A method for optimizing metal catalyst depolymerization of lignin conditions based on machine learning
By using machine learning and Bayesian optimization methods, an initial yield prediction model was constructed and an intermediate product repolymerization inhibition effect was introduced. This solved the problem of balancing activity and selectivity in the depolymerization of lignin catalyzed by metal catalysts, and achieved high-precision prediction and optimization of bio-oil monomer yield, thereby improving R&D efficiency and industrial application potential.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GUANGXI UNIV
- Filing Date
- 2026-05-08
- Publication Date
- 2026-06-23
AI Technical Summary
Existing metal catalyst-catalyzed lignin depolymerization technology suffers from problems such as difficulty in achieving both activity and selectivity, poor stability, and high cost. Furthermore, traditional research methods are insufficient to fully elucidate the reaction mechanism, resulting in low research and development efficiency and an inability to effectively guide catalyst design and reaction condition optimization.
A machine learning-based approach was used to construct an initial yield prediction model and introduce the intermediate product repolymerization inhibition effect. Combined with Bayesian optimization, the dominant factors were identified through importance analysis, and the optimal catalyst preparation and reaction conditions were iteratively searched. A feedback optimization mechanism was constructed to achieve accurate prediction and optimization.
It achieves high-precision prediction of bio-oil monomer yield, identifies key influencing factors, replaces traditional trial-and-error methods, significantly improves R&D efficiency, provides reliable data-driven optimization strategies, and is suitable for industrial applications.
Smart Images
Figure CN122266554A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of machine learning and catalytic materials technology, specifically to a machine learning-based method for optimizing the conditions for depolymerizing lignin in metal catalysts. Background Technology
[0002] Driven by the global goal of carbon neutrality, the efficient utilization of biomass resources has become a research hotspot in the energy and chemical industries. Lignin, as one of the core components of lignocellulose, is the most abundant renewable aromatic polymer in nature, accounting for 15%-30% of lignocellulose content. Through catalytic depolymerization, it can be transformed into high-value-added aromatic chemicals or biofuels, which can effectively alleviate energy pressure and reduce greenhouse gas emissions, aligning with the concept of green and low-carbon development.
[0003] Metal catalyst-catalyzed lignin depolymerization is one of the mainstream depolymerization technologies. The application of catalysts such as transition metals and noble metals can significantly reduce the oxygen content of bio-oil obtained from lignin depolymerization and reduce the generation of solid residues. However, existing research still faces many bottlenecks: on the one hand, the performance of metal catalysts is difficult to balance activity and selectivity, and they also suffer from poor stability and high cost, which restricts their industrial application; on the other hand, lignin has a complex structure, and the depolymerization reaction involves many complex chemical reactions and intermediate products. Traditional research methods are unable to fully and accurately elucidate the reaction mechanism, and the diversity of experimental conditions leads to fragmented and unsystematic research data, which cannot effectively guide the design of new catalysts and the optimization of reaction conditions.
[0004] Currently, the screening of metal catalysts and the optimization of lignin depolymerization reaction conditions still rely on traditional trial-and-error methods, requiring numerous repetitive experiments. This not only results in a significant waste of chemical reagents, energy, and time but also leads to low research and development efficiency, making it difficult to quickly find the optimal catalyst-reaction condition combination. Machine learning technology, with its ability to uncover complex nonlinear relationships and its data-driven predictive advantages, offers a new approach to solving these problems. However, its application in the field of metal-catalyzed lignin depolymerization still lacks a systematic optimization strategy, and the intelligent transformation of the entire process from performance prediction to catalyst design and reaction condition control has not yet been achieved.
[0005] Therefore, there is an urgent need for a highly efficient optimization method based on machine learning to accurately predict the effect of metal catalysts on lignin depolymerization, identify key influencing factors of the reaction, and screen the optimal combination of catalyst preparation and reaction conditions based on this, so as to replace the traditional trial-and-error mode and accelerate the research and application process of high-value utilization of lignin. Summary of the Invention
[0006] To achieve the above-mentioned objectives, and to guide the design and preparation of metal catalysts and the precise control of lignin depolymerization reaction conditions, thereby achieving a high-efficiency increase in the yield of bio-oil monomers, this invention adopts the following technical solution: A machine learning-based method for optimizing lignin depolymerization conditions using metal catalysts includes the following steps: Obtain a multidimensional dataset; the multidimensional dataset includes: the physicochemical properties and preparation parameters of the metal catalyst used in the depolymerization of lignin catalyzed by the metal catalyst, the characteristics of the lignin substrate, and the reaction conditions for the depolymerization of lignin; wherein, the characteristics of the lignin substrate include substrate type, lignin monomer composition, β-O-4 bond ratio, and C-C bond ratio; An initial yield prediction model is constructed. The inputs of the initial yield prediction model are the physicochemical properties and preparation parameters of the metal catalyst, the characteristics of the lignin substrate, and the lignin depolymerization reaction conditions. The output is the yield of bio-oil monomers obtained by the depolymerization of lignin catalyzed by the metal catalyst. Based on the inhibitory effect of intermediate product repolymerization during depolymerization, the bio-oil monomer yield output by the initial yield prediction model is corrected to obtain the corrected bio-oil monomer yield. After training, a corrected yield prediction model with the corrected bio-oil monomer yield as the output is obtained. Then, based on the corrected yield prediction model, the importance of the model input is analyzed to obtain several dominant factors affecting the bio-oil monomer yield of lignin depolymerization catalyzed by metal catalysts. Using multiple dominant factors as the search space and a modified yield prediction model as the objective function, an iterative search is performed using a Bayesian optimization-guided acquisition function. In each iteration, the next set of dominant factors to be evaluated is selected based on the posterior distribution of the current Gaussian process. After convergence, the combination of at least one dominant factor that maximizes the yield of bio-oil monomers is output. The acquisition function is either an expected improvement function or a confidence upper bound function. This combination of at least one dominant factor is used to guide the preparation of metal catalysts and / or the optimization of lignin depolymerization reaction conditions.
[0007] Explanation: The above method incorporates the inhibitory effect of intermediate product repolymerization during lignin depolymerization into the yield prediction model, significantly improving the model's ability to characterize real reaction behavior. By screening dominant factors from multidimensional inputs through importance analysis, and then using Bayesian optimization to efficiently iteratively search for the optimal parameter combination, it can quickly locate the catalyst preparation and reaction conditions that maximize the yield of bio-oil monomers under limited experimental or computational resources. This effectively overcomes the shortcomings of traditional trial-and-error methods, such as high cost, the curse of dimensionality, and neglect of the influence of repolymerization. It provides a data-driven, high-precision, interpretable, and cost-effective optimization strategy for metal-catalyzed lignin depolymerization.
[0008] Furthermore, the physicochemical properties and preparation parameters of the metal catalyst include specific surface area, pore volume, pore size, metal type, loading, catalyst-lignin ratio, and catalyst-solvent ratio; the depolymerization reaction conditions include lignin-solvent ratio, reactor volume-solvent ratio, reaction pressure, reaction time, reaction temperature, and solvent type.
[0009] Note: The clear definition of the above parameters provides a clear characteristic basis for subsequent importance analysis and screening of dominant factors, and also makes the Bayesian optimization search space more targeted and operable. The final output of the optimized parameter combination can directly guide the precise control of catalyst preparation process and depolymerization reaction conditions in industry, and enhance the engineering practical value of the technical solution and the clarity of the scope of patent protection.
[0010] Furthermore, the input to the initial yield prediction model also includes the bond weight vector of the lignin substrate; The method for obtaining the key weight vector includes: Principal component analysis was performed on the proportions of β-O-4 bonds, C-C bonds, and 5-5' bonds in the lignin substrate characteristics to construct a correlation matrix between bond type and reaction pathway. Based on the metal type of the metal catalyst and the theoretical bond breaking energy for each type of bond, the reciprocal of the theoretical bond breaking energy is normalized to obtain the theoretical contribution weight of each bond type; the theoretical contribution weight is linearly mapped according to the bond type-reaction path correlation matrix to obtain the weight coefficient of each bond type, and a bond type weight vector is formed.
[0011] In the bond type weight vector, the weights are positively correlated with the ease of bond breaking and the contribution to monomer yield. β-O-4 bonds are easily broken and contribute the most, with the highest weight, contributing approximately 70% of the monomer yield. C–C bonds are relatively difficult to break, with a medium weight, contributing approximately 20% of the monomer yield. 5–5' bonds are the most difficult to break, with the lowest weight, contributing approximately 10% of the monomer yield. This invention calculates the bond type weights using PCA combined with theoretical bond breaking energy. This vector is a three-dimensional vector with the form: W = [0.7, 0.2, 0.1]. Exploratory data analysis was conducted on the constructed multidimensional dataset. The distribution characteristics of each input parameter were visualized using cloud and rain plots. Combined with box plots and the scatter distribution of the original data, the range, distribution pattern and dispersion of key variables such as specific surface area, pore volume, pore size, reaction temperature, reaction pressure and reaction time were systematically analyzed to clarify the central trend and dispersion characteristics of the dataset.
[0012] Note: The above method combines principal component analysis with theoretical bond breaking energy to construct a bond type weight vector with clear chemical meaning, enabling the initial yield prediction model to quantify the differences in the contribution of different bond types to monomer yield, and significantly improving the physical interpretability and prediction accuracy of the model.
[0013] Furthermore, the initial yield prediction model is any one of XGBoost, KNN, SVM, LR, and RF.
[0014] The bio-oil monomer yield prediction model was selected from machine learning models such as XGBoost, KNN (K-Nearest Neighbors), SVM (Support Vector Machine), LR (Linear Regression), and RF (Random Forest). The selection method was as follows: (1) Divide the multidimensional dataset into a training set and a test set, with the training set accounting for 90% and the test set accounting for 10%; (2) Each machine learning model was trained on the training set, and hyperparameters were tuned by combining Spearman and grid search. (3) Evaluate the predictive performance of each optimized machine learning model on the test set. Evaluation metrics include the coefficient of determination (R²). 2 ) and root mean square error (RMSE); (4) Compare the evaluation metrics of each machine learning model and select the model with the highest coefficient of determination and the lowest root mean square error as the bio-oil monomer yield prediction model.
[0015] Note: The above method compares five machine learning models—XGBoost, KNN, SVM, LR, and RF—using a fixed partition (90% training set, 10% test set), Spearman spectroscopy combined with grid search for hyperparameter tuning, and R... 2 The dual evaluation of RMSE objectively selects the model with the best predictive performance as the bio-oil monomer yield prediction model, thereby avoiding the bias of subjective or single model, ensuring that the model has the highest fitting accuracy and the smallest prediction error, significantly improving the accuracy of yield prediction and the model's generalization ability, and providing a reliable foundation for subsequent iterative optimization based on the model.
[0016] Furthermore, based on the inhibitory effect of intermediate product repolymerization during depolymerization, the method for correcting the bio-oil monomer yield output by the initial yield prediction model includes: Based on the reaction temperature and the solvent's hydrogen supply capacity, the theoretical steady-state concentration threshold of the phenolic monomers obtained during the depolymerization process was calculated in the reaction system. When the concentration of bio-oil monomers corresponding to the bio-oil monomer yield output by the initial yield prediction model exceeds the theoretical steady-state concentration threshold, the bio-oil monomer yield is reduced by a set coefficient; the set coefficient is jointly determined by the mass ratio of lignin to metal catalyst and the mass ratio of solvent to lignin.
[0017] Note: The above method calculates the theoretical steady-state concentration threshold of phenolic monomers and dynamically sets the down-adjustment coefficient based on the catalyst-lignin ratio and solvent-lignin ratio. It uses the inhibitory effect of intermediate product repolymerization during depolymerization to correct the yield, effectively avoiding the overly optimistic prediction of the yield in the high-concentration region by the initial model. It significantly improves the accuracy and reliability of the corrected yield prediction model under conditions close to the actual reaction limit, and provides a more realistic optimization target for subsequent Bayesian optimization.
[0018] Furthermore, the theoretical steady-state concentration threshold of the phenolic monomers obtained during the depolymerization process in the reaction system is calculated using the following formula (1): (1) In equation (1), [H_donor] represents the theoretical steady-state concentration threshold, and [H_donor] is the solvent's hydrogen-donating capacity parameter. The rate constant for the hydrogen transfer reaction is... The repolymerization rate constant is R is the repolymerization activation energy, R is the gas constant, and T is the reaction temperature.
[0019] Note: The above method calculates the theoretical steady-state concentration threshold by introducing formulas for hydrogen transfer rate constant, repolymerization rate constant, repolymerization activation energy, solvent hydrogen supply capacity, and reaction temperature. It expresses the competitive relationship between depolymerization and repolymerization with quantifiable physicochemical parameters, which significantly improves the rigor of the yield correction mechanism and the reproducibility of the calculation, and provides a clear mathematical criterion for subsequent judgment on whether the monomer concentration exceeds the threshold.
[0020] Furthermore, when the concentration of bio-oil monomers corresponding to the bio-oil monomer yield output by the initial yield prediction model is greater than the theoretical steady-state concentration threshold, the bio-oil monomer yield is calculated using the following formula (2). (2) In equation (2), For the yield of bio-oil monomers, Here, β is the set coefficient, and 0 < β < 1. The output of the initial yield prediction model is the bio-oil monomer yield; ,in, , ε is the proportionality coefficient, L / C is the mass ratio of lignin to metal catalyst, and S / L is the mass ratio of solvent to lignin.
[0021] Note: The above formula allows the correction range to be dynamically adjusted with the amount of catalyst and the solvent ratio, thereby significantly improving the extrapolation reliability and engineering applicability of the corrected yield prediction model in the repolymerization-dominant region.
[0022] Furthermore, the importance analysis includes: The permutation importance algorithm was used to calculate the relative importance score of each input to the prediction result of bio-oil monomer yield, and to identify several candidate dominant factors with the highest importance scores. The dominant factors include substrate type, solvent-catalyst ratio, solvent-lignin ratio, catalyst-lignin ratio, and reactor volume-solvent ratio.
[0023] Note: The above method can accurately identify the key factors that truly dominate the yield of bio-oil monomers, such as substrate type, solvent-catalyst ratio, solvent-lignin ratio, catalyst-lignin ratio, and reactor volume-solvent ratio. This provides a highly efficient search space after dimensionality reduction for subsequent Bayesian optimization, significantly reducing the number of experiments and computational costs.
[0024] Furthermore, the method for determining the completion of the iteration convergence is as follows: in multiple consecutive iterations, the change in the maximum predicted yield output by the corrected yield prediction model is less than a preset tolerance threshold, or the number of iterations reaches the preset maximum number of iterations. The guidance for the preparation of metal catalysts and / or the optimization of lignin depolymerization reaction conditions includes: Based on the physicochemical properties and preparation parameters of the metal catalyst in at least one parameter combination selected through Bayesian optimization iteration, the process parameters for preparing the metal catalyst are determined; based on the depolymerization reaction condition parameters in the parameter combination selected through Bayesian optimization iteration, the reaction operation conditions for lignin depolymerization are determined, including solvent type, reaction temperature, reaction pressure, reaction time, and lignin-solvent ratio. Bayesian optimization iterative methods specifically include: (1) Set a feasible range of values for each dominant factor, which is determined based on the experimental range of existing literature and industrial production process constraints; (2) Using the bio-oil monomer yield prediction model corrected by intermediate product repolymerization as the objective function, a Bayesian optimization surrogate model is constructed. (3) Within the feasible region of parameters, the next set of parameters to be evaluated is intelligently selected through the sampling function, and areas with high predicted yield and high uncertainty are explored first. (4) Input the current parameter combination into the prediction model, calculate the objective function value, and update the surrogate model and posterior probability distribution; (5) Iteratively perform parameter sampling, model prediction, and surrogate model update until the convergence criterion is met; (6) Based on physicochemical constraints, the feasibility of candidate combinations in the iteration process is screened, and invalid combinations that do not meet the constraints of thermodynamic self-consistency, pore size-molecular size matching, and solvent-catalyst synergy are eliminated. (7) After the iteration converges, output the optimal combination of parameters that maximizes the yield of bio-oil monomers.
[0025] Note: The above method ensures the rationality of the optimization process termination by clearly defining the iterative convergence conditions, and systematically presents a complete optimization process from setting parameter ranges, constructing Bayesian surrogate models, intelligent sampling, model updates to physicochemical constraint screening. Among them, the introduction of constraints such as pore size-molecular size matching and solvent-catalyst synergy to eliminate invalid combinations significantly improves the engineering feasibility of the optimal parameter combination, so that the final output parameters can be directly used to guide the metal catalyst preparation process and the lignin depolymerization reaction operating conditions, thus closely linking data-driven optimization with actual production, and combining high efficiency and reliability.
[0026] Furthermore, it also includes building a feedback optimization mechanism: Metal catalysts were prepared and lignin depolymerization experiments were conducted according to the selected parameter combinations to obtain real bio-oil monomer yield data; the real bio-oil monomer yield data were compared with the predicted values of the bio-oil monomer yield prediction model to calculate the prediction deviation. When the prediction deviation exceeds the preset threshold, the actual bio-oil monomer yield data is added to the multidimensional dataset to retrain and update the parameters of the bio-oil monomer yield prediction model; the model prediction, experimental verification, and model update are iteratively executed until the prediction deviation stabilizes within the preset threshold, thus completing the feedback optimization.
[0027] Explanation: Through the closed-loop feedback mechanism described above, the bio-oil monomer yield prediction model can continuously absorb real experimental data for self-correction and iterative optimization until the prediction deviation stabilizes within the threshold. This effectively overcomes the systematic deviation between the initial model and the actual reaction system, significantly improves the long-term prediction accuracy and robustness of the model, and provides a continuously improving data-driven framework for the preparation of metal catalysts and the optimization of depolymerization reaction conditions.
[0028] The beneficial effects of this invention are: The optimization method of this invention can achieve high-precision prediction of bio-oil monomer yield: This invention integrates multi-dimensional data such as metal catalyst properties, lignin substrate characteristics, and depolymerization reaction conditions, and through systematic screening and hyperparameter fine-tuning of various machine learning models, the final selected XGBoost model has a high coefficient of determination (R²) for predicting bio-oil monomer yield. 2 The mean square error (RMSE) can reach 0.994, and the root mean square error (RMSE) is as low as 2.267, which is far better than the traditional model. It can accurately capture the complex nonlinear relationship between various parameters and bio-oil monomer yield, and provide a reliable data-driven basis for subsequent optimization.
[0029] Precise identification of the core dominant factors of depolymerization reaction: This invention, through substitution importance analysis, can robustly identify key factors that dominate the yield of bio-oil monomers, such as substrate type as the most significant influencing factor, solvent-catalyst ratio (5.6%), and solvent-lignin ratio (2.1%).
[0030] Replacing the traditional trial-and-error method, this invention achieves intelligent optimization of the entire process: By screening all factors and physicochemical constraints of the dominant factors, it exhaustively evaluates all feasible parameter combinations in a virtual space, avoiding the blindness of the traditional trial-and-error method and significantly reducing the number of experiments and resource waste. At the same time, it directly maps the optimal parameter combination to an executable metal catalyst preparation process and lignin depolymerization reaction conditions, realizing the intelligent transformation from performance prediction to optimization practice and significantly improving R&D efficiency.
[0031] Constructing a closed-loop optimization system enhances the adaptability and reliability of the strategy: The feedback optimization mechanism set up in this invention can supplement the dataset with real data verified by experiments and retrain the model, realizing continuous iterative upgrades of the model; at the same time, the multi-objective optimization steps take into account multiple indicators such as bio-oil monomer yield, catalyst stability, and economic cost, making the optimization strategy more in line with the actual needs of industrial production, and possessing good engineering practicality and promotion value.
[0032] This invention provides a generalizable research paradigm: The end-to-end intelligent framework of "data construction - model training - factor identification - parameter optimization - experimental verification - feedback update" is not only applicable to lignin depolymerization systems catalyzed by metal catalysts, but can also be extended to other complex chemical systems such as biomass catalytic conversion and catalyst design, providing a general research method for data analysis and optimization in related fields. Attached Figure Description
[0033] Figure 1 This is a schematic diagram of the overall framework of the method according to an embodiment of the present invention; Figure 2 Inputs used to develop machine learning models: lignin-solvent ratio, lignin-catalyst ratio, reactor volume-solvent ratio, maximum (maximum), mean, and minimum (minimum) values of the solvent-catalyst ratio. Figure 3 These are the inputs used to develop machine learning models: specific surface area, pore volume, pore capacity, reaction time, temperature, pressure, maximum (maximum), mean, and minimum (minimum).
[0034] Figure 4 This is a Spearman correlation heatmap of the parameters in the multidimensional dataset of this invention embodiment. The color intensity represents the Spearman correlation coefficient, with red indicating positive correlation and blue indicating negative correlation. Figure 5These are comparison charts of the coefficient of determination and root mean square error values predicted by various machine learning models in embodiments of the present invention. Figure 6 This is a scatter plot of the predicted bio-oil monomer yield and experimentally measured values using the XGBoost target aromatic monomer yield prediction model in this embodiment of the invention. Figure 7 This is an analysis of the importance of metal catalysts in depolymerizing lignin in the embodiments of the present invention. Detailed Implementation
[0035] 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.
[0036] Example 1: A machine learning-based method for optimizing the conditions for depolymerizing lignin in metal catalysts, such as... Figure 1 As shown, it includes the following steps: S101. Obtain a multidimensional dataset; the multidimensional dataset includes: the physicochemical properties and preparation parameters of the metal catalyst used in the depolymerization of lignin catalyzed by the metal catalyst, the characteristics of the lignin substrate, and the reaction conditions for the depolymerization of lignin; wherein, the characteristics of the lignin substrate include substrate type, lignin monomer composition, β-O-4 bond ratio, and C-C bond ratio. The physicochemical properties and preparation parameters of the metal catalyst include specific surface area, pore volume, pore size, metal type, loading, catalyst-lignin ratio, and catalyst-solvent ratio; the depolymerization reaction conditions include lignin-solvent ratio, reactor volume-solvent ratio, reaction pressure, reaction time, reaction temperature, and solvent type.
[0037] Specifically, using the Web of Science database with "metal-catalyzed lignin" as the search term, 37 relevant research papers published in the past 20 years were retrieved. From these, 198 experimental data points on metal-catalyzed lignin depolymerization were extracted, and a multidimensional dataset was constructed. The dataset includes three main categories of input parameters: physicochemical properties and preparation parameters of the metal catalyst, lignin substrate characteristics, and depolymerization reaction conditions. The output parameter is the bio-oil monomer yield (wt%). The specific parameters are as follows: Physicochemical properties and preparation parameters of metal catalysts: specific surface area (m²) 2 / g), pore volume (cm) 3 / g), pore size (nm), metal type (transition metal / noble metal), metal loading (%), catalyst-lignin ratio (mg / mL), catalyst-solvent ratio (mg / mL). Lignin substrate characteristics: substrate type (poplar lignin, poplar organic solvent lignin, guaiacol, soda lignin, diphenyl ether, etc.), lignin monomer composition (G type / S type / H type), β-O-4 bond percentage (%), C-C bond percentage (%), 5-5' bond percentage (%). Depolymerization reaction conditions and parameters: lignin-solvent ratio (mg / mL), reactor volume-solvent ratio (vol / vol), reaction pressure (MPa), reaction time (h), reaction temperature (°C), solvent type (ethanol / water / ethanol-water mixed solvent, etc.); as shown in Table 1.
[0038] Table 1 Data Acquisition Parameters
[0039] Exploratory data analysis (EDA) of the dataset (e.g.) Figure 2 , Figure 3 As shown in the figure, a rain cloud plot is used to visualize the distribution characteristics of each input variable. A box plot is used to display key statistics such as the median and quartiles, and a scatter plot is used to present the distribution of the original data points. The range, distribution pattern, and dispersion of variables such as specific surface area, pore volume, pore size, reaction temperature, pressure, and reaction time are clearly defined to guide subsequent data cleaning, parameter selection, and model construction.
[0040] The dataset was cleaned and outlier removed. Spearman rank correlation analysis was used to analyze the correlation between various parameters (e.g., Figure 4 As shown in the figure, the synergistic effect of catalyst pore structure (pore volume, pore size) and specific surface area, as well as the regulation of yield by core parameters such as temperature and solvent ratio, are clearly defined.
[0041] S102. Construct an initial yield prediction model; the input of the initial yield prediction model is the physicochemical properties and preparation parameters of the metal catalyst, the characteristics of the lignin substrate and the lignin depolymerization reaction conditions, and the output is the yield of bio-oil monomers obtained by the depolymerization of lignin catalyzed by the metal catalyst. The input to the initial yield prediction model also includes a lignin substrate bond weight vector; the method for obtaining the bond weight vector includes: ① Principal component analysis was performed on the proportions of β-O-4 bonds, C-C bonds, and 5-5' bonds in the lignin substrate characteristics to construct a correlation matrix between bond type and reaction pathway; ② Based on the metal type of the metal catalyst and the theoretical bond breaking energy for each type of bond, the reciprocal of the theoretical bond breaking energy is normalized to obtain the theoretical contribution weight of each bond type; the theoretical contribution weight is linearly mapped according to the bond type-reaction path correlation matrix to obtain the weight coefficient of each bond type, and a bond type weight vector is formed.
[0042] Specifically, in the aforementioned bond type weight vector, the weight is positively correlated with the "ease of bond breaking and contribution to monomer yield." β-O-4 is easily broken and contributes the most, with the highest weight, contributing approximately 70% of the monomer yield; C–C bonds are relatively difficult to break, with a medium weight, contributing approximately 20% of the monomer yield; 5–5' bonds are the most difficult to break, with the lowest weight, contributing approximately 10% of the monomer yield. This invention calculates the bond type weights using PCA combined with theoretical bond breaking energy. This vector is a three-dimensional vector with the form: W = [0.7, 0.2, 0.1]. Exploratory data analysis was conducted on the constructed multidimensional dataset. The distribution characteristics of each input parameter were visualized using cloud and rain plots. Combined with box plots and the scatter distribution of the original data, the range, distribution pattern and dispersion of key variables such as specific surface area, pore volume, pore size, reaction temperature, reaction pressure and reaction time were systematically analyzed to clarify the central trend and dispersion characteristics of the dataset.
[0043] The initial yield prediction model is any one of XGBoost, KNN, SVM, LR, and RF.
[0044] In this embodiment of the invention, the bio-oil monomer yield prediction model is selected from machine learning models such as XGBoost, KNN (K-Nearest Neighbors), SVM (Support Vector Machine), LR (Linear Regression), and RF (Random Forest). The selection method is as follows: (1) Divide the multidimensional dataset into a training set and a test set, with the training set accounting for 90% and the test set accounting for 10%; (2) Each machine learning model was trained on the training set, and hyperparameters were tuned by combining Spearman and grid search. (3) Evaluate the predictive performance of each optimized machine learning model on the test set. Evaluation metrics include the coefficient of determination (R²). 2 ) and root mean square error (RMSE); (4) Compare the evaluation metrics of each machine learning model and select the model with the highest coefficient of determination and the lowest root mean square error as the bio-oil monomer yield prediction model.
[0045] For example, using the Python-based Scikit-learn machine learning library, five machine learning models—XGBoost, KNN (K-Nearest Neighbors), SVM (Support Vector Machine), LR (Linear Regression), and RF (Random Forest)—are selected for regression prediction. The specific steps are as follows: The cube was randomly divided into a training set (178 data points) and a test set (20 data points) in a 9:1 ratio. Hyperparameter tuning for each model is performed using five-fold cross-validation combined with grid search. For example, hyperparameters such as learning rate, tree depth, and number of leaf nodes are optimized for the XGBoost model to avoid overfitting. On the test set, the coefficient of determination (R²) is used. 2 The predictive performance of each model is evaluated using the root mean square error (RMSE), calculated as follows: Where n is the number of test samples, These are experimental values. For predicted values, This represents the average of the experimental values.
[0046] Model evaluation results are as follows Figure 5 As shown, the XGBoost model exhibits the best predictive performance, with an R² value for predicting bio-oil monomer yield. 2 The XGBoost model achieved a performance score of 0.994 and an RMSE of 2.267, significantly outperforming other models such as Random Forest and SVM. This model utilizes a gradient boosting decision tree architecture to handle complex nonlinear relationships. Its built-in L1 and L2 regularization techniques effectively mitigate overfitting risks and can directly handle sparse data, making it well-suited to the characteristics of the dataset in this study. Therefore, the XGBoost model was selected as the bio-oil monomer yield prediction model.
[0047] To further improve model performance, features with no or low predictive power are removed based on the feature importance of the initial XGBoost model, resulting in the optimal feature combination. The optimized model's R-value is improved. 2 Increased to 0.956, RMSE decreased to 6.597 (e.g.) Figure 6 As shown in the figure, the prediction accuracy has been further improved.
[0048] S103. Based on the inhibitory effect of intermediate product repolymerization during depolymerization, the bio-oil monomer yield output by the initial yield prediction model is corrected to obtain the corrected bio-oil monomer yield. After training, a corrected yield prediction model with the corrected bio-oil monomer yield as the output is obtained. Then, based on the corrected yield prediction model, the importance of the model input is analyzed to obtain several dominant factors affecting the bio-oil monomer yield of lignin depolymerization catalyzed by metal catalyst. Methods for correcting the bio-oil monomer yield output by the initial yield prediction model include: ①Based on the reaction temperature and the solvent's hydrogen supply capacity, calculate the theoretical steady-state concentration threshold of the phenolic monomers obtained during the depolymerization process in the reaction system; ② When the concentration of bio-oil monomers corresponding to the bio-oil monomer yield output by the initial yield prediction model exceeds the theoretical steady-state concentration threshold, the bio-oil monomer yield is reduced by a set coefficient; the set coefficient is jointly determined by the mass ratio of lignin to metal catalyst and the mass ratio of solvent to lignin.
[0049] The theoretical steady-state concentration threshold of the phenolic monomers obtained during the depolymerization process in the reaction system is calculated using the following formula (1): (1) In equation (1), [H_donor] represents the theoretical steady-state concentration threshold, and [H_donor] is the solvent's hydrogen-donating capacity parameter. The rate constant for the hydrogen transfer reaction is... The repolymerization rate constant is R is the repolymerization activation energy, R is the gas constant, and T is the reaction temperature.
[0050] When the concentration of bio-oil monomers corresponding to the bio-oil monomer yield output by the initial yield prediction model is greater than the theoretical steady-state concentration threshold, the bio-oil monomer yield is calculated using the following formula (2). (2) In equation (2), Here, β is the set coefficient, and 0 < β < 1. The output of the initial yield prediction model is the bio-oil monomer yield; ,in, , ε is the proportionality coefficient, L / C is the mass ratio of lignin to metal catalyst, and S / L is the mass ratio of solvent to lignin.
[0051] The importance analysis includes: The permutation importance algorithm is used to calculate the relative importance score of each input to the predicted bio-oil monomer yield, and to identify multiple candidate dominant factors with the highest importance scores. In other embodiments, SHAP values, Gini importance built into random forests, and other methods can also be used for importance analysis. The permutation importance algorithm used in the embodiments of this invention is an existing technology. The dominant factors include substrate type, solvent-catalyst ratio, solvent-lignin ratio, catalyst-lignin ratio, and reactor volume-solvent ratio.
[0052] Combining importance score results (e.g.) Figure 7 The study identified substrate type, solvent-catalyst ratio, solvent-lignin ratio, catalyst-lignin ratio, and reactor volume-solvent ratio as the core dominant factors. Among these, substrate type accounted for approximately 42% of the total importance (poplar lignin 0.224, poplar organic solvent lignin 0.066, guaiacol 0.045, soda lignin 0.044, diphenyl ether 0.034), making it the most critical factor affecting the yield of bio-oil monomers.
[0053] The patterns of action of each dominant factor are as follows: Substrate type: The monomer composition, bond type distribution, and functional group density of lignin determine its depolymerization potential. Hardwood lignin (SG type) has a high proportion of β-O-4 bonds and a low proportion of C-C bonds, resulting in a significantly higher yield of bio-oil monomers than softwood lignin and herbaceous lignin. Organic solvent lignin retains more native β-O-4 bonds and has higher purity, thus exhibiting better depolymerization performance than sulfate lignin and soda lignin. Solvent-catalyst ratio: The hydrogen-donating capacity and anti-coking effect of the solvent can have a synergistic effect with the active sites of the catalyst. An appropriate ratio can improve the contact efficiency between the catalyst and lignin and promote the depolymerization reaction. Solvent-lignin ratio: Regulates the solubility-dispersion state of lignin and the diffusion-aggregation balance of depolymerization intermediates, avoiding increased mass transfer resistance or monomer repolymerization caused by excessively high / low lignin concentration.
[0054] S104. Using multiple dominant factors as the search space and a modified yield prediction model as the objective function, an iterative search is performed using a Bayesian optimization-guided acquisition function. In each iteration, the next set of dominant factors to be evaluated is selected based on the posterior distribution of the current Gaussian process. After the iteration converges, the combination of at least one dominant factor that maximizes the yield of bio-oil monomers is output. The acquisition function is an expected improvement function or a confidence upper bound function. The combination of at least one dominant factor is used to guide the preparation of metal catalysts and / or the optimization of lignin depolymerization reaction conditions.
[0055] The method for determining the completion of the iteration convergence is as follows: in multiple consecutive iterations, the change in the maximum predicted yield output by the corrected yield prediction model is less than the preset tolerance threshold, or the number of iterations reaches the preset maximum number of iterations. Bayesian optimization iterative methods specifically include: (1) Set a feasible range of values for each dominant factor, which is determined based on the experimental range of existing literature and industrial production process constraints; (2) Using the modified yield prediction model as the objective function, construct a Bayesian optimization surrogate model; (3) Within the feasible region of parameters, the next set of parameters to be evaluated is intelligently selected through the acquisition function, and regions with high predicted yield and high uncertainty are explored first; the acquisition function is the expected improvement function. (4) Input the current parameter combination into the prediction model, calculate the objective function value, and update the surrogate model and posterior probability distribution; (5) Iteratively perform parameter sampling, model prediction, and surrogate model update until the convergence criterion is met; (6) Based on physicochemical constraints, the feasibility of candidate combinations in the iteration process is screened, and invalid combinations that do not meet the constraints of thermodynamic self-consistency, pore size-molecular size matching, and solvent-catalyst synergy are eliminated. (7) After the iteration converges, output the optimal combination of parameters that maximizes the yield of bio-oil monomers.
[0056] The guidance for the preparation of metal catalysts and / or the optimization of lignin depolymerization reaction conditions includes: Based on the physicochemical properties and preparation parameters of the metal catalyst in at least one parameter combination selected through Bayesian optimization iteration, the process parameters for preparing the metal catalyst are determined; based on the depolymerization reaction condition parameters in the parameter combination selected through Bayesian optimization iteration, the reaction operation conditions for lignin depolymerization are determined, wherein the reaction operation conditions include solvent type, reaction temperature, reaction pressure, reaction time, and lignin-solvent ratio.
[0057] It also includes building a feedback optimization mechanism: Metal catalysts were prepared and lignin depolymerization experiments were conducted according to the selected parameter combinations to obtain real bio-oil monomer yield data; the real bio-oil monomer yield data were compared with the predicted values of the bio-oil monomer yield prediction model to calculate the prediction deviation. When the prediction deviation exceeds a preset threshold, real bio-oil monomer yield data is added to the multidimensional dataset to retrain and update the parameters of the bio-oil monomer yield prediction model. The process of iteratively executing model prediction, experimental verification, and model updates continues until the prediction deviation stabilizes within the preset threshold, completing the feedback optimization. S105. Experimental Verification and Feedback Optimization: The metal catalyst preparation and lignin depolymerization experiments were carried out according to the above optimized parameters. The relative deviation between the actual bio-oil monomer yield and the model prediction value was less than the preset threshold of 5%, indicating that the model prediction results are reliable.
[0058] If the prediction deviation exceeds the preset threshold, the real data from this experiment is added to the multidimensional dataset, the XGBoost model is retrained and the parameters are updated, and the model prediction, experimental verification and model update are iteratively executed until the prediction deviation stabilizes within the threshold, and the feedback optimization is completed.
Claims
1. A method for optimizing lignin depolymerization conditions using metal catalysts based on machine learning, characterized in that, Includes the following steps: Obtain a multidimensional dataset; the multidimensional dataset includes: the physicochemical properties and preparation parameters of the metal catalyst used in the depolymerization of lignin catalyzed by the metal catalyst, the characteristics of the lignin substrate, and the reaction conditions for the depolymerization of lignin; wherein, the characteristics of the lignin substrate include substrate type, lignin monomer composition, β-O-4 bond ratio, and C-C bond ratio; An initial yield prediction model is constructed. The inputs of the initial yield prediction model are the physicochemical properties and preparation parameters of the metal catalyst, the characteristics of the lignin substrate, and the lignin depolymerization reaction conditions. The output is the yield of bio-oil monomers obtained by the depolymerization of lignin catalyzed by the metal catalyst. Based on the inhibitory effect of intermediate product repolymerization during depolymerization, the bio-oil monomer yield output by the initial yield prediction model is corrected to obtain the corrected bio-oil monomer yield. After training, a corrected yield prediction model with the corrected bio-oil monomer yield as the output is obtained. Then, based on the corrected yield prediction model, the importance of the model input is analyzed to obtain several dominant factors affecting the bio-oil monomer yield of lignin depolymerization catalyzed by metal catalysts. Using multiple dominant factors as the search space and a modified yield prediction model as the objective function, an iterative search is performed using a Bayesian optimization-guided acquisition function. In each iteration, the next set of dominant factors to be evaluated is selected based on the posterior distribution of the current Gaussian process. After convergence, the combination of at least one dominant factor that maximizes the yield of bio-oil monomers is output. The acquisition function is either an expected improvement function or a confidence upper bound function. This combination of at least one dominant factor is used to guide the preparation of metal catalysts and / or the optimization of lignin depolymerization reaction conditions.
2. The method for optimizing lignin depolymerization conditions using a metal catalyst based on machine learning as described in claim 1, characterized in that, The physicochemical properties and preparation parameters of the metal catalyst include specific surface area, pore volume, pore size, metal type, loading, catalyst-lignin ratio, and catalyst-solvent ratio; the depolymerization reaction conditions include lignin-solvent ratio, reactor volume-solvent ratio, reaction pressure, reaction time, reaction temperature, and solvent type.
3. The method for optimizing lignin depolymerization conditions using metal catalysts based on machine learning as described in claim 1, characterized in that, The input to the initial yield prediction model also includes a lignin substrate bond weight vector; The method for obtaining the key weight vector includes: Principal component analysis was performed on the proportions of β-O-4 bonds, C-C bonds, and 5-5' bonds in the lignin substrate characteristics to construct a correlation matrix between bond type and reaction pathway. Based on the metal type of the metal catalyst and the theoretical bond breaking energy for each type of bond, the reciprocal of the theoretical bond breaking energy is normalized to obtain the theoretical contribution weight of each bond type; the theoretical contribution weight is linearly mapped according to the bond type-reaction path correlation matrix to obtain the weight coefficient of each bond type, and a bond type weight vector is formed.
4. The method for optimizing lignin depolymerization conditions using a metal catalyst based on machine learning as described in claim 1, characterized in that, The initial yield prediction model is any one of XGBoost, KNN, SVM, LR, and RF.
5. The method for optimizing lignin depolymerization conditions using a metal catalyst based on machine learning as described in claim 1, characterized in that, Based on the inhibitory effect of intermediate product repolymerization during depolymerization, methods for correcting the bio-oil monomer yield output by the initial yield prediction model include: Based on the reaction temperature and the solvent's hydrogen supply capacity, the theoretical steady-state concentration threshold of the phenolic monomers obtained during the depolymerization process was calculated in the reaction system. When the concentration of bio-oil monomers corresponding to the bio-oil monomer yield output by the initial yield prediction model exceeds the theoretical steady-state concentration threshold, the bio-oil monomer yield is reduced by a set coefficient; the set coefficient is jointly determined by the mass ratio of lignin to metal catalyst and the mass ratio of solvent to lignin.
6. The method for optimizing lignin depolymerization conditions using a metal catalyst based on machine learning as described in claim 5, characterized in that, The theoretical steady-state concentration threshold of the phenolic monomers obtained during the depolymerization process in the reaction system is calculated using the following formula (1): （1） In equation (1), This is the theoretical steady-state concentration threshold. [H_donor] This is a parameter representing the solvent's hydrogen-donating capacity. The rate constant for the hydrogen transfer reaction is... The repolymerization rate constant is For re-polymerization activation energy, R The gas constant is T The reaction temperature is denoted as .
7. The method for optimizing lignin depolymerization conditions using a metal catalyst based on machine learning as described in claim 5, characterized in that, When the concentration of bio-oil monomers corresponding to the bio-oil monomer yield output by the initial yield prediction model is greater than the theoretical steady-state concentration threshold, the corrected bio-oil monomer yield is calculated using the following formula (2). （2） In equation (2), To correct the yield of bio-oil monomers, Here, β is the set coefficient, and 0 < β < 1. The output of the initial yield prediction model is the bio-oil monomer yield; ,in, , ε is the proportionality coefficient, L / C is the mass ratio of lignin to metal catalyst, and S / L is the mass ratio of solvent to lignin.
8. The method for optimizing lignin depolymerization conditions using a metal catalyst based on machine learning as described in claim 1, characterized in that, The importance analysis includes: The permutation importance algorithm was used to calculate the relative importance score of each input to the prediction result of bio-oil monomer yield, and to identify several candidate dominant factors with the highest importance scores. The dominant factors include substrate type, solvent-catalyst ratio, solvent-lignin ratio, catalyst-lignin ratio, and reactor volume-solvent ratio.
9. The method for optimizing lignin depolymerization conditions based on machine learning as described in claim 1, characterized in that, The method for determining the completion of the iteration convergence is as follows: in multiple consecutive iterations, the change in the maximum predicted yield output by the corrected yield prediction model is less than the preset tolerance threshold, or the number of iterations reaches the preset maximum number of iterations. The guidance for the preparation of metal catalysts and / or the optimization of lignin depolymerization reaction conditions includes: Based on the physicochemical properties and preparation parameters of the metal catalyst in at least one parameter combination selected through Bayesian optimization iteration, the process parameters for preparing the metal catalyst are determined; based on the depolymerization reaction condition parameters in the combination of dominant factors selected through Bayesian optimization iteration, the reaction operation conditions for lignin depolymerization are determined, including solvent type, reaction temperature, reaction pressure, reaction time, and lignin-solvent ratio.
10. The method for optimizing lignin depolymerization conditions based on machine learning as described in claim 1, characterized in that, It also includes building a feedback optimization mechanism: Metal catalysts were prepared and lignin depolymerization experiments were conducted according to the selected combination of dominant factors to obtain real bio-oil monomer yield data; the real bio-oil monomer yield data were compared with the predicted values of the modified yield prediction model to calculate the prediction deviation. When the prediction deviation exceeds the preset threshold, the actual bio-oil monomer yield data is added to the multidimensional dataset to retrain and update the parameters of the corrected yield prediction model; the model prediction, experimental verification, and model update are iteratively executed until the prediction deviation stabilizes within the preset threshold, thus completing the feedback optimization.