A Deep Learning-Based Method for Predicting Surface Morphology and Wettability in Laser Processing

Abstract

This application discloses a deep learning-based method for predicting the surface morphology and wettability of laser-processed surfaces. The method includes: acquiring SEM morphology data of components processed with different laser processing parameters; acquiring wettability data after laser processing; constructing a mapping dataset of laser processing parameters, SEM morphology data, and wettability data for the metal components; constructing and training a surface morphology generation model based on a conditional generative adversarial network; constructing and training a surface wettability prediction model based on deep residual network branches and conditional parameter encoding branches; inputting the laser processing parameters of the sample to be tested into the surface morphology generation model to output the generated SEM image; and inputting the generated SEM image and laser processing parameters into the surface wettability prediction model to output the predicted wettability value. This method organically integrates laser processing parameter data and surface morphology data, not only improving model performance and generalization ability but also outputting highly accurate prediction results.

Description

Technical Field

[0001] This application belongs to the technical field of materials processing and machine learning, specifically involving a method for predicting the surface morphology and wettability of laser-processed surfaces based on deep learning. Background Technology

[0002] Laser micro / nanostructure processing technology is widely used in the surface functionalization of various metallic / non-metallic materials due to its advantages such as non-contact operation, high energy density, high precision, and high processing flexibility. By controlling process parameters such as laser scanning speed, scanning spacing, single pulse energy, and repetition frequency, different periodic micro / nano structures can be formed on the material surface. Combined with subsequent chemical treatments, the wettability of the material surface can be controlled, which has important engineering application value in fields such as corrosion protection, self-cleaning, anti-icing, biomedicine, and microfluidics. However, the laser processing involves highly complex thermophysical processes, including transient melting and solidification of materials, vaporization and recasting of materials, and plasma effects, among other physical mechanisms. The processing results are affected by the nonlinear coupling of various process parameters. Even under the condition of consistent scanning paths, different combinations of parameters can still lead to differences in the surface micro / nano structures, thereby causing changes in surface wettability. This complexity makes it difficult to describe the relationship between the structure and wettability of laser-processed surfaces using traditional physical models or empirical formulas.

[0003] Currently, research on the wetting properties of laser-processed surfaces mainly relies on experimental methods. This involves obtaining surface morphology and contact angle data through numerous parameter combination experiments, followed by optimization of process parameters based on empirical rules. However, this method suffers from problems such as long experimental cycles, high sample acquisition costs, and limited parameter space coverage, making it difficult to meet the needs of rapid design and mass production applications. Furthermore, the lack of unified standards for process parameters across different materials and laser systems results in poor transferability and reusability of experimental results.

[0004] In recent years, with the rapid development of artificial intelligence, some studies have attempted to introduce machine learning methods to predict laser processing results in order to reduce experimental costs and improve design efficiency. For example, regression models or shallow neural networks are used to establish a mapping relationship between laser processing parameters and surface properties to predict surface contact angles or roughness. However, such methods typically only use processing parameters as input, ignoring the key intermediate variable of surface micro / nanostructure morphology, making it difficult to fully characterize the influence of complex structural features on wetting performance, thus limiting prediction accuracy and generalization ability. On the other hand, with the development of deep learning technology, models such as generative adversarial networks have made significant progress in image generation and reconstruction, and some studies have attempted to apply them to material microstructure modeling. However, existing generative models often do not fully incorporate laser processing parameters as constraints, resulting in an unclear correspondence between the generated results and the actual process, making it difficult to directly guide the actual processing.

[0005] Based on the above problems, it is necessary to propose a new technical solution that organically combines laser processing parameters, surface morphology generation, and wetting performance prediction by introducing a deep learning model. This will enable efficient prediction and collaborative design of laser-processed surface structure and wetting performance, thereby reducing experimental costs, shortening the R&D cycle, and improving the intelligence level of laser functional surface design. Summary of the Invention

[0006] This application provides a deep learning-based method for predicting the surface morphology and wettability of laser-processed surfaces, enabling the generation of laser-processed surface images and quantitative prediction of surface wettability. This addresses the problems of high experimental trial-and-error costs and long design cycles in the process of designing a surface with specific wettability.

[0007] The technical solution provided in this application is as follows.

[0008] In a first aspect, this application provides a method for predicting the surface morphology and wettability of laser-processed surfaces based on deep learning, including: Obtain SEM morphology data of components after processing with different laser processing parameters; Acquire wettability data of laser-processed metal components after chemical-thermal treatment; the chemical-thermal treatment transforms the surface wettability through organic solution immersion treatment; Preprocessing of SEM morphology data and laser processing parameter data is performed to construct a mapping dataset of laser processing parameters, SEM morphology data and wettability data of metal components; A surface morphology generation model is constructed and trained based on a conditional generative adversarial network. Laser processing parameters are used as conditional inputs. The generator produces the corresponding surface morphology image, and the discriminator performs discrimination training. A surface wettability prediction model was constructed and trained based on deep residual network branches and conditional parameter encoding branches, with SEM morphology data and corresponding laser processing parameters as dual-branch inputs and wettability data as output. The laser processing parameters of the sample to be tested are input into the trained surface morphology generation model, which outputs the generated SEM image. The generated SEM image and the corresponding laser processing parameters are input into the trained surface wettability prediction model, which outputs the predicted wettability value.

[0009] In one possible implementation, the material of the component includes one of aluminum alloy, titanium alloy, copper alloy, stainless steel, zirconium oxide, and silicon carbide; the laser processing parameters include laser scanning speed, scanning spacing, and single pulse energy.

[0010] In one possible implementation, the surface wettability prediction model includes a generator and a discriminator; The generator sequentially comprises: an input layer, a conditional fusion module, a first deconvolution module, a first deconvolution module with residuals, a self-attention module, a second deconvolution module, a second deconvolution module with residuals, a third deconvolution module, a third convolution module with residuals, and a fourth deconvolution module; the deconvolution module includes a deconvolution layer and a ReLU function; The discriminator sequentially comprises: an input layer, a first convolutional module with residuals, a first convolutional module, a second convolutional module with residuals, a second convolutional module, a third convolutional module with residuals, an auto-attention block, a third convolutional module, a fourth convolutional module with residuals, a flattening layer, a linear transformation layer, and an output layer; the convolutional module includes a convolutional layer and a LeakyReLU function.

[0011] In one possible implementation, the model parameters of the surface topography generation model are updated using the Adam optimizer, with the loss function being a combination of Wasserstein loss and L1 loss.

[0012] In one possible implementation, the surface wettability prediction model includes a ResNet152 network branch, a conditional parameter encoding branch, a feature fusion layer, and an output layer. The ResNet152 network branch includes: stem, four residual stages, and an identity mapping layer; the condition parameter encoding branch includes multiple fully connected layers for mapping low-dimensional condition parameters to high-dimensional condition parameters; the feature fusion layer is used to concatenate the features output from the two branches to achieve joint representation of surface image information and condition parameters; the output layer includes a fully connected layer and an inverse normalization layer.

[0013] Furthermore, in the ResNet152 network branch, the stem includes 7×7 convolutions and pooling; The four residual stages include: conv2_x with 3 bottleneck blocks, conv3_x with 8 bottleneck blocks, conv4_x with 36 bottleneck blocks, and conv5_x with 3 bottleneck blocks; each bottleneck block includes 3 convolutions and residual connections.

[0014] Secondly, this application provides a deep learning-based device for predicting the surface morphology and wettability of laser-processed surfaces, including: The acquisition module is used to acquire SEM morphology data of components after processing with different laser processing parameters; acquire wettability data of laser-processed metal components after chemical-thermal treatment; the chemical-thermal treatment transforms the surface wettability through organic solution immersion treatment; The dataset construction module is used to preprocess SEM morphology data and wettability data to construct a mapping dataset of laser processing parameters, SEM morphology data and wettability data of metal components; The first model building module is used to build and train a surface morphology generation model based on a conditional generative adversarial network. It takes laser processing parameters as conditional input, generates the corresponding surface morphology image through the generator, and performs discrimination training by the discriminator. The second model building module is used to build and train a surface wettability prediction model based on deep residual network branches and conditional parameter encoding branches. It uses SEM morphology data and corresponding laser processing parameters as dual-branch inputs and wettability data as output. The output module is used to input the laser processing parameters of the sample to be tested into the trained surface morphology generation model and output the generated SEM image; and to input the generated SEM image and the corresponding laser processing parameters into the trained surface wettability prediction model and output the predicted wettability value.

[0015] Thirdly, this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the deep learning-based laser processing surface morphology and wettability prediction method as described in the first aspect.

[0016] Fourthly, this application provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the deep learning-based method for predicting the surface morphology and wettability of laser-processed surfaces as described in the first aspect.

[0017] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the deep learning-based method for predicting the surface morphology and wettability of laser-processed surfaces as described in the first aspect.

[0018] This application proposes a deep learning-based method for predicting the surface morphology and wettability of laser-processed surfaces, realizing the generation of SEM images of laser-processed surfaces and the prediction of the wettability of corresponding samples. Specifically, the main beneficial effects of this application include: (1) The technical solution provided in this application first generates SEM images based on laser processing parameter data combined with a conditional generative adversarial network. Then, it combines the SEM image data, the generated SEM images, and the corresponding fused data of laser processing parameters with surface morphology to generate a model to predict the wettability of the component. This method is based on a specialized deep learning model, which organically integrates laser processing parameter data and surface morphology data. This not only improves the model performance and generalization ability but also outputs highly accurate prediction results.

[0019] (2) The technical solution provided in this application can simultaneously achieve the synergistic prediction of the surface structure and wettability of laser processing.

[0020] (3) The technical solution provided in this application can reduce experimental costs, shorten the research and development cycle, predict results with high accuracy, and improve the intelligence level of laser functional surface design. Attached Figure Description

[0021] Figure 1 A schematic flowchart of the wettability prediction method based on a laser-processed surface morphology generation model provided in this application embodiment; Figure 2 A flowchart of laser processing and chemical treatment provided for embodiments of this application; Figure 3 This is a schematic diagram of the structure of the surface morphology generation model provided in the embodiments of this application; Figure 4 This is a schematic diagram of the structure of the self-attention module provided in an embodiment of this application; Figure 5 A schematic diagram of the bottleneck block provided in an embodiment of this application; Figure 6 SEM images of laser-processed surfaces of some samples provided in the embodiments of this application; Figure 7 This is a graph showing the performance evaluation results of the surface generation model training process provided in an embodiment of this application. Figure 8 A comparison diagram of the model-generated image results and the real image provided in the embodiments of this application; Figure 9 The model provided in this application provides the prediction results of wettability of laser-processed surfaces. Figure 10 This is a schematic diagram of the structure of the laser processing surface morphology and wettability prediction device based on deep learning provided in the embodiments of this application; Figure 11This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0022] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0024] In the description of this application, it should also be noted that, unless otherwise expressly specified and limited, the terms "set up," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this technology based on the specific circumstances.

[0025] In the description of this application, spatial relation terms such as "below," "under," "below," "below," "above," "over," etc., are used herein to describe the relationship between one element or feature shown in the figures and other elements or features. It should be understood that, in addition to the orientation shown in the figures, spatial relation terms also include different orientations of the device in use and operation. For example, if the device in the figures is flipped, an element or feature described as "below" or "under" or "below" of other elements or features will be oriented "above" other elements or features. Therefore, the exemplary terms "below" and "under" can include both upper and lower orientations. Furthermore, the device may also include other orientations (e.g., rotated 90 degrees or other orientations), and the spatial descriptive terms used herein are interpreted accordingly.

[0026] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use this application. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that this application can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid unnecessarily obscuring the description of this application. Therefore, this application is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0027] Existing generative models often fail to adequately incorporate laser processing parameters as constraints, resulting in unclear correspondence between the generated results and the actual process, making it difficult to directly guide the actual processing.

[0028] Therefore, this application provides a deep learning-based method for predicting the surface morphology and wettability of laser-processed surfaces.

[0029] See Figure 1 The deep learning-based method for predicting the surface morphology and wettability of laser-processed surfaces, as provided in this application embodiment, includes: S101. Obtain SEM morphology data of the component after processing with different laser processing parameters.

[0030] In one possible implementation, the material of the component includes one of aluminum alloy, titanium alloy, copper alloy, stainless steel, zirconium oxide, and silicon carbide; the laser processing parameters include laser scanning speed, scanning spacing, and single pulse energy.

[0031] Specifically, the laser processing parameters are as follows: The nanosecond laser has a wavelength of 1064 nm, a pulse width of 80 ns, a focused spot diameter of approximately 50 μm, a scanning speed of 20~150 mm / s, two scans, a single pulse energy of 15~160.71 μJ, a scanning interval of 4~20 μm, and adopts a linear scanning method. The scanning area is a rectangle of 20 mm × 20 mm, and the specific area size can be enlarged or reduced as needed.

[0032] S102. Obtain the wettability data of the metal component after laser processing and chemical-thermal treatment; the chemical-thermal treatment changes the wettability of the surface through organic solution immersion treatment.

[0033] In one possible implementation, the chemical-thermal treatment method includes: A silicone oil-propanol mixture is dropped onto the surface of the laser-processed sample. After the solution has soaked the sample, the surface is heated on a heating plate. After cooling naturally in the air, the sample is washed in an ethanol solution to remove contaminants. The wettability of the washed sample changes from hydrophilic to hydrophobic or superhydrophobic.

[0034] Furthermore, the volume fraction of propanol in the silicone oil-propanol mixed solution is 96%, and the volume fraction of silicone oil is 4%.

[0035] Furthermore, the heating temperature is 200°C and the heating time is 15 minutes.

[0036] See Figure 2 , Figure 2 A schematic diagram of the laser processing (a) and surface modification (b) process is shown.

[0037] S103. Preprocess the SEM morphology data and laser processing parameter data to construct a mapping dataset of laser processing parameters, SEM morphology data and wettability data of metal components.

[0038] In one possible implementation, the method for preprocessing SEM topography data includes: resizing the SEM image, random cropping, and normalization.

[0039] Specifically, the normalized grayscale range of the SEM image is (-1, 1).

[0040] In one possible implementation, the method for preprocessing laser processing parameter data includes: decomposing the single-pulse energy into power and frequency parameters and performing normalized encoding.

[0041] Specifically, the normalization method for laser processing parameter data is linear normalization, and its data range is mapped to [0,1].

[0042] Specifically, the encoding format of the laser processing parameter data is [scanning spacing (UTF-8), scanning speed (UTF-8), average power (UTF-8), single pulse energy (UTF-8), scanning frequency (UTF-8)].

[0043] S104. Construct and train a surface morphology generation model based on a conditional generative adversarial network. Using laser processing parameters as conditional input, the generator generates the corresponding surface morphology image, and the discriminator performs discrimination training.

[0044] In one possible implementation, see [link to relevant documentation] Figure 3 The surface morphology generation model includes a generator and a discriminator; The generator sequentially comprises: an input layer, a conditional fusion module, a first deconvolution module, a first deconvolution module with residuals, a self-attention module, a second deconvolution module, a second deconvolution module with residuals, a third deconvolution module, a third convolution module with residuals, and a fourth deconvolution module; the deconvolution module includes a deconvolution layer and a ReLU function; The discriminator sequentially comprises: an input layer, a first convolutional module with residuals, a first convolutional module, a second convolutional module with residuals, a second convolutional module, a third convolutional module with residuals, an auto-attention block, a third convolutional module, a fourth convolutional module with residuals, a flattening layer, a linear transformation layer, and an output layer; the convolutional module includes a convolutional layer and a LeakyReLU function.

[0045] Furthermore, the model parameters of the surface morphology generation model are updated using the Adam optimizer, and the loss function is a combination of Wasserstein loss and L1 loss.

[0046] The reasoning process of the surface morphology generation model is described below. (See also...) Figure 3 .

[0047] I. The reasoning process of the generator 1. Input layer Input data: random noise (dimension [B, n_dim]), laser processing parameters (dimension [B, params_dim]).

[0048] 2. Conditional parameter encoding layer Fully connected encoding: Laser processing parameters are mapped to new feature vectors through three fully connected layers + LeakyReLU activation, with dimensional transformation [B, params_dim] → [B, n_dim].

[0049] 3. Fusion of noise vector and conditional features (conditional fusion module) Feature concatenation: Concatenate the random noise and conditional feature vectors along the channel dimension to obtain the total generated input. Dimension transformation: [B,n_dim] + [B,n_dim] → [B, 2×n_dim].

[0050] 4. Vector Conversion Fully connected layer for dimensionality enhancement: The concatenated vector is mapped to a high-dimensional vector through a fully connected layer, with the dimensionality transformation [B, 2×n_dim] → [B, 256×16×16].

[0051] Dimension Reshaping: Reshape the high-dimensional vector into a convolutional feature map, dimensional transformation [B, 256×16×16] → [B,256, 16, 16].

[0052] 5. Deconvolution upsampling Phase 1 (1×1 sampling module): 256 channels → 64 channels (size unchanged) 1×1 deconvolution: channel dimensionality reduction without changing spatial dimensions, dimensionality transformation [B,256,16,16]→ [B,64,16,16].

[0053] Batch normalization + ReLU activation: standardizes feature distribution, improves training stability, and extracts non-linear features.

[0054] Phase 2 (4×4 sampling module): 64 channels → 128 channels 4×4 deconvolution: Spatial size upsampling by 2 times, channel dimensionality upsampling, dimensionality transformation [B,64,16,16] → [B,128,32,32]. (First upsampling) Batch normalization + ReLU activation: stabilizes training and enhances feature representation.

[0055] Residual blocks: Solve the vanishing gradient problem in deep networks by using residual connections, deepening the network while ensuring training stability.

[0056] Self-attention module (core layer): Data flow: Feature map generates query Q, key K, and value V; attention weights are calculated; global features are weighted and fused; residuals are returned. Function: Captures global spatial dependencies in an image, overcoming the limitation of ordinary convolution that can only extract local features; Advantages: Makes the structure of the generated image more reasonable and the texture more uniform, greatly improving the realism.

[0057] Phase 3 (1×1 sampling module): 128 channels → 32 channels (size unchanged) 1×1 deconvolution: channel compression and dimensionality reduction, spatial size remains 32×32, dimensionality transformation [B,128,32,32] → [B,32,32,32].

[0058] Batch normalization + ReLU activation: standardize features and extract non-linear features.

[0059] Phase 4 (4×4 sampling module): 32 channels → 64 channels (second upsampling × 2) 4×4 deconvolution: upsamples the spatial size by 2 times, increases the channel dimension, and transforms the dimension [B,32,32,32] → [B,64,64,64].

[0060] Batch normalization + ReLU activation: stabilizes training and optimizes feature distribution.

[0061] Residual blocks: Deepen the network, avoid gradient vanishing, and improve the depth of feature extraction.

[0062] Phase 5 (1×1 sampling module): 64 channels → 8 channels (size unchanged) 1×1 deconvolution: Significantly compresses the number of channels while maintaining a spatial size of 64×64. Dimension transformation [B,64,64,64] → [B,8,64,64].

[0063] Batch normalization + ReLU activation: standardize features and simplify feature dimensions.

[0064] Phase 6 (4×4 sampling module): 8 channels → 16 channels (third upsampling × 2) 4×4 deconvolution: upsamples the spatial size by 2 times, increases the channel dimension, and transforms the dimension [B,8,64,64] → [B,16,128,128].

[0065] Batch normalization + ReLU activation: stabilizes training and enhances feature representation.

[0066] Residual blocks: Deepen the network and ensure that deep networks are trainable.

[0067] 6. Output layer (1×1 sampling module + tanh activation) Final deconvolution: Fourth upsampling, the number of channels is converted to the target number of channels in the image, and the dimension is transformed [B,16,128,128] → [B,1,256,256].

[0068] tanh activation: Normalizes the output value to the range [-1, 1], matching the normalized image format.

[0069] Output: Generates the final controllable high-resolution image with dimensions [B, 1, 256, 256].

[0070] II. The reasoning process of the discriminator: 1. Input layer Input data: The generated image (img) and laser processing parameters (params) are the basic inputs for judging authenticity and regression parameters.

[0071] Data augmentation: Perform DiffAugment data augmentation on the input images (real image X and generated image G) to improve the model's generalization ability and training stability.

[0072] 2. Conditional parameter space expansion Dimension expansion: The one-dimensional laser processing parameters (params) are expanded into a two-dimensional spatial feature map, while maintaining the same size as the input image.

[0073] Feature concatenation: The enhanced image features are concatenated with the two-dimensional spatial feature map in the channel dimension to form joint input features.

[0074] 3. Convolutional Feature Extraction Backbone Layer (Stage-by-Stage Inference) Phase 1 (4×4 sampling module): Image size from 256 to 128 4×4 convolution: downsampling compresses the spatial size, converts the channels to 16, and extracts basic image features.

[0075] LeakyReLU activation: Introduces non-linear feature representation capabilities, adapting to the training characteristics of GANs.

[0076] Residual blocks: optimize gradient propagation through residual connections, deepen the network, and avoid gradient vanishing.

[0077] Phase 2 (1×1 sampling module, 4×4 sampling module): Image size from 128 to 64 1×1 convolution: Compresses the channel dimension to 8, reducing computation and refining features.

[0078] LeakyReLU activation: enhances the representation of nonlinear features.

[0079] 4×4 convolution: Downsampling again, increasing the number of channels to 64, to extract deeper structural features.

[0080] LeakyReLU activation: optimizes feature distribution.

[0081] Residual blocks: Deepen the network, improve feature extraction depth and training stability.

[0082] Phase 3 (1×1 sampling module, 4×4 sampling module): Image size reduced from 64 to 32 (core phase) 1×1 convolution: channels are compressed to 32, reducing feature dimensions.

[0083] LeakyReLU activation: enhances nonlinear expression.

[0084] 4×4 convolution: downsampling, increasing the number of channels to 128, capturing high-level semantic features.

[0085] LeakyReLU activation: stable feature distribution.

[0086] Residual blocks: Further deepen the network and enhance its feature representation capabilities.

[0087] Self-attention block: Data flow: Feature map generates query Q, key K, and value V; attention weights are calculated; global features are weighted and fused; residuals are returned. Function: To establish global spatial dependencies in an image, overcoming the limitations of the local receptive field in traditional convolution; Advantages: Accurately captures long-distance correlation information in images, improving the accuracy of authenticity judgment and parameter regression.

[0088] Phase 4 (1×1 sampling module, 4×4 sampling module): Image size from 32 to 16 1×1 convolution: channels are compressed to 64, simplifying feature dimensions.

[0089] LeakyReLU activation: optimizes nonlinear features.

[0090] 4×4 convolution: The final downsampling increases the number of channels to 256, extracting top-level abstract features.

[0091] LeakyReLU activation: Completes the feature activation of the final convolutional layer.

[0092] Residual blocks: Further deepen the network and enhance its feature representation capabilities.

[0093] Flattening operation: Flatten the 3D convolutional feature map into a 1D vector to fit the input of the fully connected layer.

[0094] 4. Parametric Regression Branch Fully connected mapping: The flattened image features are mapped to parametric features through two fully connected layers.

[0095] Regression output: Outputs the predicted conditional parameters, realizing inverse regression from the image to the input parameters.

[0096] 5. Authenticity / Counterfeit Detection Branch Joint discrimination: Nonlinear mapping and feature fusion are performed on the flattened joint features through multiple fully connected layers.

[0097] Final output: Output a single-dimensional discrimination score, used to determine whether the input image is a real image or a generated image.

[0098] 6. Final Output Discrimination score: Used to distinguish between genuine and fake images.

[0099] Regression parameters: used to inversely predict the laser processing parameters corresponding to the generated image.

[0100] The training process of the surface morphology generation model is described below.

[0101] I. Training Generator: 1. Generate random noise: Construct a new random noise tensor for the generator to generate images.

[0102] 2. Generating fake images: Input noise and real laser processing parameters into the generator to obtain fake images to be optimized.

[0103] 3. Discriminator Inference: The fake image and the real laser processing parameters are input into the discriminator to obtain the discrimination score of the generated sample and the parameter prediction result.

[0104] 4. Generator Adversarial Loss: The negative mean of the discriminator's output scores is taken, i.e., the Wasserstein generation loss, which guides the generator to deceive the discriminator. The loss is responsible for aligning the distribution of real and fake images to achieve adversarial training.

[0105] 5. Calculate L1 loss (core): Use L1 loss to calculate the error between the predicted parameters of the generated sample and the true parameters, constrain the generated image to meet the condition parameters, and L1 loss ensures the controllability of the conditions and the consistency of the parameters of the generated image.

[0106] 6. Total loss = Wasserstein generative adversarial loss + gradient penalty + dynamic weights × L1 parameter loss.

[0107] 7. Generator parameter update: Gradient clearing, backpropagation of total loss, and optimizer update of generator parameters.

[0108] II. Training the Discriminator 1. Generate random noise: Construct a random noise tensor that meets the dimensionality requirements as the basic input signal for the generator.

[0109] 2. Generating fake images: Input the noise and the real laser processing parameters into the generator to obtain the generated fake images.

[0110] 3. Calculate gradient penalty: Based on the discriminator, real image and fake image, calculate the gradient penalty term required for WGAN-GP.

[0111] 4. Input a real image and real laser processing parameters to obtain the real sample discrimination score and laser processing parameter prediction results. Input a fake image and real laser processing parameters to obtain the fake sample discrimination score and laser processing parameter prediction results.

[0112] 5. Using mean squared error loss, calculate the error between the predicted laser processing parameters of the real sample and the actual laser processing parameters, and the error between the predicted laser processing parameters of the fake sample and the actual laser processing parameters.

[0113] 6. Calculate Wasserstein loss: Based on the discrimination scores of real and fake samples, calculate Wasserstein distance loss to measure the difference in distribution between real and fake samples.

[0114] 7. Total loss = Wasserstein loss + gradient penalty + regression loss of true parameters + regression loss of dummy parameters.

[0115] 8. Discriminator parameter update: Gradient is cleared to zero, gradient is calculated through backpropagation, the optimizer updates the discriminator parameters, and the loss value is recorded.

[0116] S105. Construct and train a surface wettability prediction model based on deep residual network branches and conditional parameter encoding branches, using SEM morphology data and corresponding laser processing parameters as dual-branch inputs and wettability data as outputs.

[0117] In one possible implementation, the surface wettability prediction model includes a ResNet152 network branch, a conditional parameter encoding branch, a feature fusion layer, and an output layer. The ResNet152 network branch includes: stem, four residual stages, and an identity mapping layer; the condition parameter encoding branch includes multiple fully connected layers for mapping low-dimensional condition parameters to high-dimensional condition parameters; the feature fusion layer is used to concatenate the features output from the two branches to achieve joint representation of surface image information and condition parameters; the output layer includes a fully connected layer and an inverse normalization layer.

[0118] Furthermore, in the ResNet152 network branch, the stem includes 7×7 convolutions and pooling; The four residual stages include: conv2_x with 3 bottleneck blocks, conv3_x with 8 bottleneck blocks, conv4_x with 36 bottleneck blocks, and conv5_x with 3 bottleneck blocks; each bottleneck block includes a 1×1 convolution, a 3×3 convolution, a 1×1 convolution, and a residual connection.

[0119] It should be noted that, as Figure 5 The diagram shown is a schematic of the bottleneck block.

[0120] The reasoning process of the surface wettability prediction model is described below.

[0121] I. Input Layer Input data: Single-channel surface image (real or generated SEM image) and laser processing parameters (params).

[0122] Image dimensions: B × 1 × H × W; Parameter dimensions: B × 5.

[0123] II. Image Feature Extraction Branch (ResNet152 Backbone) 1. Stem module (preprocessing + basic feature extraction) 7×7 convolution: Converts single-channel input to 64-channel output with a stride of 2 and padding of 3, completing spatial downsampling and extracting basic features such as image edges and textures. Channel change: 1 → 64.

[0124] Pooling layer: Further downsampling, compressing spatial size, and retaining core feature information.

[0125] 2. conv2_x residual stage (3 bottleneck blocks) Three bottleneck residual blocks are stacked consecutively. Each bottleneck block is subjected to 1×1 convolution for dimensionality reduction, 3×3 convolution for feature extraction, and 1×1 convolution for dimensionality increase in sequence. Gradient passthrough is achieved through residual connections, and the output feature dimension is: B × 256 × H / 4 × W / 4.

[0126] 3. conv3_x residual stage (8 bottleneck blocks) Eight bottleneck residual blocks are stacked consecutively, with the same structure as described above, to complete downsampling and channel expansion, and output feature dimensions: B × 512 × H / 8 × W / 8.

[0127] 4. conv4_x residual stage (36 bottleneck blocks) Stacking 36 bottleneck residual blocks consecutively, the output feature dimension is: B × 1024 × H / 16 × W / 16.

[0128] 5. conv5_x residual stage (3 bottleneck blocks) Three bottleneck residual blocks are stacked consecutively to complete the final deep feature extraction, and the output feature dimension is B × 2048 × H / 32 × W / 32.

[0129] 6. Identity Mapping Layer: Replace the original classification layer (global average pooling + fully connected layer) with identity mapping, and output a fixed-dimensional 2048-dimensional image feature vector with a dimension of B × 2048. Its function is to output high-level, compact, and semantically strong image features.

[0130] III. Conditional Parameter Encoding Branch The first fully connected layer maps 5-dimensional parameters to 64-dimensional parameters, and ReLU activation introduces non-linear feature representation.

[0131] The second layer is fully connected: 64 dimensions are mapped to 512 dimensions, and ReLU is activated.

[0132] The third layer is fully connected: 512 dimensions are mapped to 2048 dimensions, and ReLU is activated.

[0133] The conditional parameter encoding branch ultimately outputs a parameter feature vector with dimensions B × 2048. This branch is used to map low-dimensional conditional parameters to high-dimensional features, aligning them with the image feature dimensions.

[0134] IV. Feature Fusion Layer Feature stitching: The 2048-dimensional image features and the 2048-dimensional parameter features are stitched together in dimension 1. After fusion, the feature dimension is B × 4096, realizing the joint representation of surface image information and condition parameters.

[0135] V. Joint discrimination and output branch The first layer is fully connected: 4096-dimensional fused features are mapped to 512 dimensions, and ReLU activation is applied.

[0136] The second fully connected layer maps 512 dimensions to a 1-dimensional output, resulting in B × 1.

[0137] Output the final prediction result: The final predicted value of wetting performance is obtained by denormalizing B × 1 through the denormalization layer.

[0138] S106. Input the laser processing parameters of the sample to be tested into the trained surface morphology generation model and output the generated SEM image; input the generated SEM image and the corresponding laser processing parameters as data into the trained surface wettability prediction model and output the predicted wettability value.

[0139] The following description, in conjunction with more detailed embodiments, provides further details.

[0140] Example 1: Step 10) Clean the aluminum alloy sample with ethanol and then with deionized water using ultrasonic cleaning, each cleaning time being 5 minutes, to remove surface contaminants. Then place the sample under a hair dryer to dry it.

[0141] Step 20) Fix the sample on the laser processing platform and perform scanning ablation using a nanosecond laser. The laser wavelength is 1064 nm, the pulse width is 80 ns, the frequency is 200 kHz, the average power is set to 15 W, the scanning rate is 20 mm / s, the scanning spacing is 20 µm, and the scanning area is 5 mm × 5 mm.

[0142] Step 30) The laser-processed sample is first soaked in a silicone oil-isopropanol solution, then placed on a heating stage at 200°C and heated for 15 minutes. After heating, the surface is cleaned with ethanol, and finally the contact angle of the surface is measured using a contact angle measuring instrument.

[0143] Step 40) Obtain the SEM morphology data and laser processing parameter data of the laser-processed surface, and use the data as input to train the surface morphology generation model. The Adam optimizer is used for updates, with its β value set to (0, 0.9) and learning rate set to 0.0001. The loss function is a combination of Wasserstein loss and L1 loss, and a self-attention mechanism is employed during model building.

[0144] Example 2: Step 10) The titanium alloy sample is ultrasonically cleaned with ethanol and deionized water in sequence, with each cleaning time being 5 minutes to remove surface contaminants.

[0145] Step 20) Fix the sample on the laser processing platform and perform scanning ablation using a nanosecond laser. The laser wavelength is 1064 nm, the pulse width is 80 ns, the frequency is 112 kHz, the average power is set to 15 W, the scanning rate is 20 mm / s, the scanning spacing is 10 µm, and the scanning area is 5 mm × 5 mm.

[0146] Step 30) The chemical treatment method is the same as in Example 1, and surface contact angle data is obtained.

[0147] Step 40) Integrate the real SEM image and the generated SEM image of the laser-processed sample with their corresponding laser processing parameters (real SEM image + processing parameters, generated SEM image + processing parameters) as a surface wettability prediction model, and output the model's prediction result for surface wettability. The model was validated on a sample using these parameters. The wettability prediction error based on the real image + processing parameters was 6.3°, while the error based on the generated image + processing parameters was 3.9°. This verifies the accuracy and practicality of this method in wettability prediction.

[0148] The following analysis is based on the relevant test results.

[0149] I. Surface morphology: Figure 6SEM images of the laser-processed surfaces of some samples are shown. It can be seen that the single-pulse energy determines the degree of surface etching, while the laser scanning speed affects the morphology of the etched surface. When the single-pulse energy is relatively low, the surface morphology is relatively smooth, and laser processing marks are difficult to discern. As the single-pulse energy increases, obvious etching streaks or pits appear on the surface. Under lower scanning speed conditions, the laser pulse has a high overlap rate between adjacent scanning tracks, making the material surface more uniformly heated, thus tending to form continuous etching streaks; while at higher scanning speeds, the pulse overlap rate decreases, local energy distribution is uneven, and the surface is more prone to forming discrete etching pits. With further increases in single-pulse energy, the thermal etching effect of the laser on the material surface is significantly enhanced. The material absorbs a large amount of energy in a short period of time and undergoes a series of complex physical and chemical processes, including evaporation, ionization, plasma jetting, and material deposition, eventually forming micro / nanoparticles on the processed surface. Under high scanning speed conditions, due to the reduced efficiency of surface coating and re-condensation of molten material, the resulting micro / nanoparticles are smaller and more densely distributed. These small and densely stacked particles cause incident electrons to undergo multiple scattering and reflections within them, inhibiting electrons from escaping from the surface, thus resulting in deeper grayscale contrast in SEM images under high scanning speed conditions.

[0150] II. Model Training and Evaluation: Figure 7 This is a graph showing the performance evaluation results of the surface generation model training process. Figure 7 (a) shows the changes in the loss values ​​of the discriminator and generator during model training. It can be seen that in the early training phase (training epochs < 50), the discriminator's loss value is a large negative value and rapidly rises towards zero, indicating that the discriminator can learn the significant differences between real and generated images. Simultaneously, the generator's loss value exhibits drastic fluctuations, reflecting that the generator has failed to form a stable output, and the quality of the generated images still lags significantly behind real images. As the number of training epochs increases, the generator's ability gradually improves, and the distribution of generated samples increasingly approaches the distribution of real data. This indicates that the discriminator struggles to distinguish between real and fake samples, causing its loss value to gradually approach zero, while the generator's loss value shows a significant upward trend and gradually stabilizes. In the later stages of training, the generator's loss value fluctuates around -200 without severe bifurcation or abnormal oscillations, while the discriminator's loss value also remains within a relatively small fluctuation range. Figure 7Figures (b) and (c) show the RMSE results of the amplitude and angle values ​​of the generated images in the training and validation sets after Fourier transform. Amplitude typically reflects the model's sensitivity to energy distribution. As shown in Figure 7(b), the root mean square error (RMSE) of the validation set is higher than that of the training set. In the early stages of model training, the RMSE of both the training and validation sets shows a rapid decreasing trend, indicating that the model can quickly capture the main relationship between input features and the target. As the training process continues, the RMSE of the training set further decreases to approximately 45, while the RMSE of the validation set fluctuates around 60. The RMSE of the angle typically reflects the difference in the phase distribution (i.e., spatial structure) of the samples in the Fourier frequency domain. It can be seen that the RMSE of the angle varies very little throughout the training process, and the errors of the training and validation sets remain stably within a narrow range (approximately 2.55~2.58). This means that the angle features of the samples are more stable than the amplitude features in the Fourier frequency domain. During training, the RMSE curves of the angles of the training and validation sets fluctuate alternately with little difference.

[0151] III. Evaluation of Generated Images: Figure 8 The comparison between the model-generated images and real images is shown. The quality of the generated images can be clearly assessed from the figures. As training progresses, the generated images gradually evolve from an initial blurry structure to images with clear texture and morphological features. Overall, the trained surface topography generation model can effectively learn the key structural and texture features of samples obtained under different laser processing parameters, including stripe structures, fine particles, pore morphology, and multi-scale particle aggregation structures. Furthermore, the generated images show a high degree of similarity to real images in terms of morphological features and statistical distribution.

[0152] IV. Wettability Prediction Analysis: Figure 9 This is an analysis of the surface wettability prediction results of the surface wettability prediction model for laser-processed surfaces. Figure 9 (a) is the prediction result of the training set without inputting laser processing parameters; Figure 9 (b) shows the predicted results of the real and generated images of the test set without inputting laser processing parameters; Figure 9 (c) is the prediction result of the training set after inputting the laser processing parameters; Figure 9 (d) shows the prediction results of the real and generated images of the test set after inputting laser processing parameters. It can be seen that without introducing laser processing parameters, the surface wettability prediction model has poor performance and generalization ability; its RMSE and R0 of the real image are significantly lower. 2The MAE values ​​were 22.395, 0.281, and 13.2°, respectively. However, the prediction results based on the generated image showed a more significant bias, with its RMSE increasing significantly to 59.044. 2 The value was negative (-3.998). After introducing the laser parameter, its generalization performance on generated samples was significantly improved. The root mean square error (RMSE) of its predictions based on real images was 4.832, with a correlation coefficient of 0.967, while the RMSE of its predictions based on generated images was only 5.765, and the correlation coefficient remained at a high level. Furthermore, the mean absolute error (MAE) values ​​for both real and generated images were relatively low, at 4.7° and 5.1°, respectively.

[0153] The following describes the deep learning-based laser processing surface morphology and wettability prediction device provided in this application. The deep learning-based laser processing surface morphology and wettability prediction device described below can be referred to in correspondence with the deep learning-based laser processing surface morphology and wettability prediction method described above.

[0154] Figure 10 This is a schematic diagram of the structure of the deep learning-based laser processing surface morphology and wettability prediction device provided in the embodiments of this application, as shown below. Figure 10 As shown, it includes: an acquisition module 101, a dataset construction module 102, a first model construction module 103, a second model construction module 104, and an output module 105, wherein: The acquisition module 101 is used to acquire SEM morphology data of the component after processing with different laser processing parameters; acquire wettability data of the laser-processed metal component after chemical-thermal treatment; the chemical-thermal treatment transforms the surface wettability through organic solution immersion treatment; The dataset construction module 102 is used to preprocess SEM morphology data and wettability data to construct a mapping dataset of laser processing parameters, SEM morphology data and wettability data of metal components. The first model building module 103 is used to build and train a surface morphology generation model based on a conditional generative adversarial network. The laser processing parameters are used as conditional inputs. The generator generates the corresponding surface morphology image, and the discriminator performs discrimination training. The second model construction module 104 is used to construct and train a surface wettability prediction model based on deep residual network branches and conditional parameter encoding branches, with SEM morphology data and corresponding laser processing parameters as dual-branch inputs and wettability data as output. The output module 105 is used to input the laser processing parameters of the sample to be tested into the trained surface morphology generation model and output the generated SEM image; and to input the generated SEM image and the corresponding laser processing parameters into the trained surface wettability prediction model and output the predicted wettability value.

[0155] Figure 11 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 11 As shown, the electronic device may include a processor 1110, a communications interface 1120, a memory 1130, and a communications bus 1140. The processor 1110, communications interface 1120, and memory 1130 communicate with each other via the communications bus 1140. The processor 1110 can call logic instructions from the memory 1130 to execute a deep learning-based method for predicting the surface morphology and wettability of laser-processed surfaces.

[0156] Furthermore, the logical instructions in the aforementioned memory 1130 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0157] On the other hand, this application also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the deep learning-based laser processing surface morphology and wettability prediction method provided by the above methods.

[0158] In another aspect, this application also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, is implemented to perform the deep learning-based laser processing surface morphology and wettability prediction methods provided by the above methods.

[0159] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0160] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0161] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for predicting laser processing surface topography and wettability based on deep learning, characterized in that, include: Obtain SEM morphology data of components after processing with different laser processing parameters; Acquire wettability data of laser-processed metal components after chemical-thermal treatment; the chemical-thermal treatment transforms the surface wettability through organic solution immersion treatment; Preprocessing of SEM morphology data and laser processing parameter data is performed to construct a mapping dataset of laser processing parameters, SEM morphology data and wettability data of metal components; A surface morphology generation model is constructed and trained based on a conditional generative adversarial network. Laser processing parameters are used as conditional inputs. The generator produces the corresponding surface morphology image, and the discriminator performs discrimination training. A surface wettability prediction model was constructed and trained based on deep residual network branches and conditional parameter encoding branches, with SEM morphology data and corresponding laser processing parameters as dual-branch inputs and wettability data as output. The laser processing parameters of the sample to be tested are input into the trained surface morphology generation model, which outputs the generated SEM image. The generated SEM image and the corresponding laser processing parameters are input into the trained surface wettability prediction model, which outputs the predicted wettability value.

2. The method of claim 1, wherein, The material of the component includes one of aluminum alloy, titanium alloy, copper alloy, stainless steel, zirconium oxide and silicon carbide; the laser processing parameters include laser scanning speed, scanning spacing and single pulse energy.

3. The method of claim 1, wherein, The surface wettability prediction model includes: a generator and a discriminator; The generator sequentially comprises: an input layer, a conditional fusion module, a first deconvolution module, a first deconvolution module with residuals, a self-attention module, a second deconvolution module, a second deconvolution module with residuals, a third deconvolution module, a third convolution module with residuals, and a fourth deconvolution module; the deconvolution module includes a deconvolution layer and a ReLU function; The discriminator sequentially comprises: an input layer, a first convolutional module with residuals, a first convolutional module, a second convolutional module with residuals, a second convolutional module, a third convolutional module with residuals, an auto-attention block, a third convolutional module, a fourth convolutional module with residuals, a flattening layer, a linear transformation layer, and an output layer; the convolutional module includes a convolutional layer and a LeakyReLU function.

4. The method of claim 1, wherein, The model parameters of the surface morphology generation model are updated using the Adam optimizer, and the loss function is a combination of Wasserstein loss and L1 loss.

5. The method of claim 1, wherein, The surface wettability prediction model includes a ResNet152 network branch, a conditional parameter encoding branch, a feature fusion layer, and an output layer. The ResNet152 network branch includes: stem, four residual stages, and an identity mapping layer; the condition parameter encoding branch includes multiple fully connected layers for mapping low-dimensional condition parameters to high-dimensional condition parameters; the feature fusion layer is used to concatenate the features output from the two branches to achieve joint representation of surface image information and condition parameters; the output layer includes a fully connected layer and an inverse normalization layer.

6. The method of claim 5, wherein, In the ResNet152 network branch, stem includes 7×7 convolution and pooling; The four residual stages include: conv2_x with 3 bottleneck blocks, conv3_x with 8 bottleneck blocks, conv4_x with 36 bottleneck blocks, and conv5_x with 3 bottleneck blocks; each bottleneck block includes 3 convolutions and residual connections.

7. A device for predicting the surface topography and wettability of a laser-processed surface based on deep learning, characterized by, include: The acquisition module is used to acquire SEM morphology data of components after processing with different laser processing parameters; acquire wettability data of laser-processed metal components after chemical-thermal treatment; the chemical-thermal treatment transforms the surface wettability through organic solution immersion treatment; The dataset construction module is used to preprocess SEM morphology data and wettability data to construct a mapping dataset of laser processing parameters, SEM morphology data and wettability data of metal components; The first model building module is used to build and train a surface morphology generation model based on a conditional generative adversarial network. It takes laser processing parameters as conditional input, generates the corresponding surface morphology image through the generator, and performs discrimination training by the discriminator. The second model building module is used to build and train a surface wettability prediction model based on deep residual network branches and conditional parameter encoding branches. It uses SEM morphology data and corresponding laser processing parameters as dual-branch inputs and wettability data as output. The output module is used to input the laser processing parameters of the sample to be tested into the trained surface morphology generation model and output the generated SEM image; and to input the generated SEM image and the corresponding laser processing parameters into the trained surface wettability prediction model and output the predicted wettability value.

8. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the deep learning-based method for predicting the surface morphology and wettability of laser-processed surfaces as described in any one of claims 1 to 6. 9.A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by the processor, it implements the deep learning-based method for predicting the surface morphology and wettability of laser-processed surfaces as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the deep learning-based method for predicting the surface morphology and wettability of laser-processed surfaces as described in any one of claims 1 to 6.

Images