Autonomous preparation method of rotating electrode film based on image segmentation and Bayesian optimization

Abstract

The invention provides a rotating electrode film autonomous preparation method based on image segmentation and Bayesian optimization. The method comprises the steps of constructing a film image data set and training an image segmentation model, calculating a film quality score based on an image segmentation result, and iteratively searching an optimal preparation parameter based on Bayesian optimization. According to the method, a trial and error process depending on subjective experience is converted into a closed-loop automatic optimization process driven by objective data, uncertainty of manual operation can be eliminated, and standardization of the preparation process is achieved; by means of the efficient global search capability of the Bayesian optimization algorithm, the parameter optimization process originally needing several days or even several weeks is shortened to be completed within several hours, and the research and development efficiency can be greatly improved; the high-quality thin film can be stably prepared by directly associating preparation parameters with quantitative image quality scores, so that the accuracy and the reliability of subsequent electrochemical testing are improved.

Description

Technical Field

[0001] This invention belongs to the field of electrochemical rotating electrode thin film preparation technology, specifically involving an autonomous method for preparing rotating electrode thin films based on image segmentation and Bayesian optimization. Background Technology

[0002] In electrocatalysis research (such as oxygen reduction reaction (ORR) and oxygen evolution reaction (OER), rotating electrodes are a key tool for evaluating the intrinsic properties of catalysts. However, the accuracy and reproducibility of the test results are highly dependent on the quality of the catalyst film coated on the electrode surface. Currently, traditional catalyst film preparation methods heavily rely on the operational experience of researchers. Ink formulations (such as solvent ratios and additive content), coating parameters (such as spin coating speed), and drying conditions are adjusted manually through trial and error to achieve uniformity, defect-free, and consistent films. This process typically takes several days or even weeks, is extremely inefficient, and has become a bottleneck restricting the screening and development of high-throughput catalysts.

[0003] Although machine learning and optimization algorithms are increasingly used in materials science, their application in guiding and optimizing physical preparation processes, especially the preparation of complex catalyst films, remains an underdeveloped area. Current technologies lack a closed-loop system capable of real-time, objectively evaluating film quality and automatically and efficiently optimizing preparation parameters accordingly. Summary of the Invention

[0004] (a) Technical problems to be solved This invention proposes an autonomous preparation method for rotating electrode thin films based on image segmentation and Bayesian optimization, in order to solve the technical problem of achieving automated and standardized preparation of catalyst thin films in electrochemical testing.

[0005] (II) Technical Solution To address the aforementioned technical problems, this invention proposes a method for autonomous fabrication of rotating electrode thin films based on image segmentation and Bayesian optimization. This method includes the following steps: S1. Constructing a thin-film image dataset and training an image segmentation model S1-1. Multiple rotating electrode thin film samples were prepared using an automated rotating electrode thin film preparation equipment with multiple different combinations of preparation parameters; S1-2. Use an image acquisition device to acquire optical images of the rotating electrode thin film sample; S1-3. Label the optical images of the rotating electrode thin film samples with feature regions, which include at least the effective coverage area, the overly thin area, the defective area, and the overflow area, and generate the corresponding mask files to construct the training dataset; S1-4. A neural network model is used and trained using a training dataset to obtain an image segmentation model capable of segmenting optical images of rotating electrode thin films. S2. Calculate film quality score based on image segmentation results. S2-1. The prepared electrode film image is segmented using the trained image segmentation model to identify each feature region; S2-2. Extract multiple feature parameters from the segmentation results, including at least the standard deviation of the gray value of the effective coverage area, the proportion of the area of ​​the effective coverage area to the entire electrode area, the standard deviation of the gray value of the overly thin area, the proportion of the area of ​​the overly thin area to the entire electrode area, the proportion of the area of ​​the defective area to the entire electrode area, and the proportion of the area of ​​the overflow area to the entire electrode area. S2-3. Substitute the extracted feature parameters into the preset quality scoring function to calculate the film quality score; S3. Finding the optimal preparation parameters based on Bayesian optimization iteration S3-1. Determine the preparation parameters of the rotating electrode thin film to be optimized and their value range, and use the thin film quality score as the optimization target; S3-2. Generate multiple sets of initial preparation parameter combinations and execute step S2 to obtain the initial preparation parameter-scoring dataset; S3-3. Based on the initial parameter-scoring dataset, establish a surrogate model; S3-4. Use the acquisition function to select the next candidate preparation parameter combination from the surrogate model; send the selected preparation parameter combination to the rotary electrode thin film automated preparation equipment to perform thin film preparation, image acquisition and quality scoring, and add the new data points to the dataset to update the surrogate model; S3-5: Repeat step S3-4 until the termination condition is met, and output the optimal rotating electrode thin film preparation parameters.

[0006] Furthermore, in steps S1-4, a neural network model with an encoder-decoder structure is adopted.

[0007] Furthermore, in steps S1-4, the image segmentation model is a U-Net architecture, and the training process includes a frozen training phase in which the encoder weights are frozen and only the decoder is trained, and a thawed training phase in which the encoder is unfrozen and the entire model is fine-tuned.

[0008] Furthermore, in steps S2-3, the quality scoring function Score for:

[0009] In the formula, To effectively cover the standard deviation of gray values ​​in the area, The effective coverage area is the proportion of the total electrode area. The standard deviation of gray values ​​in the excessively thin region. This refers to the proportion of the area of ​​the excessively thin region to the total electrode area. c This represents the proportion of the defective area to the total electrode area. d This represents the proportion of the overflow area to the total electrode area. K 1. K 2. K 3. K 4. K 5. K 6 represents a predefined weighting coefficient.

[0010] Furthermore, K 1=8, K 2=0.1, K 3=2, K 4 = 0.1, K 5 = 0.2, K 6 = 0.05.

[0011] Furthermore, in step S3-1, the preparation parameters include at least the solvent ratio, spin coating speed, and additive content.

[0012] Furthermore, in step S3-2, multiple sets of initial preparation parameter combinations are generated using the Latin hypercube sampling method.

[0013] Furthermore, in step S3-3, the surrogate model is a Gaussian process regression model.

[0014] Furthermore, in steps S3-4, the acquisition function is the desired improvement function.

[0015] Furthermore, in steps S3-5, the termination condition is that the quality score converges or the preset maximum number of iterations is reached.

[0016] (III) Beneficial Effects This invention proposes an autonomous fabrication method for rotating electrode thin films based on image segmentation and Bayesian optimization. The method includes constructing a thin film image dataset and training an image segmentation model, calculating thin film quality scores based on the image segmentation results, and iteratively finding the optimal fabrication parameters using Bayesian optimization. This invention transforms a trial-and-error process relying on subjective experience into a closed-loop automated optimization process driven by objective data, eliminating the uncertainty of human operation and standardizing the fabrication process. Utilizing the efficient global search capability of the Bayesian optimization algorithm, the parameter optimization process, which originally required days or even weeks, is shortened to within hours, significantly improving R&D efficiency. By directly correlating the fabrication parameters with quantitative image quality scores, high-quality thin films can be stably fabricated, thereby improving the accuracy and reliability of subsequent electrochemical tests.

[0017] The self-developed preparation method of this invention is applicable to different types of catalysts (such as platinum-carbon and nickel-iron catalysts), and has good versatility and promotion value. It is particularly suitable for high-throughput catalyst screening and the development of clean energy materials. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the self-preparation method according to an embodiment of the present invention; Figure 2a The image before segmentation. Figure 2b The image segmentation model is used to segment the thin film image. Figure 3a This is a loss rate graph for each epoch during the training of the image segmentation model. Figure 3b Miou plot for every five epochs; Figure 4a Optimization curves for commercial 60% platinum-carbon catalyst thin films. Figure 4b A graph showing the optimization process for preparing platinum-carbon catalysts in the laboratory. Figure 4c Optimization curves for commercial nickel-iron catalysts. Detailed Implementation

[0019] To make the objectives, contents, and advantages of the present invention clearer, the specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples.

[0020] This embodiment proposes an autonomous fabrication method for rotating electrode thin films based on image segmentation and Bayesian optimization, the process of which is as follows: Figure 1 As shown, the specific steps include the following: First, a dataset for training the image segmentation model was constructed. A large number of rotating electrode films were fabricated using the automated rotation platform autoRDE-film with different parameter combinations. After fabrication, images of the rotating electrode films were acquired using a high-resolution industrial camera. Then, the LabelMe tool was used to label each acquired rotating electrode film image, defining four regions: effective coverage area, overly thin area, defective area, and overflow area, and generating corresponding mask files. The dataset was then randomly divided into training and test sets in a 9:1 ratio.

[0021] Subsequently, a training method was adopted that uses the U-Net architecture as the segmentation model and fine-tunes the backbone network (VGG-16) with ImageNet pre-trained weights and a custom dataset.

[0022] Model training is divided into two stages: In the frozen training phase, leveraging the powerful feature extraction capabilities of the ImageNet pre-trained model, all parameters of the backbone network are fixed, and only the randomly initialized decoder part is trained. This phase utilizes general visual features learned from large-scale datasets to quickly adapt to the object segmentation task, while effectively preventing overfitting that may occur on small-scale datasets by limiting the number of trainable parameters. This phase accounts for 2 / 3 of the total training epochs, ensuring that the decoder can fully learn how to map high-level semantic features to specific segmentation masks.

[0023] In the unfreezing phase, the parameters of the backbone network are unfrozen, allowing the parameters of the entire model to participate in optimization. This phase further refines the pre-trained general features into task-specific features, achieving collaborative optimization between the encoder and decoder. Through such joint training, segmentation accuracy and generalization ability on the target dataset can be improved.

[0024] Freeze-train: Fix the backbone network weights and train only the decoder; this phase accounts for 2 / 3 of the total epochs. Unfreeze-train: Release all network parameters for fine-tuning; this phase accounts for 1 / 3 of the total epochs. Figure 3a and 3b As shown.

[0025] The rotating electrode thin film image is segmented based on the trained image segmentation model, and the following feature parameters are extracted: To effectively cover the standard deviation of gray values ​​in the area, The effective coverage area is the proportion of the total electrode area. The standard deviation of gray values ​​in the excessively thin region. This refers to the proportion of the area of ​​the excessively thin region to the total electrode area. c This represents the proportion of the defective area to the total electrode area. d This represents the proportion of the overflow area to the total electrode area.

[0026] Substitute the above parameters into the following quality scoring formula to calculate the quality score:

[0027] The higher the score, the better the film quality; the ideal score is 10.

[0028] The effects of segmenting the thin film image using an image segmentation model before and after are shown below. Figure 2a and 2b As shown.

[0029] Three key process parameters in thin film preparation were used as optimization parameters for Bayesian optimization: the water-to-isopropanol solvent ratio (7:3~9:1), spin coating speed (0~800rpm), and DMF content (0~1% v / v). Quality score was used as the optimization objective. The surrogate model used in the Bayesian optimization process was Gaussian Process Regression (GPR) with a Matern kernel; the acquisition function was Expected Improvement (EI). Before Bayesian optimization began, Latin hypercube sampling (LHS) was used to generate multiple initial parameter combinations. These initial schemes were prepared and scored using the automated preparation rotation platform autoRDE-film and an image segmentation model. Subsequently, Bayesian optimization began. In each iteration, the set of parameters with the highest EI value was selected. The quality score was prepared and calculated using autoRDE-film and the image segmentation model. This data point was added to the observation set, and the surrogate model was updated. This iteration cycle continued until the score converged or the maximum number of iterations was reached.

[0030] Figure 4a Optimization curves for commercial 60% platinum-carbon catalyst thin films. Figure 4b A graph showing the optimization process for preparing platinum-carbon catalysts in the laboratory. Figure 4c Table 1 shows the optimization process curves for commercial nickel-iron catalysts. Taking 60% platinum-carbon catalyst A as an example, the above system was used for optimization. Before iterative optimization, 12 sets of random parameter combinations were set. In these 12 random schemes, the prepared films generally exhibited significant defects such as agglomeration and coffee ring effects, with an average score of 6.5 ± 2.5, and their oxygen reduction reaction (ORR) half-wave potentials were all below 0.87 V vs. RHE. At the beginning of iterative optimization, to accumulate data, the three parameter combinations with the highest predicted scores were selected for each iteration. After the iteration began, excellent convergence efficiency was demonstrated: after 5 iterations, the film score improved to 9.3, and the film score prepared under these parameters remained stable above 9.3, with its oxygen reduction reaction (ORR) half-wave potential reaching 0.92 V vs. RHE, an improvement of approximately 50 mV compared to before optimization, significantly improving the accuracy of catalyst performance evaluation. This process took only 5 hours, and subsequent scores were all above 9.2.

[0031] For catalyst A, the final parameter combination was determined as follows: water to isopropanol ratio 8.1:1.9, drying speed 400 rpm, and additive content 0.5%.

[0032] Table 1 Partial optimization process of catalyst A

[0033] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for autonomously fabricating rotating electrode thin films based on image segmentation and Bayesian optimization, characterized in that, The method for independently preparing the rotating electrode thin film includes the following steps: S1. Constructing a thin-film image dataset and training an image segmentation model S1-1. Multiple rotating electrode thin film samples were prepared using an automated rotating electrode thin film preparation equipment with multiple different combinations of preparation parameters; S1-2. Use an image acquisition device to acquire optical images of the rotating electrode thin film sample; S1-3. Label the optical images of the rotating electrode thin film samples with feature regions, which include at least the effective coverage area, the overly thin area, the defective area, and the overflow area, and generate the corresponding mask files to construct the training dataset; S1-4. A neural network model is used and trained using a training dataset to obtain an image segmentation model capable of segmenting optical images of rotating electrode thin films. S2. Calculate film quality score based on image segmentation results. S2-1. The prepared electrode film image is segmented using the trained image segmentation model to identify each feature region; S2-2. Extract multiple feature parameters from the segmentation results, including at least the standard deviation of the gray value of the effective coverage area, the proportion of the area of ​​the effective coverage area to the entire electrode area, the standard deviation of the gray value of the overly thin area, the proportion of the area of ​​the overly thin area to the entire electrode area, the proportion of the area of ​​the defective area to the entire electrode area, and the proportion of the area of ​​the overflow area to the entire electrode area. S2-3. Substitute the extracted feature parameters into the preset quality scoring function to calculate the film quality score; S3. Finding the optimal preparation parameters based on Bayesian optimization iteration S3-1. Determine the preparation parameters of the rotating electrode thin film to be optimized and their value range, and use the thin film quality score as the optimization target; S3-2. Generate multiple sets of initial preparation parameter combinations and execute step S2 to obtain the initial preparation parameter-scoring dataset; S3-3. Based on the initial parameter-scoring dataset, establish a surrogate model; S3-4. Use the acquisition function to select the next candidate preparation parameter combination from the surrogate model; send the selected preparation parameter combination to the rotary electrode thin film automated preparation equipment to perform thin film preparation, image acquisition and quality scoring, and add the new data points to the dataset to update the surrogate model; S3-5: Repeat step S3-4 until the termination condition is met, and output the optimal rotating electrode thin film preparation parameters.

2. The method for autonomous fabrication of rotating electrode thin films based on image segmentation and Bayesian optimization as described in claim 1, characterized in that, In steps S1-4, a neural network model with an encoder-decoder structure is adopted.

3. The method for autonomous fabrication of rotating electrode thin films based on image segmentation and Bayesian optimization as described in claim 1, characterized in that, In steps S1-4, the image segmentation model is a U-Net architecture. The training process includes a frozen training phase where the encoder weights are frozen and only the decoder is trained, and a thawed training phase where the encoder is unfrozen and the entire model is fine-tuned.

4. The method for autonomous fabrication of rotating electrode thin films based on image segmentation and Bayesian optimization as described in claim 1, characterized in that, In steps S2-3, the quality scoring function Score for: In the formula, To effectively cover the standard deviation of gray values ​​in the area, The effective coverage area is the proportion of the total electrode area. The standard deviation of gray values ​​in the excessively thin region. This refers to the proportion of the area of ​​the excessively thin region to the total electrode area. c This represents the proportion of the defective area to the total electrode area. d This represents the proportion of the overflow area to the total electrode area. K 1. K 2. K 3. K 4. K 5. K 6 represents a predefined weighting coefficient.

5. The method for autonomous fabrication of rotating electrode thin films based on image segmentation and Bayesian optimization as described in claim 4, characterized in that, K 1=8, K 2=0.1, K 3=2, K 4=0.1, K 5=0.2, K 6=0.05。 6. The method for autonomous fabrication of rotating electrode thin films based on image segmentation and Bayesian optimization as described in claim 1, characterized in that, In step S3-1, the preparation parameters include at least the solvent ratio, spin coating speed, and additive content.

7. The method for autonomous fabrication of rotating electrode thin films based on image segmentation and Bayesian optimization as described in claim 1, characterized in that, In step S3-2, the Latin hypercube sampling method is used to generate multiple sets of initial preparation parameter combinations.

8. The method for autonomous fabrication of rotating electrode thin films based on image segmentation and Bayesian optimization as described in claim 1, characterized in that, In step S3-3, the surrogate model is a Gaussian process regression model.

9. The method for autonomous fabrication of rotating electrode thin films based on image segmentation and Bayesian optimization as described in claim 1, characterized in that, In step S3-4, the acquisition function is the desired improvement function.

10. The method for autonomous fabrication of rotating electrode thin films based on image segmentation and Bayesian optimization as described in claim 1, characterized in that, In steps S3-5, the termination condition is that the quality score converges or the preset maximum number of iterations is reached.

Images