A data-driven closed-loop iterative high-quantum-yield and high-luminous-intensity carbon quantum dot preparation method and system

Abstract

The application belongs to the technical field of intelligent green synthesis of carbon quantum dots, and particularly relates to a data-driven closed-loop iterative high-quantum-yield and high-luminous-intensity carbon quantum dot preparation method and system, which comprises the following steps: based on carbon quantum dot synthesis parameters and corresponding luminous performance indexes, an initial data set is constructed; a multi-target prediction model is constructed, and the initial data set is used to train the multi-target prediction model; based on a global optimization algorithm, the trained multi-target prediction model is iteratively optimized to obtain carbon quantum dot synthesis parameters satisfying preset performance conditions; and based on the carbon quantum dot synthesis parameters satisfying the preset performance conditions, carbon quantum dots satisfying the preset performance conditions are prepared. Through the closed-loop evolution mechanism of data feedback and experimental verification, the application realizes intelligent optimization of the carbon quantum dot preparation process.

Description

Technical Field

[0001] This invention belongs to the field of intelligent green synthesis technology of carbon quantum dots, specifically relating to a data-driven closed-loop iterative method and system for preparing carbon quantum dots with high quantum yield and high luminescence intensity. Background Technology

[0002] Carbon quantum dots (CQDs) have attracted widespread attention in fields such as bioimaging, sensing, photocatalysis, and light-emitting devices due to their excellent optical properties, low toxicity, good water solubility, and surface modifiability. Currently, the mainstream preparation routes for CQDs still rely on trial-and-error experimental models such as hydrothermal / solvothermal methods and microwave methods. Researchers manually adjust more than ten interdependent parameters, including carbon source, nitrogen source, dopant, pH, and temperature program, and screen for formulations with relatively superior luminescence performance (quantum yield QY, maximum photoluminescence intensity MPI) through multiple parallel experiments. This process has accumulated a large amount of empirical data and, with the help of single-factor or orthogonal experimental designs, has narrowed the search range to some extent. However, it still remains in a linear stage of "experiment-experiment-characterization," lacking the ability to systematically optimize high-dimensional parameter spaces, resulting in long development cycles, high material consumption, and low experimental reproducibility.

[0003] Although the above empirical methods can occasionally yield samples with QY>30%, the luminescence performance of CQDs is extremely sensitive to the coupling of multiple variables such as raw material ratio, temperature control program, and insulation platform. Traditional methods are difficult to achieve synergistic improvement of QY and MPI within a huge parameter space. More importantly, the lack of closed-loop feedback between experiment, data, and decision-making makes it impossible to use historical data for self-evolution, making it difficult to break through the "performance ceiling". The discovery of high-performance carbon quantum dots still has a significant element of chance, which seriously restricts the batch stability and performance predictability required for their high-end applications. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention provides a data-driven closed-loop iterative method and system for preparing carbon quantum dots with high quantum yield and high luminescence intensity. The aim is to achieve intelligent optimization of the carbon quantum dot preparation process through a closed-loop evolution mechanism of data feedback and experimental verification.

[0005] To achieve the above objectives, the present invention provides the following solution: A data-driven closed-loop iterative method for preparing carbon quantum dots with high quantum yield and high luminescence intensity, the method comprising: An initial dataset was constructed based on the synthesis parameters of carbon quantum dots and their corresponding luminescence performance indicators. Construct a multi-objective prediction model and train it using the initial dataset; Based on the global optimization algorithm, the trained multi-objective prediction model is iteratively optimized to obtain carbon quantum dot synthesis parameters that meet the preset performance conditions; Based on the carbon quantum dot synthesis parameters that meet the preset performance conditions, carbon quantum dots that meet the preset performance conditions are prepared.

[0006] Preferably, the method for constructing the initial dataset based on carbon quantum dot synthesis parameters and corresponding luminescence performance indicators includes: The raw materials for preparing carbon quantum dots were selected, including carbon source, nitrogen source, dopant source, pH adjuster and metal ion dopant elements; Several initial synthesis experiments were designed based on different synthesis conditions. Combined with sample raw materials, initial carbon quantum dots were synthesized through hydrothermal reaction using a preset temperature control program. The quantum yield and maximum photoluminescence intensity of the initial carbon quantum dots were determined to form an initial dataset containing the synthesis parameters of the carbon quantum dots and the corresponding luminescence performance indicators.

[0007] Preferably, the multi-objective prediction model is the CEM-MultiMax Net neural network model, which includes two branch output layers; The two branch output layers use the Sigmoid activation function and the Identity activation function, respectively, to output predicted values ​​of quantum yield and maximum photoluminescence intensity.

[0008] Preferably, the method for iteratively optimizing the trained multi-objective prediction model based on a global optimization algorithm to obtain carbon quantum dot synthesis parameters that meet preset performance conditions includes: Candidate synthesis parameters are generated by using the trained multi-objective prediction model and combining it with a global optimization algorithm. Experiments were conducted to verify the candidate synthesis parameters and obtain real luminescence performance data. The actual luminescence performance data is added to the initial dataset to obtain the supplementary dataset; The multi-objective prediction model was retrained using a supplementary dataset and iteratively optimized until carbon quantum dot synthesis parameters that meet the preset performance conditions were obtained.

[0009] The present invention also provides a data-driven closed-loop iterative system for preparing carbon quantum dots with high quantum yield and high luminescence intensity. The system is used to implement the aforementioned method and includes: a dataset construction module, a model construction and training module, an iterative optimization module, and a preparation module. The dataset construction module is used to construct an initial dataset based on the carbon quantum dot synthesis parameters and corresponding luminescence performance indicators; The model building and training module is used to build a multi-objective prediction model and train the multi-objective prediction model using the initial dataset. The iterative optimization module is used to iteratively optimize the trained multi-objective prediction model based on a global optimization algorithm to obtain carbon quantum dot synthesis parameters that meet preset performance conditions. The preparation module is used to prepare carbon quantum dots that meet preset performance conditions based on carbon quantum dot synthesis parameters.

[0010] Preferably, the dataset construction module includes: selection unit, synthesis unit, and construction unit; The selection unit is used to select the sample raw materials for preparing carbon quantum dots. The sample raw materials include carbon source, nitrogen source, dopant source, pH adjuster and metal ion dopant elements. The synthesis unit is used to design several sets of initial synthesis experiments based on different synthesis conditions, and synthesize initial carbon quantum dots through hydrothermal reaction using sample raw materials and a preset temperature control program. The building blocks are used to determine the quantum yield and maximum photoluminescence intensity of the initial carbon quantum dots, forming an initial dataset containing the carbon quantum dot synthesis parameters and corresponding luminescence performance indicators.

[0011] Preferably, the multi-objective prediction model is the CEM-MultiMax Net neural network model, which includes two branch output layers; The two branch output layers use the Sigmoid activation function and the Identity activation function, respectively, to output predicted values ​​of quantum yield and maximum photoluminescence intensity.

[0012] Preferably, the iterative optimization module includes: candidate units, verification units, supplementary units, and iterative units; Candidate units are used to generate candidate synthesis parameters by utilizing the trained multi-objective prediction model and combining it with a global optimization algorithm. The verification unit is used to perform experimental verification based on the candidate synthesis parameters and obtain real luminescence performance data. The supplementary unit is used to supplement the initial dataset with the actual luminescence performance data to obtain a supplementary dataset. The iterative unit is used to retrain the multi-objective prediction model using a supplementary dataset and iteratively optimize it until carbon quantum dot synthesis parameters that meet the preset performance conditions are obtained.

[0013] The present invention also provides an electronic device, including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor executes the computing program to implement the aforementioned method.

[0014] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned method.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention replaces the traditional, inefficient "trial and error" method with data-driven intelligent optimization. Through a closed-loop iterative approach, it can systematically and efficiently explore a broad and complex synthesis parameter space, significantly shortening the R&D cycle and saving experimental costs.

[0016] The multi-objective machine learning model constructed in this invention can simultaneously handle two key performance indicators, quantum yield (QY) and photoluminescence intensity (MPI), effectively solving the performance bias problem that may be caused by single-objective optimization and achieving a synergistic improvement in overall luminescence performance.

[0017] This invention employs an iterative feedback mechanism of "prediction-validation-retraining," continuously enhancing the model's capabilities. This mechanism guides experiments beyond the limitations of the initial dataset, enabling the exploration and discovery of novel synthesis conditions that can produce breakthrough performance. As demonstrated in the examples, this method successfully increased the QY of carbon quantum dots from an initial peak of 34.47% to 80.65%, and significantly enhanced the MPI, achievements that are difficult to attain with traditional methods or non-iterative machine learning approaches.

[0018] This invention, through data augmentation, standardization, a reasonable model architecture design, and optimization strategies (such as early stopping and learning rate scheduling), achieves a model with good prediction accuracy (R²). 2 The high value and anti-overfitting ability ensure the reliability of the optimization direction.

[0019] This invention, combined with interpretable machine learning tools (such as SHAP), not only finds the optimal formulation but also reveals key influencing factors, providing strong support for a deeper understanding of the structure-property relationship of materials and for conducting mechanistic studies. Attached Figure Description

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

[0021] Figure 1 A flowchart illustrating the dataset construction process in an embodiment of the present invention; Figure 2 This is a data processing flowchart of an embodiment of the present invention; Figure 3 This diagram illustrates the model training configuration and the first-round iterative optimization process based on CEM, as shown in this embodiment of the invention. Figure 4The above are correlation analysis diagrams of the predicted and actual values ​​of QY and MPI intensity based on the CEM-MultiMax Net model in this embodiment of the invention. (A) is a scatter plot of the correlation between the predicted and actual values ​​of QY, and (B) is a scatter plot of the correlation between the predicted and actual values ​​of MPI. Figure 5 The following is a variation analysis of the val loss and training loss based on the CEM-MultiMax Net model in this embodiment of the invention, and a variation analysis diagram of MAE. Among them, (A) is a variation analysis verification diagram of val loss and training loss, and (B) is a variation analysis diagram of MAE. Figure 6 This is a Shape-beeswarm plot analysis of the influence of 15 synthetic parameter features of CQDs on QY in an embodiment of the present invention; Figure 7 This is a Shape beeswarm plot analysis of the influence of 15 synthetic parameter features of CQDs on MPI intensity in an embodiment of the present invention; Figure 8 The image shows the ultraviolet-fluorescence spectrum and excitation-dependent fluorescence spectrum of sample V3-1 in this embodiment of the invention. (A) is a schematic diagram showing bright blue fluorescence under a 365 nm ultraviolet lamp, and (B) is a schematic diagram showing the redshift phenomenon of the maximum emission wavelength of the sample as the excitation wavelength increases. Figure 9 The images show the SEM analysis and high-resolution SEM analysis of sample V3-1 in this embodiment of the invention, where (A) is the SEM particle size distribution map of the sample and (B) is the high-resolution SEM lattice stripe map of the sample. Figure 10 This is a schematic diagram of the carbon quantum dot preparation method according to an embodiment of the present invention. Detailed Implementation

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

[0023] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0024] Example 1 This invention provides a data-driven closed-loop iterative method for preparing carbon quantum dots with high quantum yield and high luminescence intensity, comprising: An initial dataset was constructed based on the synthesis parameters of carbon quantum dots and their corresponding luminescence performance indicators. Construct a multi-objective prediction model and train it using the initial dataset; Based on the global optimization algorithm, the trained multi-objective prediction model is iteratively optimized to obtain carbon quantum dot synthesis parameters that meet the preset performance conditions; Based on the carbon quantum dot synthesis parameters that meet the preset performance conditions, carbon quantum dots that meet the preset performance conditions are prepared.

[0025] like Figure 10 As shown, the specific implementation process of the present invention is as follows: Based on the synthesis parameters of carbon quantum dots and their corresponding luminescence performance indicators, an initial dataset is constructed, specifically including: (1) Select sample raw materials for preparing carbon quantum dots. The sample raw materials include carbon source, nitrogen source, dopant source, pH adjuster and metal ion dopant element. Specifically: citric acid is selected as carbon source, urea as nitrogen source, cysteine ​​as dopant source, phosphoric acid / ammonia water as pH adjuster, and Zn as the dopant element. 2+ and Fe 3+ As a metal doping element.

[0026] (2) Several initial synthesis experiments were designed based on different synthesis conditions. Combined with the sample raw materials, the initial carbon quantum dots were synthesized through hydrothermal reaction using a preset temperature control program. Specifically: Design 61 initial synthesis experiments under different synthesis conditions, systematically varying the mass ratio of each raw material, the volume ratio of solvent, the volume ratio of ammonia, the pH value (pH=1.7-10.8), and the mass ratio of metal ions (0-0.63%). Initial carbon quantum dots were synthesized via hydrothermal reaction using a one- to three-stage temperature control program (T1=160-235℃, t1=26 h; T2=160-230℃, t2=14 h; T3=0-200℃, t3=0-4 h).

[0027] The initial synthesis experiment design follows these principles: 1) The mass percentage range of carbon source / nitrogen source is 10.04-53.57% and 17.57-82.99%, respectively; 2) The mass percentage of dopant source cysteine ​​used is 0-14.88%; 3) The pH adjuster, phosphoric acid, is used at a mass percentage of 0-48.10%, and the ammonia solution is used at a volume percentage of 0-6.98%. 4) Metal ions Zn 2+The dosage is 0-0.63% by mass, Fe 3+ The dosage is 0-0.47% by mass.

[0028] (3) Measure the quantum yield (QY) and maximum photoluminescence intensity (MPI) of each initially synthesized sample (initial carbon quantum dots) and establish an initial dataset containing 15 input features (carbon quantum dot synthesis parameters: raw material parameters, temperature program) and 2 output targets (corresponding luminescence performance indicators: QY, MPI).

[0029] Specifically, the parameters for carbon quantum dot synthesis include: the mass percentages of citric acid, urea, and cysteine; the volume percentages of phosphoric acid and ammonia; the volume percentage of solvent; and the Zn content. 2+ and Fe 3+ The percentage of doping, and the temperature control program consisting of one to three stages, each stage including heating temperature and holding time.

[0030] Furthermore, a multi-objective prediction model is constructed and trained using the initial dataset.

[0031] The multi-objective prediction model used in this invention is the CEM-MultiMax Net neural network model, which is a multi-objective optimization neural network model combining transentropy method (CEM) and regression analysis, specifically designed to solve multi-objective problems in experimental optimization tasks. This model aims to simultaneously optimize multiple objective functions, such as quantum yield and photoluminescence intensity, through an iterative optimization process. The core objective of CEM-MultiMax Net is to approach the optimal solution in multi-objective optimization problems by continuously adjusting the optimization strategy and incorporating experimental feedback. Specifically, the optimization process of CEM-MultiMax Net can be divided into the following main steps: (1) Initial Training and Prediction. The model is first trained on the existing dataset, and the trained model is used to predict a set of experimental synthetic parameters. These synthetic parameters are then used in the actual experiments to generate feedback data for the objective function.

[0032] (2) Data feedback and retraining. Experimental results were added to the training dataset to expand the existing data and reflect the actual experimental results. The model was retrained based on the newly added data to further optimize the prediction accuracy of the objective function.

[0033] (3) Iterative optimization. In each iteration, CEM-MultiMax Net adjusts its optimization strategy based on the current experimental data and target prediction results. In each iteration, the model is retrained by introducing new data, gradually improving the prediction accuracy of each target.

[0034] (4) CEM optimization. CEM plays a crucial role in this model. It optimizes the performance of the objective function step by step by sampling the input space, selecting the recipe that meets the expected goal, and filtering and updating.

[0035] Through this iterative feedback mechanism, CEM-MultiMax Net can effectively solve multi-objective optimization problems, especially in fields such as chemical synthesis parameter optimization, materials research and development, and drug discovery, promoting the synergistic optimization of multiple objectives. The model's multi-objective nature enables it to handle complex multi-dimensional optimization tasks, improving each objective while gradually approaching the global optimum.

[0036] First, the boundaries and constraints of the input features are determined based on historical data. To avoid local optima in the synthesis parameters, the sampling distribution is initialized to a relatively wide range to ensure diversity in the exploration space. Simultaneously, to expand the recipe search boundary, an upper bound expansion ratio is set, allowing sampling points to exceed historical maximum values, thereby discovering potential improvement schemes.

[0037] In each iteration, a new batch of synthetic parameters is generated using a sampling algorithm. These synthetic parameters are sampled using a normal distribution, ensuring that each synthetic parameter satisfies the constraint of the historical maximum value for each feature. If the sampling results do not meet the feasibility requirements, the optimization algorithm adjusts the standard deviation of the sampling distribution and resamples to expand the solution space.

[0038] The synthesized parameters sampled in each batch are input into the pre-trained CEM-MultiMax Net model for prediction. The model outputs the target value for each synthesized parameter and evaluates it according to a preset optimization direction (e.g., "higher is better"). The prediction results of the objective function are normalized and weighted to balance the optimization needs among multiple objectives.

[0039] In each iteration, the best-performing synthetic parameters, called "elite synthetic parameters," are selected based on the target score. These elite synthetic parameters influence the mean and standard deviation of the sampling distribution, thus guiding the sampling process in the next round. To avoid getting trapped in local optima, a momentum factor is introduced when updating the distribution to improve search efficiency and expand the search space.

[0040] At the end of each optimization round, the target score of the current synthesis parameters is compared with the historical best result. If the current target score is better than the historical best synthesis parameters, the optimal synthesis parameters are updated and the current formulation is recorded as the elite synthesis parameters (elite formulation). The optimization process will continue until a preset stopping criterion is met, such as the maximum number of iterations or the convergence of the objective function.

[0041] After the optimization process is complete, all synthesis parameters that meet the optimization objective will be saved. The optimal synthesis parameters, elite synthesis parameters, and other synthesis parameters that meet the conditions will be recorded and output to provide basic data support for further experiments and applications.

[0042] By continuously evaluating and updating the objective function, CEM-MultiMax Net can gradually approach the global optimum and achieve optimization of multiple objectives.

[0043] The CEM-MultiMax Net model architecture includes: (1) Input layer: Receives 15-dimensional feature vectors (raw material mass ratio, liquid raw material and solvent addition volume ratio, pH, ion concentration, temperature parameters, etc.).

[0044] (2) Shared Encoder: Two linear layers (hidden layer dimensions 128 and 64), each followed by ReLU activation, batch normalization, and dropout (p=0.3). The shared encoder is a key component of the model, and its core function is to extract general feature representations that are valuable for multiple tasks from the input features. This design allows the model to capture common patterns in the input data, thereby improving overall performance.

[0045] (3) Two branch output layers: For the QY branch, a linear layer + Sigmoid activation is used to output the predicted value of quantum yield, with an output range of [0, 1]; For the MPI branch, a linear layer + Identity activation is used to output the predicted value of maximum photoluminescence intensity.

[0046] (4) Training hyperparameters: learning rate 0.001, batch size 32, maximum number of epochs 200, early stopping patience value 20 including two branch output layers.

[0047] The process of training a multi-objective prediction model using the initial dataset includes: (1) Preprocessing the initial dataset: removing outliers (bias > 5%), standardization (Z-score normalization), and data augmentation (Gaussian noise amplification by 10 times), specifically including: Outlier detection: The 3σ principle is used to remove samples that deviate from the mean by more than 3 times the standard deviation. Standardization processing: Input feature x is normalized using Z-score; QY target maintains the [0, 1] scale; MPI target is first transformed using log1p and then normalized using Z-score; Data augmentation: Gaussian noise with a mean of 0 and a standard deviation of 5% of the original value was added to each original sample to generate 610 augmented samples.

[0048] (2) Divide the preprocessed initial dataset into training set and test set in an 8:2 ratio.

[0049] (3) Construct the CEM-MultiMax Net multi-objective neural network model, adopt the shared encoder structure, and set the QY output layer (Sigmoid activation) and MPI output layer (Identity activation).

[0050] (4) The loss function is a weighted combination of MSE Loss (QY) and Smooth L1 Loss (MPI), and the optimizer uses the Adam algorithm combined with ReduceLROnPlateau learning rate scheduling.

[0051] (5) Train the model until convergence and verify on the test set: QY predicts R 2 =0.9246 (MAE=0.0168), MPI prediction R 2 =0.9840 (MAE=7921.1).

[0052] Furthermore, based on a global optimization algorithm, the trained multi-objective prediction model is iteratively optimized to obtain carbon quantum dot synthesis parameters that meet preset performance conditions, specifically including: Candidate synthesis parameters are generated by using the trained multi-objective prediction model and combining it with a global optimization algorithm. Experiments were conducted to verify the candidate synthesis parameters and obtain real luminescence performance data. The actual luminescence performance data is added to the initial dataset to obtain the supplementary dataset; The multi-objective prediction model was retrained using a supplementary dataset and iteratively optimized until carbon quantum dot synthesis parameters that meet the preset performance conditions were obtained.

[0053] Specifically, The optimization process includes: (1) Initialization: Generate N=100 candidate synthesis parameters by randomly sampling from the training set distribution; (2) Evaluation: Predict the QY and MPI of each candidate synthesis parameter using the trained model; (3) Selection: Select the 10 elite recipes with the highest overall score (QY×MPI) K; (4) Update: Update the sampling distribution based on the mean and covariance matrix of the elite formulation; (5) Iteration: Repeat steps (2)-(4) until convergence or the maximum number of iterations (T=50) is reached; (6) Experimental verification: The final elite formula was synthesized in actual experiments to determine the actual QY and MPI.

[0054] The iterative process includes: (1) First iteration: Candidate synthesis parameters were generated using the transentropy method (CEM), and the top 10 elite formulations were screened for experimental verification. The QY range was 23.08-24.18%, and the optimal sample was V1-3 (QY=23.73%). This method generates candidate synthetic parameters with optimal overall performance in the parameter space by sampling, evaluating, selecting and updating the distribution.

[0055] (2) Second iteration: The experimental data from the first round were added to the training set to retrain the model, generate new formulas and screen the first 6 groups for verification. The QY range was 21.55-45.77%, and the best sample was V2-3 (QY=45.77%, which is 32.78% higher than the initial value). (3) Third iteration: Integrate the data from the first two rounds for retraining, screen the top 4 groups of formulas for verification, QY range 40.27-80.65%, the best sample V3-1 (QY=80.65%, MPI=455412 au, which are 133.97% and 14.85% higher than the initial values, respectively). The multi-round iteration mechanism satisfies the following conditions: (1) Iteration termination condition: The iteration terminates when the QY of the best sample in the new round is less than 5% and the MPI is less than 10%, or when QY is greater than 80% and MPI is greater than 400,000 au. (2) Data feedback: The experimental data from each round is added to the training set, and the model is retrained to update its predictive ability; (3) Confidence assessment: The relative error between the model prediction and the experimental value is less than 15% and is considered a valid prediction.

[0056] The process of generating candidate synthesis parameters specifically includes: randomly sampling from the training set distribution to generate initial candidate synthesis parameters; using the trained model to predict the luminescence performance of each candidate synthesis parameter; selecting several elite formulations with the highest comprehensive scores; updating the sampling strategy based on the distribution of elite formulations, and repeating the iteration until convergence.

[0057] The methods for obtaining real performance data (experimental verification) include: preparing the reaction system according to the candidate synthesis parameters; carrying out the hydrothermal reaction under the set multi-stage temperature program; purifying the reaction products; and measuring the quantum yield and maximum photoluminescence intensity of the obtained carbon quantum dots to obtain real performance data.

[0058] The optimal carbon quantum dots V3-1 prepared have the following performance characteristics: (1) Quantum yield: QY=80.65%, which exceeds all environmentally friendly N-doped carbon quantum dots reported in the literature (typical QY<30%). (2) Maximum photoluminescence intensity: MPI=455412 au, which is 4.5 times that of commercial N-doped carbon quantum dots (approximately 100000 a.u.); (3) Optimal excitation wavelength: λex = 365 nm; (4) Optimal emission wavelength: λem = 439 nm; (5) Excitation dependence: The emission peak exhibits a redshift as the excitation wavelength increases; (6) Appearance characteristics: It appears pale yellow under natural light and exhibits bright blue fluorescence under 365 nm ultraviolet light.

[0059] Furthermore, the SHAP value analysis method was used to interpret the model decisions and identify key parameters affecting QY and MPI: the first-stage temperature (T1) and the second-stage temperature (T2) both made significant positive contributions to both objectives, with SHAP values ​​distributed within the ranges of ±0.1 (QY) and ±0.8 (MPI). Specifically, these included: (1) Calculate the marginal contribution (SHAP value) of each feature to the predicted output. (2) Draw a beeswarm plot to show the importance ranking of features. The top 3 features with the greatest impact on QY are: first stage temperature T1, second stage temperature T2, and urea mass; the top 3 features with the greatest impact on MPI are: first stage temperature T1, second stage temperature T2, and pH value. (3) Characteristic dependence analysis: High temperature (T1>230℃, T2>200℃) has a positive promoting effect on both QY and MPI.

[0060] Furthermore, the final preparation process for high-performance carbon quantum dots was determined to be as follows: citric acid 16.53% by mass, urea 82.64% by mass, cysteine ​​0%, phosphoric acid 0%, solvent volume percentage 94.34%, ammonia volume percentage 5.66%, pH=9.8, Zn 2+ 0.83% by mass, Fe-free 3+ Three-stage temperature program (235℃ / 6 h → 230℃ / 4 h → room temperature cooling).

[0061] This invention is applicable to, but not limited to, the following carbon quantum dot preparation systems: (1) Carbon source: citric acid, glucose, ascorbic acid, ethylenediaminetetraacetic acid, dopamine; (2) Nitrogen source: urea, ethylenediamine, triethanolamine, polyethyleneimine; (3) Dopant source: cysteine, glutathione, thiourea, boric acid; (4) Metal ion: Zn 2+ Fe 3+ Cu 2+ Mn 2+ Co 2+ (5) Synthesis methods: hydrothermal method, solvothermal method, microwave-assisted method.

[0062] Furthermore, the prepared high-fluorescence carbon quantum dots include: (1) quantum yield ≥80%; (2) maximum photoluminescence intensity ≥400000 au; (3) average particle size 2-5 nm; (4) surface containing abundant functional groups such as -OH, -COOH, and -NH2; and (5) good water solubility, photostability, and biocompatibility.

[0063] Applications of highly fluorescent carbon quantum dots in the following fields: (1) Bioimaging: cell / tissue fluorescent labeling, in vivo imaging; (2) Photocatalysis: photodegradation of organic pollutants, photocatalytic hydrogen production; (3) Sensing and detection: metal ion detection (Fe2+). 3+ Cu 2+ (4) Photoelectric devices: light-emitting diodes (LEDs), solar cells; (5) Drug delivery: carriers for photothermal therapy and photodynamic therapy.

[0064] In summary, this invention discloses a data-driven closed-loop iterative method for preparing carbon quantum dots with high quantum yield and high luminescence intensity, comprising: systematically controlling the feeding ratio of carbon / nitrogen sources such as citric acid, urea, and cysteine, pH value, and metal ion doping (Zn). 2+ Fe 3+ An initial synthetic dataset was constructed using a multi-stage temperature control program. A multi-objective neural network model was employed, with quantum yield (QY) and maximum photoluminescence intensity (MPI) as optimization targets, combined with the transentropy method for iterative optimization. After three rounds of iteration, high-performance carbon quantum dots were successfully prepared, achieving improvements of 133.97% and 14.85% respectively compared to the initial optimal sample. This invention achieves intelligent optimization of the carbon quantum dot preparation process through a closed-loop evolution mechanism of data feedback and experimental verification.

[0065] Example 2 Based on the same inventive concept, this embodiment provides a machine learning-driven optimization method for the preparation of high-luminescence carbon quantum dots, including: By constructing a closed-loop iterative framework that combines machine learning models with experimental verification, the synergistic optimization of two key luminescence performance indicators of carbon quantum dots—quantum yield (QY) and maximum photoluminescence intensity (MPI)—is achieved. Specifically, the following steps are included: S1. Obtaining the Initial Dataset. First, a series of initial synthesis experiments are conducted, systematically adjusting the synthesis parameters to prepare multiple sets of carbon quantum dot samples. The synthesis parameters include at least the raw material composition (e.g., the type and proportion of carbon source, nitrogen source, and dopant source), pH value, and a multi-stage temperature control program (e.g., heating temperature and holding time at each stage). Then, the prepared samples are characterized by measuring their corresponding luminescence performance indicators, primarily quantum yield (QY) and maximum photoluminescence intensity (MPI). Each set of synthesis parameters and its corresponding luminescence performance indicator is treated as a data point, and these are compiled to form the initial dataset.

[0066] S2. Train a multi-objective machine learning model. Based on the initial dataset obtained in step S1, construct and train a machine learning model capable of handling multi-objective regression tasks. This model aims to learn and establish a complex, non-linear mapping relationship between synthetic parameters (input) and luminescence performance metrics (output). After training, an initial prediction model is obtained.

[0067] S3. Generate candidate synthesis parameters. Using the initial prediction model obtained in step S2, and combining it with a global optimization algorithm (such as cross-entropy CEM), optimization is performed throughout the entire synthesis parameter space. The goal of this process is to find one or more sets of candidate synthesis parameters that optimize the combined performance of QY and MPI predicted by the model. These candidate parameters represent the "new recipe" that the model believes is most likely to achieve a performance breakthrough.

[0068] S4. Experimental Verification and Data Expansion. Based on the candidate synthesis parameters generated in step S3, conduct actual experimental verification, synthesize new carbon quantum dot samples, and measure their actual QY and MPI values. Use this set of new synthesis parameters and their corresponding actual performance indicators as one or more new data points.

[0069] S5. Model Iteration and Retraining. The new data points obtained in step S4 are added to the existing dataset, forming an updated dataset with richer information and a wider scope. Using this updated dataset, the multi-objective machine learning model is retrained, thereby updating the model's weights and making its understanding of parameter-performance relationships more accurate and comprehensive. This step achieves self-learning and iterative optimization of the model.

[0070] S6. Iterative Optimization. Repeat the "prediction-experiment-retraining" cycle from S3 to S5. In each iteration, the model makes predictions based on a more complete dataset, guiding the experiment towards a more promising parameter region. This iterative process continues until the luminescence performance of the newly synthesized carbon quantum dots meets the preset performance requirements (e.g., QY reaches a certain threshold), or the performance no longer shows significant improvement after multiple iterations, indicating that it has approached or reached the stable optimal value for this system. At this point, the iteration is terminated.

[0071] Furthermore, the synthesis parameters include: citric acid as a carbon source, urea as a nitrogen source, cysteine ​​as a sulfur dopant source, phosphoric acid as a pH adjuster, and the metal dopant element Zn. 2+ and Fe 3+ The percentage of the total mass of the reaction system, the percentage of ammonia and solvent in the total volume of the liquid phase, and a multi-stage temperature control program consisting of heating temperature and holding time, comprising one to three stages.

[0072] Furthermore, the multi-objective machine learning model is a multi-objective optimized neural network model that combines regression analysis, such as a model named CEM-MultiMax Net. This model employs a shared encoder-multiple-heads architecture. The shared encoder consists of several linear layers, activation functions (such as ReLU), batch normalization layers, and dropout layers, used to extract deep features from the input synthesis parameters. Two independent output heads are connected after the encoder: one for predicting QY, with its terminal using a sigmoid activation function to ensure the output value is in the range [0, 1]; the other for predicting MPI, with its terminal using an identity activation function (i.e., no activation). The model's loss function consists of two parts: mean squared error loss (MSE Loss) for QY and smooth L1 loss for MPI.

[0073] Furthermore, before training the model, the dataset is preprocessed, including: 1) Data augmentation: To prevent the model from overfitting on small datasets, Gaussian noise or other methods can be used to augment the original data. 2) Data standardization: The input synthetic parameter features are processed using Z-score standardization; for the target value MPI, due to its wide numerical range, a log1p transformation (log(1+x)) is first performed to compress its dynamic range, and then Z-score standardization is performed.

[0074] Furthermore, after completing the iterative optimization, the complete dataset containing the initial data and new data from all iteration rounds is input into the finally trained model. Interpretable machine learning methods such as SHAP (SHapley Additive exPlanations) are used to analyze and quantify the contribution and influence patterns of each synthesis parameter on QY and MPI, thereby providing data-driven insights for understanding the intrinsic mechanism of high-luminescence carbon quantum dot formation.

[0075] The present invention also provides a highly luminescent carbon quantum dot prepared by the above method, which exhibits a quantum yield (over 80%) far exceeding that achievable by conventional methods and an extremely high maximum photoluminescence intensity (over 400,000 au) at a specific excitation wavelength (365 nm).

[0076] Example 3 Based on the same inventive concept, this embodiment provides a machine learning-driven optimization method for the preparation of high-luminescence carbon quantum dots, including: This embodiment details the construction process of the initial dataset, including raw material preparation, experimental design, synthesis steps, purification procedures, and performance measurement methods. Through the systematic design of 61 initial synthesis experiments (Tables 1 and 2), covering different raw material ratios, pH values, metal ion doping concentrations, and multi-stage temperature control procedures, diverse and high-quality data were provided for subsequent machine learning model training.

[0077] (1) In this embodiment, citric acid (≥99.5%) was used as the main carbon source, urea (≥99.0%) was used as the nitrogen source, and the purity was analytical grade. L-cysteine ​​(≥98%) was used as the sulfur dopant source, phosphoric acid was used as the acidic pH adjuster, ammonia was used as the alkaline pH adjuster, and zinc sulfate was provided as ZnSO4·7H2O. 2+ Ferric chloride provides Fe ions in the form of FeCl3·6H2O. 3+ All experimental water was deionized.

[0078] (2) The experimental design followed the principle of system optimization, and a total of 61 initial synthesis experiments were designed, with the following key parameters varied systematically: citric acid mass range of 1.2-3.5 g, urea mass range of 1.5-6 g, phosphoric acid mass range of 0-4.9 g, solvent volume range of 20-60 mL, ammonia volume range of 0-3 mL, pH value range covering a wide range of 1.7-10.8, and Zn... 2+ The concentration range was 0-60 mg, and the Fe3+ concentration range was 0-45 mg. The temperature program adopted a one- to three-stage design. The first stage temperature T1 ranged from 160-235℃, and the reaction time t1 ranged from 2-6 h; the second stage temperature T2 ranged from 160-230℃, and the reaction time t2 ranged from 1-4 h; the third stage temperature T3 ranged from 0-200℃, and the reaction time t3 ranged from 0-4 h.

[0079] (3) Taking sample 58, which showed the best performance in the initial dataset, as an example, its preparation and characterization steps are as follows. First, 28.57% citric acid and 71.43% urea were dissolved in 50 mL of deionized water, and the initial pH of the solution was approximately 5.0. Then, the solution was transferred to a 100 mL high-pressure reactor for hydrothermal reaction. The reaction program was set as follows: heating to 230℃ and holding for 3 h; then cooling to 200℃ and holding for 3 h; finally cooling to 160℃ and holding for 2 h. After the reaction, the reactor was allowed to cool naturally to room temperature. The resulting reaction product underwent a series of purification treatments. The supernatant was collected by centrifugation at 4000 rpm for 10 min, filtered through a 0.22 μm filter membrane, and dialyzed with a dialysis bag with a molecular weight cutoff of 1000 Da for 72 h (during which the deionized water was replaced every 8 h). Finally, the dialyzed solution was lyophilized to obtain pale yellow powdered carbon quantum dots.

[0080] (4) Performance characterization of the obtained carbon quantum dots. The quantum yield (QY) of the samples was determined by absolute method using the Quantaurus-QY Plus quantum yield and transient fluorescence testing system. At the optimal excitation wavelength of 365 nm, the quantum yield of sample 58 was calculated to be 34.47%. Under the same excitation conditions, its maximum photoluminescence intensity (MPI) was measured to be 347450 au, and the corresponding emission peak was located at 439 nm (Table 3).

[0081] (5) To construct the initial dataset for machine learning, another 60 sets of samples were prepared using a similar method, with changes to raw material ratios and temperature programs. The performance range of these 61 sets of samples was: quantum yield 1.00-34.47%, maximum photoluminescence intensity 14520-396525 au (Table 3), providing a diverse data foundation for model training. Preliminary analysis revealed that the highest quantum yield (sample 58) and the highest maximum photoluminescence intensity (sample 61) were both obtained under conditions of low citric acid content, no metal doping, and a specific three-stage high-temperature program. This reveals a favorable direction for the synthesis of high-performance carbon quantum dots, and their precise optimal combination will be further explored using the machine learning method of this invention.

[0082] Table 1 Table 2 Table 3 Secondly, this embodiment also provides the construction and training of the CEM-MultiMax Net model. This embodiment describes the construction and training process of the CEM-MultiMax Net multi-objective neural network model. Figure 1 , Figure 2 and Figure 3 The goal is to achieve accurate predictions of the quantum yield (QY) and maximum photoluminescence intensity (MPI) of carbon quantum dots, specifically including: (1) Data preprocessing and dataset construction; First, the 61 initial experimental data sets obtained in Example 1 were preprocessed. The data were found to be complete and without missing values, and the QY and MPI values ​​were within reasonable ranges, requiring no data cleaning. Z-score normalization was applied to the 15 input features (such as raw material quality and reaction temperature). No transformation was performed on the output target QY (value range 0-1); for the output target MPI, a log(1+MPI) transformation was first performed to compress the numerical range, followed by Z-score normalization. To expand the dataset and improve the model's generalization ability, Gaussian noise with a mean of 0 and a standard deviation of 5% of the feature value was added to the input features of each original sample, expanding each original sample into 10 enhanced samples, resulting in a total of 610 samples. The 610 enhanced samples were randomly divided into a training set (488 samples) and a test set (122 samples) in an 8:2 ratio, using a fixed random seed to ensure reproducibility.

[0083] (3) Model training configuration Loss function: Weighted combined loss Ltotal = α·MSE(QY) + β·SmoothL1(MPI). Where QY is the loss calculated using mean squared error (MSE), and MPI is the loss calculated using smooth L1 loss. The weighting coefficients are set to α = 0.8 and β = 0.2. Figure 3 ).

[0084] Optimizer: Adam optimizer was selected, with a learning rate (lr) of 0.001, β1 = 0.9, and β2 = 0.999. Figure 3 ).

[0085] Learning rate scheduling: The ReduceLROnPlateau strategy is adopted. When the validation set loss does not decrease for 10 consecutive periods (patience=10), the learning rate is multiplied by 0.5.

[0086] Early stopping mechanism: To prevent overfitting, training is terminated early when the validation set loss does not improve for 20 consecutive epochs (patience=20), and the best-performing model is saved.

[0087] (4) Training results and performance evaluation: The model training triggered an early stopping mechanism at the 156th epoch, indicating that it had converged to the optimal state. The performance of the saved best model was evaluated using a reserved test set, and the results are shown in Table 4. Quantum Yield (QY) Prediction Performance: Coefficient of Determination (R²) 2 =0.9246, mean absolute error (MAE) =0.0168 (i.e. 1.68%), root mean square error (RMSE) =0.0231 (i.e. 2.31%). Maximum photoluminescence intensity (MPI) prediction performance: coefficient of determination (R) 2 =0.9840, mean absolute error (MAE) =7921.1 au (relative error approximately 3.96%), root mean square error (RMSE) =11249.7 au (relative error approximately 5.62%).

[0088] Table 4 Scatter plot of predicted and actual values ​​(e.g.) Figure 4 A and Figure 4 As shown in Figure B, the data points are closely distributed near the diagonal, further confirming the model's high accuracy and excellent fitting performance. Furthermore, after a period of training, both the training loss and the value loss tend to stabilize, validating that the loss stabilizes after 50 epochs, indicating that the model is neither overfitting nor underfitting and is well-trained. Figure 5 A and Figure 5 B). By Figure 6 It can be seen that for the numerous parameters affecting QY, the SHAP values ​​of each factor are distributed around 0, indicating that the same parameter may have a positive or negative impact on QY at different values, reflecting the complexity of the synthesis process. Furthermore, the SHAP values ​​are mainly distributed between -0.05 and 0.1, indicating that the contribution of a single parameter is limited, but the cumulative effect of multiple parameters can significantly affect QY. The point distribution is relatively uniform with few outliers, indicating good dataset quality and reliable model interpretation. Notably, higher first-stage and second-stage temperatures positively increase quantum yield, while lower temperatures have a negative impact, indicating that temperature may promote carbonization or doping reactions and is the primary control variable for synthesis optimization. In actual preparation, the first and second-stage temperatures should be preferentially adjusted to higher levels to maximize QY. Figure 7It can be seen that for the numerous parameters affecting MPI intensity, the SHAP values ​​of each factor are relatively evenly distributed around 0. Many features show that both high values ​​(red dots) and low values ​​(blue dots) can produce positive / negative contributions (MPI intensity is affected by exciton recombination, surface traps, etc.). Furthermore, the SHAP values ​​are mainly distributed within ±0.8, indicating that MPI intensity depends more on a single dominant factor than on a combination of parameters. The relatively scattered point distribution suggests more nonlinear or interactive effects. Notably, higher first-stage and second-stage temperatures generally positively increase MPI intensity, indicating that temperature may promote uniform carbonization and the formation of more luminescent centers. Simultaneously, similar to the QY trend, it indicates that first-stage and second-stage temperatures are key to both objectives, but their impact on luminescence intensity is more "sensitive."

[0089] Then came the first round of iterative optimization; This embodiment details the first round of iterative optimization process based on CEM. Figure 3 Through a systematic process of candidate synthesis parameter generation, elite formulation screening, experimental verification, and result analysis, combined with model prediction and experimental feedback, feasible optimization directions were explored, laying an important foundation for subsequent iterations.

[0090] CEM, as a highly efficient global optimization algorithm, is particularly suitable for optimization problems in high-dimensional parameter spaces. Its core mechanism lies in iteratively updating the parameter distribution, gradually concentrating the sampling towards the optimal solution region. The specific implementation process includes four key steps: first, establishing the parameter distribution based on initial samples to generate candidate synthetic parameters; then, evaluating the candidate synthetic parameters using a trained prediction model; next, selecting elite formulations based on comprehensive scores; and finally, updating the parameter distribution based on the elite formulations to complete the iterative cycle.

[0091] In the experimental verification phase, the complete synthesis and testing process was demonstrated using Elite Formulation V1-3 as an example (Tables 5 and 6). This formulation, through precise raw material preparation, strict hydrothermal reaction control, and standardized purification treatment, ultimately yielded carbon quantum dot samples.

[0092] The first iteration yielded 10 sets of valid experimental data, with the optimal formulation V1-3 receiving the highest overall score. Performance test results showed a quantum yield of 23.73% and a maximum photoluminescence intensity of 284317 au, highly consistent with the model predictions, validating the reliability of the prediction model (Table 7). However, compared to the initial optimal sample, its quantum yield decreased. This result indicates that the first round of optimization failed to achieve the expected goals. Despite the underperformance, this iteration still holds significant value. The newly obtained experimental data significantly expanded the parameter space coverage, particularly revealing the negative impacts of parameters such as phosphoric acid dosage and metal ion concentration on performance. These findings provide crucial guidance for subsequent iterations, necessitating improvements in model training strategies, optimization of the comprehensive evaluation function, adjustment of parameter sampling range, and strengthened synergistic optimization between experimental verification and theoretical prediction. The lessons learned from this iteration provide ample basis for adjusting subsequent optimization directions, demonstrating the practical value of the transentropy method in achieving a balance between "exploration and utilization" in materials synthesis optimization.

[0093] Table 5 Table 6 Table 7 Then comes model retraining and the second round of iterative optimization: This embodiment illustrates the model retraining process after incorporating the first round of iteration data into the training set, and the second round of iterative optimization based on the updated model. Through strategy adjustment and empirical feedback, the second round of iteration successfully achieved a significant improvement in the performance of carbon quantum dot synthesis, verifying the effectiveness of the iterative optimization method.

[0094] The model retraining was based on 71 sets of sample data accumulated in the early stage, including the initial 61 sets and 10 sets of experimental data from the first round. While maintaining the original CEM-MultiMax Net model architecture and hyperparameters, the expanded dataset underwent standardization preprocessing and data augmentation to generate 710 training samples. During training, a comprehensive loss function and Adam optimizer were used, combined with dynamic adjustment of the learning rate and an early stopping mechanism to ensure that the model fully learned the data features. The second round of iterations made four key adjustments in strategy: first, optimizing the initial distribution, focusing on high quantum yield samples in the early stage to improve sampling efficiency; second, improving the comprehensive scoring function, adopting a geometric weighting form to highlight the dominant weight of quantum yield; third, introducing parameter constraints to limit the amount of phosphate and metal ions to avoid adverse effects; and fourth, expanding the sampling scale, generating 200 candidate synthesis parameters and selecting 20 elite samples in each round to enhance spatial exploration capabilities. After 50 rounds of iterations, the distribution converged to the high-performance region, and finally, the 6 formulations with the highest comprehensive scores (V2-1~V2-6) were selected for experimental verification (Tables 8 and 9).

[0095] Taking V2-3 as an example, its synthesis and performance are demonstrated. This formulation uses only citric acid and urea to react in deionized water, with only ammonia water introduced to adjust the pH to 8.7. A three-stage temperature program is used for hydrothermal synthesis (Tables 8 and 9). After purification, the quantum yield reached 45.77%, and the maximum photoluminescence intensity was 322884 au (Table 10). Compared with the model predictions, the accuracy of quantum yield and photoluminescence intensity predictions reached 117.38% and 94.42%, respectively, showing that the model has excellent predictive ability in the high-performance region. The success of V2-3 reveals the key synthetic elements of high-quantum-yield carbon quantum dots: a lower amount of citric acid helps to form small-sized carbon cores, a moderate amount of urea provides optimal nitrogen doping, a simple formulation reduces impurity interference, and a weakly acidic environment and a three-stage temperature program promote carbon core formation and surface passivation. These findings provide important basis for subsequent optimization and industrial applications. The success of the second iteration verifies the effectiveness of the technical route of this invention: through the close combination of machine learning and experimental iteration, significant performance breakthroughs were achieved within a limited number of experiments. However, the QY value of the V2-3 performance parameter is 32.78% higher than the initial highest experimental group, but the MPI value is still some distance from the maximum value of the initial single experimental group, so iterative optimization is needed.

[0096] Table 8 Table 9 Table 10 Then comes the third round of iterative optimization and determination of the optimal formula: This embodiment details the third round of iterative optimization, which further explores the limits of carbon quantum dot synthesis performance through model retraining and strategy fine-tuning, ultimately determining the synthesis formulation with optimal overall performance (Tables 11 and 12). Building upon the successful experience of the second round, this iteration focuses on the fine-tuning of key parameters, achieving a further improvement in quantum yield and verifying the model's predictive stability in the high-performance region.

[0097] Model retraining was performed using an expanded dataset containing 77 original samples, covering all experimental data from the initial and first two iterations. After standardization preprocessing and data augmentation, 770 training samples were generated. The model was retrained while maintaining the CEM-MultiMax Net model architecture and training configuration. The training process showed a steady decrease in the loss function, triggering an early stopping mechanism at the 138th iteration. The third iteration made four key adjustments to the strategy: the initial distribution further focused on the top 10 samples with the highest quantum yield, strengthening the sampling orientation of the high-performance region; the comprehensive scoring function maintained a geometric weighted form, keeping the quantum yield as the dominant weight; the parameter constraints were further tightened, fixing the range of key parameters based on previous experimental statistics to narrow the search space; and the sampling scale and number of iterations were increased to fully explore the local optimum region. After iteration, the distribution converged to the high-performance region, and the four formulations with the highest comprehensive scores (V3-1 to V3-4) were finally selected for experimental verification (Tables 11 and 12).

[0098] Taking V3-1 as an example, its synthesis and performance are demonstrated. This formulation uses 16.53% citric acid, 82.64% urea (mass fraction), and no cysteine ​​or phosphoric acid added. The reaction is carried out in 86.96% deionized water at a natural pH of 9.8. 0.83% Zn²⁺ and 13.04% ammonia are introduced, and hydrothermal synthesis is completed through a three-stage temperature program (Tables 11 and 12). After standard purification, the quantum yield reached 80.65%, and the maximum photoluminescence intensity was 455412 au, making it the optimal formulation in this round (Table 13). Compared with the model predictions, the accuracy of quantum yield and photoluminescence intensity predictions reached 144.02% and 96.87%, respectively, showing excellent predictive stability of the model in the high-performance region. The experimental results of all four groups of samples in the third round are shown in Table 13. The results show that the QY range is 40.27-80.65% and the MPI intensity range is 277061-455412 au, which is a significant improvement compared to the second round. This indicates that the optimization process has strong robustness and successfully stabilized the performance at an extremely high level.

[0099] The V3-1 formulation exhibits excellent simplicity, reproducibility, and structural properties. It uses only three inexpensive raw materials, requires no complex additives, and the synthesis process is green and economical. Three batches of replicate experiments showed an average quantum yield of 80.65% and a relative standard deviation of only 0.5%, demonstrating the high reproducibility of the formulation. Optical analysis shows ( Figure 8 The optimal excitation wavelength for V3-1 is 365 nm, and the optimal emission wavelength is 439 nm. The maximum ultraviolet absorption wavelength almost coincides with its maximum excitation wavelength. The sample appears pale yellow to the naked eye, but exhibits bright blue fluorescence under a 365 nm ultraviolet lamp. Figure 8 A). Furthermore, the maximum emission wavelength of the sample exhibits a redshift as the excitation wavelength increases. Figure 8 B), consistent with the fluorescence spectral characteristics of CQD. Structural characterization indicates that ( Figure 9 V3-1 carbon quantum dots have a narrow particle size distribution (average 2.13 nm) and a distinct lattice, which are consistent with the structural characteristics of CQDs.

[0100] In summary, the V3-1 formulation exhibits excellent performance and a simple preparation process, demonstrating strong industrialization potential. Further optimization of reaction equipment and production processes holds promise for achieving high-quality, large-scale preparation of carbon quantum dots, providing a material foundation for their practical applications in fields such as bioimaging, optoelectronic devices, and sensing.

[0101] Table 11 Table 12 Table 13 Example 4 Based on the same inventive concept, the present invention also provides a data-driven closed-loop iterative high quantum yield and high luminescence intensity carbon quantum dot preparation system for implementing the method described in the foregoing embodiments. The system includes: a dataset construction module, a model construction and training module, an iterative optimization module, and a preparation module. The dataset construction module is used to construct an initial dataset based on the carbon quantum dot synthesis parameters and corresponding luminescence performance indicators; The model building and training module is used to build a multi-objective prediction model and train the multi-objective prediction model using the initial dataset. The iterative optimization module is used to iteratively optimize the trained multi-objective prediction model based on a global optimization algorithm to obtain carbon quantum dot synthesis parameters that meet preset performance conditions. The preparation module is used to prepare carbon quantum dots that meet preset performance conditions based on carbon quantum dot synthesis parameters.

[0102] Preferably, the dataset construction module includes: selection unit, synthesis unit, and construction unit; The selection unit is used to select the sample raw materials for preparing carbon quantum dots. The sample raw materials include carbon source, nitrogen source, dopant source, pH adjuster and metal ion dopant elements. The synthesis unit is used to design several sets of initial synthesis experiments based on different synthesis conditions, and synthesize initial carbon quantum dots through hydrothermal reaction using sample raw materials and a preset temperature control program. The building blocks are used to determine the quantum yield and maximum photoluminescence intensity of the initial carbon quantum dots, forming an initial dataset containing the carbon quantum dot synthesis parameters and corresponding luminescence performance indicators.

[0103] Preferably, the multi-objective prediction model is the CEM-MultiMax Net neural network model, which includes two branch output layers; The two branch output layers use the Sigmoid activation function and the Identity activation function, respectively, to output predicted values ​​of quantum yield and maximum photoluminescence intensity.

[0104] Preferably, the iterative optimization module includes: candidate units, verification units, supplementary units, and iterative units; Candidate units are used to generate candidate synthesis parameters by utilizing the trained multi-objective prediction model and combining it with a global optimization algorithm. The verification unit is used to perform experimental verification based on the candidate synthesis parameters and obtain real luminescence performance data. The supplementary unit is used to supplement the initial dataset with the actual luminescence performance data to obtain a supplementary dataset. The iterative unit is used to retrain the multi-objective prediction model using a supplementary dataset and iteratively optimize it until carbon quantum dot synthesis parameters that meet the preset performance conditions are obtained.

[0105] Example 5 The present invention also provides an electronic device, including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor executes the computing program to implement the aforementioned method.

[0106] Example 6 The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned method.

[0107] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A data-driven closed-loop iterative method for preparing carbon quantum dots with high quantum yield and high luminescence intensity, characterized in that, The method includes: An initial dataset was constructed based on the synthesis parameters of carbon quantum dots and their corresponding luminescence performance indicators. Construct a multi-objective prediction model and train it using the initial dataset; Based on the global optimization algorithm, the trained multi-objective prediction model is iteratively optimized to obtain carbon quantum dot synthesis parameters that meet the preset performance conditions; Based on the carbon quantum dot synthesis parameters that meet the preset performance conditions, carbon quantum dots that meet the preset performance conditions are prepared.

2. The method according to claim 1, characterized in that, Methods for constructing an initial dataset based on carbon quantum dot synthesis parameters and corresponding luminescence performance indicators include: The raw materials for preparing carbon quantum dots were selected, including carbon source, nitrogen source, dopant source, pH adjuster and metal ion dopant elements; Several initial synthesis experiments were designed based on different synthesis conditions. Combined with sample raw materials, initial carbon quantum dots were synthesized through hydrothermal reaction using a preset temperature control program. The quantum yield and maximum photoluminescence intensity of the initial carbon quantum dots were determined to form an initial dataset containing the synthesis parameters of the carbon quantum dots and the corresponding luminescence performance indicators.

3. The method according to claim 1, characterized in that, The multi-objective prediction model is a CEM-MultiMax Net neural network model, which includes two branch output layers; The two branch output layers use the Sigmoid activation function and the Identity activation function, respectively, to output predicted values ​​of quantum yield and maximum photoluminescence intensity.

4. The method according to claim 1, characterized in that, Methods for obtaining carbon quantum dot synthesis parameters that meet preset performance conditions by iteratively optimizing a trained multi-objective prediction model based on a global optimization algorithm include: Candidate synthesis parameters are generated by using the trained multi-objective prediction model and combining it with a global optimization algorithm. Experiments were conducted to verify the candidate synthesis parameters and obtain real luminescence performance data. The actual luminescence performance data is added to the initial dataset to obtain the supplementary dataset; The multi-objective prediction model was retrained using a supplementary dataset and iteratively optimized until carbon quantum dot synthesis parameters that meet the preset performance conditions were obtained.

5. A data-driven closed-loop iterative system for preparing carbon quantum dots with high quantum yield and high luminescence intensity, said system being used to implement the method according to any one of claims 1-4, characterized in that, The system includes: a dataset construction module, a model construction and training module, an iterative optimization module, and a preparation module; The dataset construction module is used to construct an initial dataset based on the carbon quantum dot synthesis parameters and corresponding luminescence performance indicators; The model building and training module is used to build a multi-objective prediction model and train the multi-objective prediction model using the initial dataset. The iterative optimization module is used to iteratively optimize the trained multi-objective prediction model based on a global optimization algorithm to obtain carbon quantum dot synthesis parameters that meet preset performance conditions. The preparation module is used to prepare carbon quantum dots that meet preset performance conditions based on carbon quantum dot synthesis parameters.

6. The system according to claim 5, characterized in that, The dataset construction module includes: selection unit, synthesis unit, and construction unit; The selection unit is used to select the sample raw materials for preparing carbon quantum dots. The sample raw materials include carbon source, nitrogen source, dopant source, pH adjuster and metal ion dopant elements. The synthesis unit is used to design several sets of initial synthesis experiments based on different synthesis conditions, and synthesize initial carbon quantum dots through hydrothermal reaction using sample raw materials and a preset temperature control program. The building blocks are used to determine the quantum yield and maximum photoluminescence intensity of the initial carbon quantum dots, forming an initial dataset containing carbon quantum dot synthesis parameters and corresponding luminescence performance indicators.

7. The system according to claim 5, characterized in that, The multi-objective prediction model is a CEM-MultiMax Net neural network model, which includes two branch output layers; The two branch output layers use the Sigmoid activation function and the Identity activation function, respectively, to output predicted values ​​of quantum yield and maximum photoluminescence intensity.

8. The system according to claim 5, characterized in that, The iterative optimization module includes: candidate units, verification units, supplementary units, and iterative units; Candidate units are used to generate candidate synthesis parameters by utilizing the trained multi-objective prediction model and combining it with a global optimization algorithm. The verification unit is used to perform experimental verification based on the candidate synthesis parameters and obtain real luminescence performance data. The supplementary unit is used to supplement the initial dataset with the actual luminescence performance data to obtain a supplementary dataset. The iterative unit is used to retrain the multi-objective prediction model using a supplementary dataset and iteratively optimize it until carbon quantum dot synthesis parameters that meet the preset performance conditions are obtained.

9. An electronic device comprising a memory, a processor, and a computing program stored in the memory and executable on the processor, characterized in that, When the processor executes the computing program, it implements the method described in any one of claims 1-4.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1-4.

Images