optimization; S200, network training and optimization The single-layer curing area binary image obtained in step S100 is encoded into light field information through the face projection light feature extraction module, iteratively adjusted through the ConvGRU recurrent optimization module, the error is calculated combined with the physical prior constraint loss function and back propagated, and the network parameters are optimized; S300, output the optimized face projection light pattern, and the face projection light pattern comprises a gray gradient distribution.