A method for credible attribution and recommendation of catalyst formulations under low prediction accuracy conditions

Abstract

The application discloses a catalyst formula credible attribution and recommendation method under low prediction accuracy condition, and belongs to the field of catalyst design and data-driven material development. In view of the problems that machine learning has limited prediction ability when training samples are scarce, and the existing global optimization recommendation optimal formula paradigm lacks landing credibility guarantee, the application comprises the following steps: constructing and managing a catalyst performance data set, training a prediction model in a non-leakage manner and determining the ability boundary of the prediction model; constructing a candidate formula set by grid enumeration of a design space; calculating the nearest neighbor distance of each candidate formula as an extrapolation risk measure, and dividing the support degree grades according to the distribution position; only outputting the recommendations of the sufficient grades supported by data, and giving the prediction interval and the nearest real sample evidence for each recommendation for rechecking. The application can give a recommendation result with credible attribution and recheckable without relying on high prediction accuracy, and is suitable for dehydrogenation catalysts such as propane dehydrogenation and other material formula screening.

Description

Technical Field

[0001] This invention belongs to the field of data-driven material design and catalyst development technology, specifically involving a reliable attribution and recommendation method for catalyst formulations under low prediction accuracy conditions, applicable to formulation screening scenarios for propane dehydrogenation catalysts and various heterogeneous catalytic materials. Background Technology

[0002] Propylene is a key raw material for bulk chemicals such as polypropylene, acrylonitrile, and propylene oxide. Propane dehydrogenation without oxygen, with its advantages of inexpensive raw materials and the potential to produce high-purity hydrogen as a byproduct, has become an important technological route for the high-value utilization of light hydrocarbon resources. Existing industrial dehydrogenation catalysts are mainly platinum-based and chromium-based systems. Platinum-based catalysts are expensive and prone to deactivation due to carbon deposition, while chromium-based catalysts suffer from high toxicity and environmental risks. Zinc-based catalysts possess moderate alkane activation capabilities and are a highly promising alternative system; however, their catalytic performance is highly dependent on the dispersion state of the active component, the properties of the support, and the preparation process, creating a high-dimensional, highly nonlinear formulation design space.

[0003] Currently, data-driven machine learning methods are widely used in intelligent catalyst design. They typically involve constructing positive performance prediction models, combining them with interpretability tools to mine feature importance, and then performing multi-objective optimization to obtain formulation synthesis schemes, forming a conventional design process of "prediction-interpretation-optimization." However, this existing technology system has significant technical shortcomings: On the one hand, in scenarios where training samples are scarce, the overall fitting ability of prediction models built without leakage is limited. Taking zinc-based propane dehydrogenation catalysts as an example, the performance prediction determination coefficient obtained through random cross-validation is only 0.4–0.45. Existing technologies generally determine the usability of feature attribution results based on the level of model prediction accuracy, either assuming the model has high accuracy or directly rejecting the attribution results when the model accuracy is low, lacking objective criteria that do not rely on accuracy values ​​and can independently verify the credibility of attribution. At the same time, if there is a problem of sample reuse in the training data, random data partitioning will artificially inflate the model's prediction accuracy, leading to misjudgment of the model's true capabilities.

[0004] On the other hand, even if the model's predictive ability is truly reliable, the optimal formula obtained using global optimization methods such as genetic algorithms mostly falls within the design space region where training data is sparse and prediction reliability is low. This is a distributional extrapolation result, which carries a high risk of practical application. Existing methods only output a single formula prediction result and do not quantify the data support and extrapolation risk of the recommended formula. This makes it difficult for researchers to effectively distinguish between reliable recommendations and exploratory recommendations, resulting in poor practicality. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a reliable attribution and recommendation method for catalyst formulations under conditions of low prediction accuracy. It aims to solve the technical problems of unreliable catalyst formulation attribution results and poor implementability of recommended solutions in scenarios with scarce training samples and limited capabilities of leak-free prediction models. Traditional techniques cannot independently determine attribution reliability without considering model accuracy, nor can they quantify the extrapolation risk of formulation recommendations. This results in unreliable attribution results and poor implementability of recommended solutions. This invention does not rely on model performance values ​​or assume high model accuracy. Through a complete logical system that includes model capability boundary confirmation, independent attribution reliability determination, hierarchical constraints on the data support of recommended formulations, and verifiable recommendation output, it can still output reliable and implementable catalyst formulation recommendations even with limited model prediction capabilities.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A reliable attribution and recommendation method for catalyst formulations under conditions of low prediction accuracy includes the following steps: Step S1: Collect and organize the raw data of catalyst performance, and construct a dataset with catalyst composition parameters, preparation process parameters, and reaction condition parameters as input features and catalytic performance indicators as output labels; train the prediction model using a leak-free training method, and determine and output the capability boundary of the prediction model for each catalytic performance indicator.

[0007] Step S2: Perform grid enumeration on the preset catalyst formulation design space to construct a batch of candidate formulation sets that cover all aspects.

[0008] Step S3: Calculate the nearest neighbor distance between each candidate recipe and the training set samples, and use this distance as a quantitative measure of the recipe extrapolation risk; combine the interval position of this distance in the nearest neighbor distance distribution within the training set, and divide all candidate recipes into two levels: those with sufficient data support and those with insufficient data support.

[0009] Step S4: Select candidate formulations with sufficient data support as recommended candidates. For each recommended formulation, output the corresponding performance prediction range and the most recent real sample evidence to provide a basis for formulation review and verification.

[0010] Furthermore, after completing model training and capability boundary confirmation in step S1, this invention calculates the feature importance ranking results by combining multiple cross-validation methods with full fitting. By comparing the compositional consistency and overlap stability of feature ranking under different calculation methods, independent judgment of attribution credibility is achieved. This judgment process does not rely on the model capability boundary value and is suitable for attribution verification scenarios of low-precision prediction models.

[0011] Furthermore, this invention employs two algorithms with different principles to cross-verify the feature attribution results, thereby further improving the accuracy and stability of the attribution results.

[0012] Compared with the prior art, the present invention has the following beneficial effects: First, it enables independent and reliable verification of attribution results for low-precision models. This invention relies on the stability of feature importance ranking across multiple scenarios to complete attribution discrimination, eliminating the dependence of attribution verification on the numerical value of model prediction accuracy. It can determine the attribution validity of low-performance prediction models and also perform secondary verification of the attribution results of conventional high-precision models, avoiding judgment bias caused by single-precision evaluation.

[0013] Secondly, it effectively reduces the experimental implementation risk of formulation recommendations. This invention quantifies the extrapolation risk of candidate formulations using nearest neighbor distance, performs hierarchical screening of formulation data support, and only outputs formulation schemes from regions with sufficient data support. It avoids using extrapolation results outside the design spatial distribution as recommended schemes, thereby improving the reliability and practicality of formulation recommendations from the source.

[0014] Third, it enables the recommendation results to be traceable and verifiable. This invention simultaneously outputs the performance prediction range and real sample comparison evidence for each recommended formula, making it easy for experimental personnel to intuitively determine the credibility and suitability of the formula, thus solving the technical shortcoming of traditional single prediction results that cannot be verified.

[0015] Fourth, it ensures the physical consistency of the technical solution. The feature attribution results and formulation recommendations of this invention are both consistent with the physical mechanism of catalytic reactions, and the results are logically consistent, effectively avoiding unreasonable recommendations caused by algorithm optimization.

[0016] Fifth, the technology has strong versatility. The method of this invention is not limited to a single catalytic system, but can be widely adapted to the formulation screening of various dehydrogenation catalysts and heterogeneous catalytic materials, and can be extended to various data-driven intelligent material design scenarios. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the method of the present invention, showing the overall implementation process of data governance, capability boundary confirmation, credible attribution judgment, candidate formulation support level classification, and recommendation delivery.

[0018] Figure 2 This is a schematic diagram showing the spatial distribution and reliability classification of candidate formulations. Detailed Implementation

[0019] The present invention will be further described below with reference to embodiments. These embodiments are for illustrative purposes only and do not constitute a limitation thereof.

[0020] Dataset construction and governance: The original data consisted of experimental data from the zinc-based catalyst for propane dehydrogenation. Each document covered in the survey was individually verified against its original text, and the original entries were extracted. These entries underwent a three-stage processing: source verification to eliminate entries with untraceable origins; standardization of physical definitions to eliminate unsteady-state records, converting space velocities to a uniform order of magnitude, identifying and eliminating duplicated filler values, and narrowing the operational boundaries, retaining only records with true target values; and selection of modeling samples, requiring that both the continuous operational variables and target values ​​used as features be true values. The resulting modeling dataset, used in this embodiment, represents the zinc main phase as the active component in each record.

[0021] This dataset contains 208 records, sourced from 39 independent publications, with operating temperatures ranging from 400 °C to 620 °C. The values ​​of the variables in the dataset are as follows: zinc loading mass fraction 0.06% to 40.0%, temperature mean 557.8 °C, median 550 °C, and space velocity 0.19 to 22.7 h⁻¹. -1 (Median 2.36), propane partial pressure 2.5% to 100% (median 10.0%); propane conversion 2.0% to 55.0% (mean 27.2%, median 28.0%), propylene selectivity 9.1% to 99.5% (mean 87.2%, median 91.8%).

[0022] The dataset is organized according to the active component main phase, support and preparation method. In terms of stoichiometry, four continuous operating variables are used as core features (zinc loading, temperature, space velocity and propane partial pressure). Pairwise interaction terms and square terms are constructed. The support and preparation method are categorized according to a set frequency threshold. Low-frequency categories are assigned to the baseline class, and finally a 40-dimensional feature space is formed.

[0023] Model training and capability boundary confirmation: This embodiment uses a gradient boosting regression model to predict two targets: propane conversion rate and propylene selectivity. The model hyperparameters and the number of training epochs are fixed, early stopping is not used, and the random seed is fixed. Cross-validation is performed in two ways: random partitioning and partitioning by data source.

[0024] The model's capability boundary is characterized by the prediction results of randomized cross-validation: In this embodiment, under randomized partitioning, the coefficient of determination for conversion rate regression is 0.4462, the mean absolute error is 6.99 percentage points, and the standard deviation of the residuals is 9.46; the coefficient of determination for selective regression is 0.4455, the mean absolute error is 5.85 percentage points, and the standard deviation of the residuals is 9.75; under the condition of grouping by data source, the coefficient of determination for conversion rate regression is 0.2034, and the coefficient of determination for selective regression is −0.1547. Variance decomposition shows that the learnable information of the model mainly comes from the differences between data source groups, and the learnable signal within groups is limited. The above capability boundary indicates that the model's capability is limited under the training sample conditions; therefore, it is recommended to deliver the model in an "interval, hierarchical, evidence-based" manner, rather than delivering it as a single prediction point.

[0025] Credible attribution judgment: For each target, feature importance is calculated in three ways (full fit, random partition cross-validation, and source group cross-validation), and the composition and overlap of the top features in each ranking are compared.

[0026] In this embodiment, the pairwise overlap of the top 10 features in the three ranking groups of zinc-based propane dehydrogenation conversion rate and propylene selectivity was mostly 9 / 10 (range 8 / 10 to 10 / 10), significantly higher than the random expectation; the dominant feature remained consistently at the top in all three methods, with only the weaker features at the bottom showing slight fluctuations. Based on this, the above attribution results are considered reliable.

[0027] Furthermore, a different algorithm (random forest) was used for cross-validation: the rank correlation of the full feature ranking of the two algorithms was 0.864 (conversion rate) and 0.885 (selectivity); the overlap of the top few feature sets of each algorithm was higher than the random expectation. The two methods with different principles obtained consistent attribution results, indicating that the attribution reflects a stable structure-activity relationship rather than the random bias of a single algorithm.

[0028] Physically, the conversion rate is mainly affected by the interaction between the number of zinc active sites and the reaction conditions, while the selectivity is mainly controlled by the reaction conditions. This attribution direction is consistent with the physical understanding of the dehydrogenation reaction mechanism.

[0029] Specifically, the top 10 characteristics of the conversion rate target in terms of importance across the three methods are as follows: Under the full-scale fitting method, from highest to lowest importance, the order is: Zn×temperature, space velocity×propane partial pressure, temperature×space velocity, propane partial pressure, Zn×propane partial pressure, Zn×space velocity, space velocity, zinc loading, carrier SiO2, temperature; Under the randomized cross-validation method, the order is: Zn×temperature, Zn×propane partial pressure, space velocity×propane partial pressure, Zn×space velocity, temperature×space velocity, carrier SiO2, propane partial pressure, space velocity, zinc loading, temperature×propane partial pressure; Under the source-grouped cross-validation method, the order is: Zn×propane partial pressure, Zn×temperature, space velocity×propane partial pressure, carrier SiO2, propane partial pressure, Zn×space velocity, zinc loading, space velocity, temperature×propane partial pressure, temperature×space velocity. The overlap of the top 10 features under the three methods is as follows: 9 features overlap between random partition cross-validation and full fit; 9 features overlap between source group cross-validation and full fit, and 10 features overlap with random partition cross-validation.

[0030] The top 10 most important selective targets across the three methods are as follows: Under full-fit testing, from highest to lowest importance: space velocity × propane partial pressure, space velocity, temperature × space velocity, temperature × propane partial pressure, Zn × space velocity, Zn × temperature, Zn × propane partial pressure, zinc loading, zinc square, temperature; Under randomized cross-validation, the order is: space velocity, temperature × space velocity, Zn × space velocity, space velocity × propane partial pressure, temperature × propane partial pressure, zinc loading, Zn × temperature, Zn × propane partial pressure, temperature, propane partial pressure; Under source-grouped cross-validation, the order is: space velocity, temperature × space velocity, space velocity × propane partial pressure, Zn × temperature, Zn × space velocity, temperature × propane partial pressure, zinc loading, Zn × propane partial pressure, propane partial pressure, carrier ZSM-5. The overlap of the top 10 features under the three methods is as follows: 9 features overlap between random partition cross-validation and full fit; 8 features overlap between source group cross-validation and full fit, and 9 features overlap with random partition cross-validation.

[0031] As can be seen from the above ranking results, the top 10 features of both conversion rate and selectivity overlap by 8 to 10 points in all three methods, which is significantly higher than the expected overlap when randomly selected (the expected overlap of the top 10 features in two random columns of 40-dimensional features is 2.5 points). Furthermore, the dominant feature consistently ranks high in all methods. Therefore, the attribution results in the design space described in this embodiment are reliable.

[0032] Furthermore, under the condition of full fit, the feature importance was recalculated using the bagging random forest algorithm, which has a different principle, to complete the cross-validation at the algorithm level. The rank correlation of the full feature ranking of the two algorithms (gradient boosting and bagging) for full fit was 0.864 (conversion rate) and 0.885 (selectivity). The top 6 important features overlapped by 5 (conversion rate) and 4 (selectivity) respectively, and the top 10 features overlapped by 8 (conversion rate) and 7 (selectivity) respectively. The consistent attribution ranking obtained by the two methods with different principles indicates that the attribution reflects a stable structure-activity relationship rather than the accidental result of a single algorithm. Together with the ranking stability between the above cross-validation methods, this provides numerical evidence for the credibility of the attribution results.

[0033] Candidate formulation space construction and support level classification: The values ​​of the four continuous operational variables are divided into several representative levels, and then Cartesian products are performed with several frequently occurring carrier and preparation combinations to construct a fully enumerable set of candidate formulations. In this embodiment, each of the four continuous variables is assigned 6 levels and multiplied with 19 carrier and preparation combinations, resulting in a total of 24,624 candidates, covering the main body of the design space.

[0034] This embodiment does not employ a global optimization approach using genetic algorithms: under conditions of limited capabilities and scarce data, global optimization would push solutions towards sparsely populated regions, where predictions are least reliable. Grid enumeration ensures complete coverage and quantifies the data support for each candidate point.

[0035] For each candidate recipe, its nearest neighbor distance to the complete records of continuous variables in the training set (183 records in this example) is calculated as a measure of extrapolation risk. The smaller the distance, the closer the candidate is to existing experimental data. The percentile of this distance in the nearest neighbor distance distribution within the training set is selected as a threshold to classify the candidate recipes. High support is defined as the nearest neighbor distance not exceeding the 75th percentile (threshold 0.228 in this embodiment). Such candidates are close to the real samples, and there are a total of 318 such candidates in this embodiment. The middle support is the nearest neighbor distance not greater than the 99th percentile (threshold 1.606 in this embodiment), located within the coverage area of ​​the data body, and there are a total of 21476 in this embodiment; The exploratory measure is the nearest neighbor distance greater than the 99th percentile, which is far from the training sample. In this embodiment, there are a total of 2830 such nearest neighbors.

[0036] Candidates whose predicted values ​​fall outside the reasonable physical range of 0% to 100% are downgraded to the exploratory category. In this embodiment, a total of 782 candidates were downgraded due to their predicted values ​​falling outside the reasonable physical range, and these have been included in the aforementioned exploratory category. The nearest neighbor distance is calculated based on continuous operational variables and does not include category features such as carrier or preparation. Category differences are presented separately when comparing with the nearest sample.

[0037] Recommended delivery: This method selects candidates only from high and medium support levels as recommendations. Each recommendation delivers: a dual-objective predicted value; a prediction interval, which is the residual band of the predicted value plus or minus a set multiple σ, where σ is the standard deviation of the cross-validation residuals (in this embodiment, the standard deviation of the 208 cross-validation residuals is 9.5 percentage points in the conversion direction and 9.8 percentage points in the selectivity direction; when ±1.96σ is taken, i.e., approximately 95% coverage, the interval width is approximately ±18.5 percentage points in the conversion direction and ±19.1 percentage points in the selectivity direction); and the most recent real sample evidence, which traces back to the most recent real sample in the training set and its distance and performance for verification and comparison.

[0038] In this embodiment, the recommended objects meet the set thresholds of conversion rate not less than 38% and selectivity not less than 95%, and the ranking recommendation list is output according to the support level. There are a total of 172 candidates that meet the conditions, of which 134 are recommended in the high support and medium support levels, and 38 are not recommended in the exploratory level.

[0039] Example 1: The carrier was TS-1, the impregnation method was used, the zinc loading was 4.12%, the temperature was 600 °C, and the space velocity was 0.46 h⁻¹. -1 The propane partial pressure is 5%. The predicted conversion rate is 44.3%, the selectivity is 98.7%, and the nearest neighbor distance is 0.181, which is considered high. The nearest real sample is a sample with a zinc loading of 3%, dealuminated β-zeolite support, 600 °C, a measured conversion rate of 37.9%, a selectivity of 95.6%, and a distance of 0.181. The prediction is consistent with the nearest real sample.

[0040] Example 2: The carrier was Beta-BEA, the impregnation method was used, the zinc loading was 4.12%, the temperature was 600 °C, and the space velocity was 0.46 h⁻¹. -1 The propane partial pressure was 5%. The predicted conversion rate was 39.0%, the selectivity was 99.3%, and the nearest neighbor distance was 0.181, placing it in the high confidence range. This is anchored to samples near the same operating conditions as Example 1. Examples 1 and 2 are both in the highest confidence range, with conversion and selectivity values ​​respectively, allowing researchers to choose based on their objectives. Furthermore, under the same operating conditions as Examples 1 and 2, the candidate predicted conversion rate of the impregnation method using Al2O3 as a support was 42.2%, the selectivity was 98.9%, and the nearest neighbor distance was 0.18, also placing it in the high confidence range.

[0041] Example 3: The following candidate vectors have relatively high predicted conversion rates, but their support is only moderate and involves more extrapolation. Experimental data for the adjacent operating range of this vector should be supplemented before further experiments. For example: ZSM-5 vector (impregnation, zinc 12.28%, 520 °C, space velocity 1.90 h⁻¹). -1The predicted conversion was 46.4% (propane partial pressure 5%), with a selectivity of 98.2% and a distance of 0.39. The support was Silicalite-1 (hydrothermal support with impregnation and doping, zinc 6.84%), at 540 °C and a space velocity of 0.46 h⁻¹. -1 The predicted conversion rate was 46.7% (propane partial pressure 5%), selectivity was 96.5%, distance was 0.41, and medium support was used; MFI-zeolite support (hydrothermal method, zinc 6.84%, 560 ℃, space velocity 0.46 h⁻¹) was used. -1 (Propane partial pressure 5%) Predicted conversion rate 47.4%, selectivity 96.4%, distance 0.41, medium support.

[0042] Example 4: The recommendations output by this method are arranged in tiers according to their support level. High-support candidates that can be used directly, medium-support candidates that require supplementary data, and exploratory candidates that are only used for directional reference are listed separately, so that experimenters can arrange experiments accordingly. In this example, high-support and most dual-excellent candidates are concentrated in the low space velocity and low propane partial pressure range, which is consistent with the physical common sense that reducing space velocity prolongs contact and reducing partial pressure suppresses side reactions to improve conversion and selectivity.

[0043] In the above embodiments, each sub-carrier is recommended to carry a ±1.96σ residual band. Only the high and medium support levels are listed, and the exploratory level is not recommended.

[0044] The above description is merely an embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A reliable attribution and recommendation method for catalyst formulations under conditions of low prediction accuracy, characterized in that, Includes the following steps: Step S1: Acquire and process catalyst performance data, construct a dataset with catalyst composition parameters, preparation process parameters, and reaction condition parameters as features and catalytic performance indicators as labels, train a prediction model in a leak-free manner, and confirm and output the capability boundary of the prediction model for each performance indicator. Step S2: Perform grid enumeration on the preset catalyst formulation design space to construct a fully covered candidate formulation set; Step S3: For each candidate recipe, calculate its nearest neighbor distance with the training set samples and use it as an extrapolation risk measure. Based on the position of this distance in the nearest neighbor distance distribution within the training set, classify the candidate recipes into two levels: one with sufficient data support and one with insufficient data support. Step S4: Select recommended candidates only from the candidate formulations with sufficient data support, and for each recommendation result, output the corresponding performance prediction range and the most recent real sample evidence.

2. The reliable attribution and recommendation method for catalyst formulation under low prediction accuracy conditions as described in claim 1, characterized in that, After model training and capability boundary confirmation are completed in step S1, the attribution credibility determination step is also included: Feature importance is calculated using various cross-validation methods and full-fitting methods. The consistency and overlap stability of feature importance ranking under different calculation methods are compared. The reliability of feature attribution results is determined based on whether the feature ranking overlap is higher than the random expectation and whether the dominant feature is stably reproduced. The determination process is executed independently and does not depend on the capability boundary value of the prediction model.

3. The reliable attribution and recommendation method for catalyst formulation under low prediction accuracy conditions as described in claim 2, characterized in that: Two machine learning algorithms with different principles were used to calculate feature importance, and the consistency of feature ranking output by different algorithms was compared. The attribution results of multiple algorithms were cross-validated to confirm that the final attribution result can reflect the true correlation between catalyst formulation and performance.

4. The reliable attribution and recommendation method for catalyst formulation under low prediction accuracy conditions as described in claim 1, characterized in that, The leak-free training method described in step S1 includes: The model training rounds are fixed, the early stopping mechanism is removed, and two verification methods are adopted: random partitioning cross-validation and grouping cross-validation according to data source. The out-of-sample prediction results are used to characterize the model's capability boundary, and this capability boundary is used as the upper limit of the reliability of subsequent formulation recommendations.

5. The reliable attribution and recommendation method for catalyst formulation under low prediction accuracy conditions as described in claim 1, characterized in that: Step S2 uses a grid enumeration method to traverse the design space, achieving complete coverage of regions with sufficient data support under the condition that the model capability boundary is known and limited, thus avoiding the omission of effective candidate formulations.

6. The reliable attribution and recommendation method for catalyst formulation under low prediction accuracy conditions as described in claim 1, characterized in that, The classification method for step S3 includes: Based on the percentile of the nearest neighbor distance of the candidate recipe in the distance distribution within the training set, a grading threshold is set, and all candidate recipes are divided into three levels: high support, medium support, and exploratory. At the same time, candidate recipes whose predicted values ​​exceed the reasonable physical range are downgraded to the exploratory level. Recipe recommendations only select candidate recipes in the high support and medium support levels.

7. The reliable attribution and recommendation method for catalyst formulation under low prediction accuracy conditions as described in claim 1, characterized in that: The prediction interval mentioned in step S4 is the interval formed by superimposing the predicted values ​​with the standard deviation of the cross-validation residuals by a set multiple; The evidence of the most recent real sample is the most recent real sample matched by each candidate formulation in the training set and the corresponding distance. The distance is calculated only based on continuous operational variables to eliminate interference from category features.

8. The reliable attribution and recommendation method for catalyst formulation under low prediction accuracy conditions as described in claim 1, characterized in that: The catalytic performance indicators include the catalyst conversion rate and propylene selectivity of the propane dehydrogenation reaction; the prediction model adopts a gradient boosting regression model; the raw data is used for model construction after source verification, physical standardization, and modeling sample screening and preprocessing; for candidate formulations whose performance indicators meet the set thresholds, a sorted recommendation list is output according to the data support level.

9. The reliable attribution and recommendation method for catalyst formulation under low prediction accuracy conditions as described in claim 1, characterized in that: The model capability boundary obtained in step S1 serves as a preliminary criterion for the reliability of subsequent attribution analysis and formulation recommendation. The true capability of the model is quantified by fixing the number of training rounds, canceling early stopping, and using dual-mode cross-validation for verification. This effectively identifies and corrects the problem of overestimation of model capability caused by the reuse of data source information.

Images