Biomass-based bifunctional catalyst screening method based on machine learning

Abstract

The invention provides a biomass-based bifunctional catalyst screening method based on machine learning. The biomass-based bifunctional catalyst screening method comprises the following steps: acquiring multi-source catalyst data, and cleaning, dividing and standardizing the multi-source catalyst data; feature engineering of mechanism and data fusion is carried out, and a core feature set containing specific derivative features is constructed; adopting a machine learning framework integrating transfer learning, meta learning and Bayesian optimization, and utilizing the core feature set to train a high-performance catalyst performance prediction model; and finally, screening candidate catalysts and outputting results by utilizing the model. According to the method, efficient and accurate screening of the Ni-Ca bifunctional catalyst in a small sample scene is realized through multi-source data integration, mechanism-data fusion feature engineering and a small sample adaptive multi-algorithm fusion framework.

Description

Technical Field

[0001] This invention relates to the field of catalyst screening and machine learning, specifically providing a machine learning-based biomass-based bifunctional catalyst screening method. Background Technology

[0002] Biomass-based (e.g., Ni-Ca) bifunctional catalysts have broad application prospects in the field of biomass resource utilization due to their excellent catalytic activity and stability. Traditional catalyst screening methods mainly rely on experimental trial and error, requiring numerous orthogonal experiments to optimize catalyst formulations and preparation parameters, resulting in long experimental cycles, high costs, and low efficiency. With the development of machine learning technology, data-driven catalyst screening methods have emerged, which achieve rapid prediction of catalyst performance by constructing predictive models, reducing the number of experiments required.

[0003] However, existing data-driven catalyst screening techniques still have the following shortcomings: (1) Poor adaptability to small samples: The preparation and performance testing of biomass-based Ni-Ca catalysts are difficult and costly, and the number of available samples is usually small (less than 30). Traditional machine learning models are prone to overfitting under small sample data, resulting in poor generalization ability and difficulty in guaranteeing prediction accuracy. (2) Single algorithm: Existing technologies mostly use a single machine learning algorithm, such as neural network and random forest, which do not fully combine the catalyst reaction mechanism and small sample learning characteristics, making it difficult to balance the model's fitting ability and generalization ability. (3) Poor interactive experience: The interface design of the existing screening system is mostly adapted to industrial production scenarios, and has not been optimized for the usage habits of researchers. It lacks intuitive process monitoring, parameter configuration and result visualization functions, and has a high operating threshold. (4) Weak data integration capability: It is difficult to effectively integrate multi-source data such as laboratory data and literature data, resulting in low data utilization and further limiting the improvement of model performance.

[0004] Therefore, proposing a screening method for biomass-based Ni-Ca bifunctional catalysts that is adaptable to small sample scenarios, integrates the advantages of multiple algorithms, is interactive, and supports the integration of multi-source data has become an urgent need in the current catalyst research and development field. Summary of the Invention

[0005] To overcome the above-mentioned shortcomings, this invention is proposed to provide solutions or at least partially solve the problems of poor small sample adaptability, simple algorithms, poor interactive experience, and weak data integration capabilities in the prior art.

[0006] This invention provides a machine learning-based method for screening biomass-based bifunctional catalysts, comprising the following steps: We acquire multi-source catalyst data, including biomass-based catalyst component formulations, preparation parameters, structural characteristics, and performance indicators. We then clean, segment, and standardize the data to form a training dataset. The preprocessed data is used to extract and screen fundamental features related to catalyst performance and generate derived features based on the reaction mechanism of the bifunctional catalyst to enhance the correlation between features and performance indicators, thus constructing a core feature set for model training; the fundamental features include at least raw material characteristics, preparation process characteristics, structural characteristics, and performance-related features. A machine learning framework integrating transfer learning, meta-learning, and Bayesian optimization is adopted, and the model is trained and hyperparameters are optimized using the core feature set to construct a catalyst performance prediction model. The characteristic data of the candidate catalysts are input into the catalyst performance prediction model to obtain the performance prediction results. The target catalysts are then selected based on the prediction results, and the selection results are output.

[0007] Preferably, the data cleaning includes: identifying outliers using the interquartile range method and verifying them in conjunction with knowledge of the catalyst field; The K-nearest neighbor algorithm is used to match the nearest samples of missing data based on feature similarity, calculate the missing values ​​by weighted average, and fill in the missing values.

[0008] Preferably, the data partitioning includes: dynamically selecting a partitioning strategy based on the sample size to partition the cleaned data; specifically: when the total sample size is less than a predetermined threshold, the leave-one-out method is used to partition the training set and the test set; when the total sample size is greater than or equal to the predetermined threshold, the training set and the test set are randomly partitioned according to a preset ratio.

[0009] Preferably, the core feature set for model training is constructed by extracting and screening fundamental and derived features related to catalyst performance from the preprocessed data, and includes at least the following: employing a strategy combining mechanism-oriented and data-driven approaches to preprocess the data, specifically: Based on the catalyst reaction mechanism, preset features that are directly related to catalytic activity and selectivity are retained from the basic and derived features; Secondly, using the SHAP value analysis method, features whose contribution to the prediction results exceeds a preset threshold are selected from the remaining features to jointly constitute the core feature set.

[0010] Preferably, the method further includes the generation of derived features, specifically: Based on Ni 0 Synergistic effect with CaO to generate Ni 0 Site density / CaO content ratio; Based on the correlation between the preparation process and structural properties, a correlation value between specific surface area and activation temperature is generated; Based on the correlation between feedstock characteristics and catalytic performance, an interaction term for C / H ratio and ash content was generated.

[0011] Preferably, a machine learning framework integrating transfer learning, meta-learning, and Bayesian optimization is used to train and optimize the model using the core feature set to construct a catalyst performance prediction model. This includes a transfer learning training process, specifically: The XGBoost regression model was selected as the base model, and catalyst data covering the characteristic-performance correlation of various metal oxide catalysts were used for pre-training. The core structural parameters of the pre-trained model are fixed, and the model is further trained by adjusting the learning rate and the number of new trees, using the core feature set as input. The core structural parameters include the learning rate and the number of new trees.

[0012] Preferably, a machine learning framework integrating transfer learning, meta-learning, and Bayesian optimization is used to train and optimize the model using the core feature set to construct a catalyst performance prediction model. This includes a meta-learning training process, specifically: A catalyst task generator is constructed to generate multiple simulated catalyst small-sample tasks, each task including a support set for rapid model fine-tuning and a query set for evaluation; The meta-parameters are updated by alternating between inner and outer loops. The inner loop quickly fine-tunes the model parameters and calculates the loss based on the support set, while the outer loop calculates the meta-loss and updates the meta-parameters based on the query set. Preferably, the fine-tuning includes dividing the real catalyst small sample data into a support set and a query set, and using the meta-parameters obtained from meta-training to fine-tune the model through an inner loop to adapt it to the real task.

[0013] Preferably, the method of using a machine learning framework that integrates transfer learning, meta-learning, and Bayesian optimization to train the model and optimize its hyperparameters using the core feature set to construct a catalyst performance prediction model further includes: using a Bayesian optimization algorithm to automatically optimize the hyperparameters in the transfer learning training process and / or the meta-learning training process with the model prediction accuracy as the optimization objective.

[0014] Preferably, the Bayesian optimization algorithm is used, with model prediction accuracy as the optimization objective, to automatically optimize the hyperparameters in the transfer learning training process and / or the meta-learning training process, specifically including: The coefficient of determination R of the prediction model 2 The core optimization objective is set as the forward prediction error root mean square error (RMSE) and used as the constraint condition. Construct a hyperparameter-performance probability model based on Gaussian processes; The expected improvement criterion is adopted as the sampling strategy, and the optimal hyperparameter combination is obtained by iterative search within the preset hyperparameter space. The obtained optimal hyperparameter combination is automatically configured into the corresponding training process.

[0015] The beneficial effects of this invention are as follows: 1. Significantly improves prediction accuracy and generalization ability under small sample data: By innovatively adopting the fusion framework of "transfer learning pre-training-meta-learning fine-tuning", it effectively utilizes general data prior knowledge and endows the model with the ability to quickly adapt to extremely small sample tasks, fundamentally overcoming the core problem of traditional machine learning models being prone to overfitting and poor generalization in small sample (especially less than 30 samples) scenarios.

[0016] 2. It achieves a deep integration of mechanistic knowledge and data-driven approaches, enhancing the interpretability and reliability of the model: In the feature engineering stage, a dual screening strategy of "mechanism-oriented + data-driven" is adopted, based not only on catalyst reaction mechanisms (such as Ni...) 0 The synergistic effect with CaO constructed derived features with clear physical meaning, and the SHAP value was used to quantify the feature contribution, ensuring the scientific rigor and efficiency of the core feature set. This makes the model prediction results not only accurate but also consistent with catalysis theory, providing researchers with a clear physical explanation and optimization direction.

[0017] 3. It achieves automated and efficient hyperparameter optimization, reducing the barrier to entry and trial-and-error costs: By introducing a Bayesian optimization process, it automatically seeks the optimal hyperparameters for the complex hyperparameter spaces of transfer learning and meta-learning using efficient probabilistic models and sampling strategies. This significantly reduces the number of hyperparameter adjustments to within 10-20 iterations to find the optimal combination. This process is fully automated, avoiding tedious manual parameter tuning, making the method easy to apply even for researchers in the field of catalysis with limited machine learning experience.

[0018] 4. Improved practicality and research efficiency: From multi-source data integration and automated preprocessing to the final visualization results and structured report output, this invention provides a complete end-to-end solution. This method significantly reduces the large number of experiments required by traditional trial-and-error methods, transforming catalyst screening from a time-consuming and labor-intensive experimental process into a highly efficient computationally assisted process guided by model prediction, greatly accelerating the R&D cycle and reducing R&D costs. Attached Figure Description

[0019] The disclosure of this invention will become more readily understood with reference to the accompanying drawings. It will be readily understood by those skilled in the art that these drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. Furthermore, similar numbers in the drawings are used to denote similar components, wherein: Figure 1 This is a framework diagram of a machine learning-based biomass-based bifunctional catalyst screening method according to an embodiment of the present invention; Figure 2 This is a migration mechanism according to one embodiment of the present invention; Figure 3 This is a meta-learning mechanism according to an embodiment of the present invention; Figure 4 This is a Bayesian optimization mode according to an embodiment of the present invention; Figure 5 This invention relates to a biomass-based Ni-Ca bifunctional catalyst and its biomass gasification performance. Detailed Implementation

[0020] Some embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0021] like Figure 1-5 As shown, this invention provides a machine learning-based method for screening biomass-based bifunctional catalysts, comprising: Step S1: Obtain multi-source catalyst data containing biomass-based catalyst component formulations, preparation parameters, structural characteristics, and performance indicators. Clean, divide, and standardize the data to form a training dataset.

[0022] Specifically, step S11, multi-source data acquisition and import: Step S111, Laboratory Data: This includes measured data on the raw material characteristics of Ni-Ca catalyst (C / H ratio of biochar, ash content, etc.), preparation parameters (activation temperature, Ni loading, etc.), and performance indicators (catalytic hydrogen production efficiency, stability, etc.), which can be imported in batches in Excel, CSV, and TXT formats. Step S112, Literature Data: By crawling relevant literature from academic databases such as CNKI and Web of Science using keywords (e.g., "biochar-based Ni-Ca catalyst"), structured catalyst formulations, preparation processes, and performance data are extracted to generate a standardized literature dataset. Through custom data field mapping rules, the two types of data are automatically matched and fused to form a multi-source integrated dataset.

[0023] Step S12, Data Cleaning: Construct a hierarchical data cleaning mechanism to ensure data quality. Step S121, Outlier Identification and Removal: The interquartile range (Q3-Q1) of each feature data is calculated using the IQR (interquartile range) method. Data exceeding the range of [Q1-1.5×IQR, Q3+1.5×IQR] are marked as outliers. Secondary verification is performed using knowledge from the catalyst field (e.g., ash content is usually no more than 25%, and Ni loading is generally 5%-15%) to avoid mistakenly deleting valid data. Finally, outlier samples are removed. Step S122, Missing value imputation: The K-Nearest Neighbors (KNN) algorithm is used to match the nearest samples of the missing data based on feature similarity. The missing value is calculated by weighted average. Compared with the traditional mean / median imputation, it can better preserve the data distribution characteristics.

[0024] Step S13, Data partitioning and standardization: Step S131, Data Splitting: Dynamically select a splitting strategy based on the sample size to ensure the rationality of the training and test sets: When the sample size is less than 10, use the leave-one-out method (reserve 1 sample as the test set each time, and use the rest as the training set, and verify cyclically); When the sample size is ≥10, randomly split the training and test sets in an 8:2 ratio, and fix the random seed (random_state=42) to ensure the reproducibility of the data splitting. Step S132, Standardization: The Z-score standardization method is used, and the calculation formula is as follows: X_std=(X-μ) / σ Where μ is the mean of the training set and σ is the standard deviation of the training set. The test set is transformed strictly based on the statistical parameters of the training set to avoid data leakage from the source. After standardization, all data have a mean of 0 and a standard deviation of 1, eliminating the difference in units and ensuring the convergence of subsequent model training.

[0025] Step S2: Extract and screen the basic features associated with catalyst performance from the preprocessed data and generate derived features that can enhance the correlation between features and performance indicators based on the reaction mechanism of the bifunctional catalyst, and construct a core feature set for model training; the basic features include at least raw material characteristics, preparation process characteristics, structural characteristics, and performance-related features.

[0026] This step, based on the reaction mechanism of Ni-Ca bifunctional catalysts, constructs a highly discriminative feature system to improve the learning efficiency and prediction accuracy of the model under small sample data. The specific implementation process is as follows: Step S21, Basic Feature Extraction: Construct a "four-dimensional feature system" to comprehensively cover factors affecting catalyst performance and achieve automatic extraction and quantification of basic features: Step S211, Raw material characteristics: C / H ratio, ash content, specific surface area, pore size distribution, etc. of biochar; Step S212, Preparation process characteristics: activation temperature, holding time, Ni loading, CaO content, Ni / Ca molar ratio, preparation method (impregnation / co-precipitation), etc. Step S213, Structural characteristics: Ni 0 Site density, Ni 0 Grain size, CaO dispersion, catalyst specific surface area, etc.; Step S214, Performance-related characteristics: Catalytic activity and cycle stability in preliminary experiments.

[0027] Step S22, Core Feature Screening: A dual-dimensional screening strategy of "mechanism-oriented + data-driven" is adopted to eliminate redundant features and retain key information: Step S221, Mechanism-guided screening: Based on the reaction mechanism of Ni-Ca bifunctional catalysts (Ni 0 The site is the catalytic active center, and CaO enables CO2 adsorption and anti-carbon deposition. Features directly related to catalytic activity (such as Ni) are preferentially retained. 0 Site density, CaO content, activation temperature), and other irrelevant features such as raw material collection time were excluded; Step S222, Data-driven screening: The SHAP (SHapley Additive exPlanations) value analysis method is used to calculate the contribution of each feature to the catalyst performance prediction results and quantify the importance of the features; the absolute value of the SHAP value ≥ 0.1 is set as the screening threshold (optimized according to the characteristics of small sample data), and finally 5-8 core features are retained to balance the feature information content and model complexity.

[0028] Step S23, Derivative Feature Generation: Based on the reaction mechanism, design targeted derived features to strengthen the correlation between features and performance indicators and improve the model's predictive ability. Step S231, based on Ni 0 Synergistic effect with CaO to generate "Ni 0 The "site density / CaO content ratio" is a characteristic that directly reflects the matching degree between the catalytic active center and the CO2 adsorption site, and is positively correlated with the hydrogen production efficiency. Step S232: Based on the correlation between the preparation process and structural characteristics, generate a "specific surface area-activation temperature correlation value" (specific surface area × ln(activation temperature)) to quantify the effect of activation temperature on the pore structure of the catalyst. Step S233: Based on the correlation between raw material characteristics and catalytic performance, a "C / H ratio-ash content interaction term" is generated to reflect the influence of biochar raw materials on catalyst stability; all derived characteristics have been verified by domain knowledge to ensure clear physical meaning.

[0029] Step S24, Feature Standardization: The Z-score standardization method is used to uniformly process the core features and derived features. The calculation formula is X_std=(X-μ) / σ (μ is the mean of the training set features, and σ is the standard deviation of the training set features). The training set and the test set are strictly distinguished, and only the training set parameters are used to transform the test set features to avoid data leakage. After standardization, all features are on the same order of magnitude, ensuring the effectiveness of subsequent model training.

[0030] like Figure 2-4 As shown, in step S3, a machine learning framework that integrates transfer learning, meta-learning and Bayesian optimization is adopted to train the model and optimize its hyperparameters using the core feature set, so as to construct a catalyst performance prediction model.

[0031] This step is the core of the screening method. It adopts a three-level fusion framework of "transfer learning + meta-learning + Bayesian optimization". Through the synergistic complementarity of multiple algorithms, it makes full use of general data prior knowledge and the characteristics of small sample tasks to achieve high-precision prediction of catalyst performance in small sample scenarios. The specific implementation process is as follows: Step S31, Transfer Learning Training Process (Applicable to scenarios with 5-30 samples): Employ a two-stage strategy of "pre-training-fine-tuning" to compensate for the shortcomings of small-sample Ni-Ca data by utilizing prior knowledge from general catalyst data. Step S311, Pre-training stage: The XGBoost regression model is selected as the base model (strong anti-overfitting ability and good adaptability to small sample data). The pre-training dataset uses general catalyst data (1000+ samples) from public materials science databases such as OQMD and Materials Project, covering the characteristic-performance correlation of various metal oxide catalysts. Pre-training parameter configuration: learning rate 0.1, maximum tree depth 6, subsample ratio 0.8, column sampling ratio 0.8, number of iterations 500, and an early stopping strategy (early_stopping_rounds=50) is adopted to avoid pre-training overfitting. Step S312, Fine-tuning stage: Fix the core parameters of the pre-trained model (tree structure, leaf node weights), and only adjust the learning rate (0.001~0.05, much smaller than the pre-training learning rate to avoid covering general knowledge) and the number of new trees (50~200, dynamically adjusted according to the small sample size); use the Ni-Ca core feature dataset output from step S2 as input to continue training the model to adapt to the specific characteristics and performance rules of Ni-Ca catalysts; the fine-tuning process adopts an early stopping strategy (early_stopping_rounds=30), and terminates training when the RMSE of the test set does not decrease for 30 consecutive rounds to ensure the model's generalization ability.

[0032] Step S32, Meta-learning Training Process (Applicable to scenarios with ≤5 samples): Based on the Model Independent Meta-Learning (MAML) algorithm, achieves fast model adaptation with a very small number of samples: Step S321, Meta-training Stage: Construct a catalyst task generator to generate 50-100 simulated catalyst small-sample tasks. Each task includes a "support set (6-8 samples, for rapid fine-tuning)" and a "query set (4-6 samples, for evaluating adaptation effect)". The feature distribution of the simulated tasks is generated based on the distribution of real Ni-Ca catalyst data to ensure task authenticity. Meta-parameters are updated through alternating training using inner and outer loops: the inner loop uses the support set to perform 1-2 steps of rapid fine-tuning of the model parameters (learning rate 0.001-0.02) and calculates the inner loop loss (MSE); the outer loop uses the query set to calculate the meta-loss (MSE) and updates the meta-parameters through the Adam optimizer (learning rate 0.0001-0.001), enabling the model to "rapidly adapt to new catalyst tasks". Step S322, Meta-fine-tuning stage: Divide the real Ni-Ca small sample data into a support set (3~6 samples) and a query set (2~4 samples). Using the meta-parameters obtained from meta-training, the model can be quickly adapted to the real Ni-Ca task and achieve high-precision performance prediction through only 1~2 steps of internal loop fine-tuning.

[0033] Step S33: Bayesian Hyperparameter Optimization: For the hyperparameter optimization problem in transfer learning and meta-learning, an efficient Bayesian optimization algorithm is adopted to reduce the number of hyperparameter adjustments (adapting to the limited experimental characteristics of small sample data). Step S331, Optimization objective: Based on the model's predicted R... 2 The score is the core optimization objective (R 2 The closer to 1, the higher the prediction accuracy), while constraining RMSE ≤ 5% (to match the accuracy requirements of catalyst performance testing). Step S332, Probabilistic Model Construction: A Gaussian Process (GP) is used as the hyperparameter-performance probabilistic model. The nonlinear relationship between hyperparameters and model performance is modeled through the prior distribution. The posterior distribution is updated based on the searched sample points to predict the performance of unsearched points. Step S333, Sampling Strategy: The next hyperparameter sampling point is selected using the Expected Improvement (EI) criterion, balancing "exploring unknown regions" and "utilizing known optimal regions" to improve search efficiency; Step S334, Optimization Scope: Cover the key hyperparameters of transfer learning and meta-learning (fine-tuning learning rate, maximum tree depth, etc. for transfer learning, inner / outer loop learning rate, hidden layer dimension, etc. for meta-learning); initially sample 3~5 points, iterate and optimize 10~20 times to find the optimal combination of hyperparameters; Step S335, Hyperparameter Synchronization: The optimized hyperparameters are automatically synchronized to the corresponding training algorithm to achieve adaptive improvement of model performance.

[0034] Step S34, Training process monitoring: Real-time display of training logs, loss curves, and training progress. When the training loss does not decrease for 30 consecutive rounds, the early stopping mechanism is triggered to terminate training and avoid overfitting. The model training status can be intuitively evaluated by the changing trends of the loss curves of the training set and the test set.

[0035] Step S4: Input the characteristic data of the candidate catalysts into the catalyst performance prediction model to obtain the performance prediction results, and screen out the target catalysts based on the prediction results, and output the screening results.

[0036] This step transforms model predictions into intuitive and usable data for scientific research decision-making. By combining visualization analysis with structured reports, it lowers the barrier to interpreting results and provides clear guidance for experimental design. The specific implementation process is as follows: Step S41, Performance Prediction and Screening: Input the characteristic data of the candidate Ni-Ca catalysts (after standardization in step S2) into the trained performance prediction model to obtain the performance prediction results of each candidate catalyst (such as hydrogen production efficiency, stability, etc.); sort the candidate catalysts according to the performance prediction values ​​and screen out the catalyst scheme with the best performance.

[0037] Step S42, Visual Analysis: Generate multi-dimensional visualization charts to intuitively present the filtering results and the correlation between features and performance. Step S421, Feature Importance Bar Chart: Based on the SHAP value analysis results, this chart shows the degree of influence of each core feature on the performance of the Ni-Ca catalyst (the horizontal axis represents the feature name, and the vertical axis represents the absolute value of the SHAP value), clarifying key influencing factors (such as Ni). 0 Site density, activation temperature); Step S422, Performance Prediction Curve: Display the distribution of the true performance values ​​and model prediction values ​​of the test set samples in the form of a scatter plot (horizontal axis is the true value, vertical axis is the predicted value), add the fitted line and R². 2 RMSE annotation provides an intuitive assessment of the model's prediction accuracy. Step S423, Structure-Performance Relationship Heatmap: Displays the correlation strength between different combinations of core features (Ni loading, CaO content, activation temperature) and catalyst performance (the darker the color, the better the performance), quickly locating the optimal feature combination area; all visualization charts support interactive operations (zooming, hovering to display detailed data), facilitating in-depth analysis.

[0038] Step S43, Structured Report Output: Automatically generate a standardized structured screening report, covering key information throughout the screening process: Step S431, core content includes: basic project information (selection task name, date), data overview (sample size, number of features, data source), feature engineering results (list of core features, description of derived features), model configuration (selected algorithm, optimal hyperparameter combination), prediction accuracy evaluation (R... 2 RMSE and MAE indices), optimal catalyst scheme (formulation: Ni loading, CaO content, Ni / Ca ratio; preparation parameters: activation temperature, holding time; performance prediction), screening conclusions and suggestions (proposing experimental verification directions based on results). Step S432: The report format supports exporting to two mainstream formats: PDF and Word. Charts and graphs are automatically embedded, and the layout is standardized. It can be directly used for scientific research records or results presentations.

[0039] like Figure 5 To verify the effectiveness of this method (TL-ML-BO), it was compared with Random Forest (RF), Transfer Learning (TL), and a fusion model of transfer learning and meta-learning (TL-ML). Under 100 sets of experimental data, the tar degradation rate (RMSE and ) was used as the benchmark. Using the metric , we can illustrate the advantages of the proposed method.

[0040] Table 1. Evaluation results of model prediction performance

[0041] Table 2 Actual Performance Tests

[0042] As shown in Tables 1 and 2, this method exhibits a significant advantage in tar degradation rate. Therefore, the proposed method demonstrates the best test results.

[0043] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the original technical features, and the technical solutions resulting from these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A method for screening biomass-based bifunctional catalysts based on machine learning, characterized in that, Includes the following steps: We acquire multi-source catalyst data, including biomass-based catalyst component formulations, preparation parameters, structural characteristics, and performance indicators. We then clean, segment, and standardize the data to form a training dataset. The preprocessed data is used to extract and screen fundamental features related to catalyst performance and generate derived features based on the reaction mechanism of the bifunctional catalyst to enhance the correlation between features and performance indicators, thus constructing a core feature set for model training; the fundamental features include at least raw material characteristics, preparation process characteristics, structural characteristics, and performance-related features. A machine learning framework integrating transfer learning, meta-learning, and Bayesian optimization is adopted, and the model is trained and hyperparameters are optimized using the core feature set to construct a catalyst performance prediction model. The characteristic data of the candidate catalysts are input into the catalyst performance prediction model to obtain the performance prediction results. The target catalysts are then selected based on the prediction results, and the selection results are output.

2. The method according to claim 1, characterized in that, The data cleaning includes: using the interquartile range method to identify outliers and verifying them in conjunction with knowledge of the catalyst field; The K-nearest neighbor algorithm is used to match the nearest samples of missing data based on feature similarity, calculate the missing values ​​by weighted average, and fill in the missing values.

3. The method according to claim 2, characterized in that, The data partitioning includes: dynamically selecting a partitioning strategy based on the sample size to partition the cleaned data. Specifically: when the total sample size is less than a predetermined threshold, the leave-one-out method is used to partition the training set and the test set; when the total sample size is greater than or equal to the predetermined threshold, the training set and the test set are randomly partitioned according to a preset ratio.

4. The method according to claim 1, characterized in that, Extracting and filtering fundamental and derived features related to catalyst performance from the preprocessed data, and constructing a core feature set for model training, includes at least the following: employing a strategy combining mechanism-oriented and data-driven approaches to preprocess the data, specifically: Based on the catalyst reaction mechanism, preset features that are directly related to catalytic activity and selectivity are retained from the basic and derived features; Secondly, using the SHAP value analysis method, features whose contribution to the prediction results exceeds a preset threshold are selected from the remaining features to jointly constitute the core feature set.

5. The method according to claim 1, characterized in that, The method also includes the generation of derived features, specifically: Based on Ni 0 Synergistic effect with CaO to generate Ni 0 Site density / CaO content ratio; Based on the correlation between the preparation process and structural properties, a correlation value between specific surface area and activation temperature is generated; Based on the correlation between feedstock characteristics and catalytic performance, an interaction term for C / H ratio and ash content was generated.

6. The method according to claim 1, characterized in that, A machine learning framework integrating transfer learning, meta-learning, and Bayesian optimization is employed. The model is trained and its hyperparameters optimized using the core feature set to construct a catalyst performance prediction model. This includes a transfer learning training process, specifically: The XGBoost regression model was selected as the base model, and catalyst data covering the characteristic-performance correlation of various metal oxide catalysts were used for pre-training. The core structural parameters of the pre-trained model are fixed, and the model is further trained by adjusting the learning rate and the number of new trees, using the core feature set as input. The core structural parameters include the learning rate and the number of new trees.

7. The method according to claim 6, characterized in that, A machine learning framework integrating transfer learning, meta-learning, and Bayesian optimization is employed. The model is trained and hyperparameters optimized using the core feature set to construct a catalyst performance prediction model. This includes a meta-learning training process, specifically: A catalyst task generator is constructed to generate multiple simulated catalyst small-sample tasks, each task including a support set for rapid model fine-tuning and a query set for evaluation; The meta-parameters are updated by alternating between inner and outer loops. The inner loop quickly fine-tunes the model parameters and calculates the loss based on the support set, while the outer loop calculates the meta-loss and updates the meta-parameters based on the query set.

8. The method according to claim 7, characterized in that, The fine-tuning involves dividing the real catalyst small sample data into a support set and a query set, using the meta-parameters obtained from meta-training, and fine-tuning the model through an inner loop to adapt it to the real task.

9. The method according to claim 6, characterized in that, The model for predicting catalyst performance is constructed by using a machine learning framework that integrates transfer learning, meta-learning, and Bayesian optimization, and by training and optimizing the model using the core feature set. The model is further improved by using a Bayesian optimization algorithm to automatically optimize the hyperparameters in the transfer learning training process and / or the meta-learning training process with the model prediction accuracy as the optimization objective.

10. The method according to claim 9, characterized in that, Using a Bayesian optimization algorithm, with model prediction accuracy as the optimization objective, the automatic optimization of hyperparameters in the transfer learning training process and / or the meta-learning training process specifically includes: The coefficient of determination R of the prediction model 2 The core optimization objective is set as the forward prediction error root mean square error (RMSE) and used as the constraint condition. Construct a hyperparameter-performance probability model based on Gaussian processes; The expected improvement criterion is adopted as the sampling strategy, and the optimal hyperparameter combination is obtained by iterative search within the preset hyperparameter space. The obtained optimal hyperparameter combination is automatically configured into the corresponding training process.

Images