An online evaluation method and system for new materials based on artificial intelligence
By constructing a unified embedding representation and introducing uncertainty, combined with an incremental training mechanism, the problems of long cycle and decreased reliability in traditional new material evaluation methods are solved, and the stability and credibility of online evaluation of new materials are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- LONGYAN UNIV
- Filing Date
- 2026-04-17
- Publication Date
- 2026-06-26
AI Technical Summary
Traditional methods for evaluating the performance of new materials rely on offline experimental testing, which is time-consuming and difficult to achieve in real time. Furthermore, the reliability of prediction results decreases when production conditions change, making it difficult to adapt to dynamic changes.
By acquiring raw material information, real-time online process data, and environmental and equipment metadata, a unified embedded representation is constructed. Performance prediction is performed using an integrated sub-model, and an uncertainty quantity and applicable domain determination mechanism are introduced. Combined with an incremental training mechanism, model parameters are updated to achieve online evaluation.
It improves the stability and reliability of new material evaluation, maintains stable evaluation capabilities under complex production conditions, adapts to changes in data distribution, and achieves continuity and reliability.
Smart Images
Figure CN122067660B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of intelligent material evaluation technology, specifically relating to a new online material evaluation method and system based on artificial intelligence. Background Technology
[0002] With the development of new materials technologies, materials research and development and production processes are increasingly characterized by large data scales, multiple parameter dimensions, and dynamic changes in operating status. In materials research and development, process optimization, and quality control, it is often necessary to evaluate material performance to determine whether the material meets the requirements of the target application. Traditional methods for evaluating the performance of new materials mainly rely on experimental testing or offline analysis, which involve testing material samples and analyzing them using empirical models to obtain performance evaluation results. However, these methods typically require long experimental cycles and are difficult to adapt to the real-time changes required during the materials production process, making it difficult to achieve online evaluation of the performance of new materials.
[0003] With the development of computing technology, data-driven computational models are increasingly being applied to the field of material performance prediction and evaluation. For example, predictive models are constructed to model multi-source data such as material composition, process parameters, and environmental conditions, thereby predicting material properties. While existing methods can improve evaluation efficiency to some extent, some problems still exist in practical applications. For instance, when production conditions, material formulations, or equipment status change, the distribution of input data may deviate from the data range during model training, leading to a decrease in the reliability of model prediction results. Furthermore, in long-term online operating environments, models struggle to promptly identify changes in data distribution and adaptively update, thus affecting the stability of evaluation results. Summary of the Invention
[0004] This invention provides an online evaluation method and system for new materials based on artificial intelligence, which solves the technical problems in related technologies where the performance evaluation process of new materials relies on offline experimental testing, has a long evaluation cycle, and is difficult to conduct real-time evaluation during the production process.
[0005] This invention provides a novel online evaluation method for materials based on artificial intelligence, comprising the following steps:
[0006] Step 1: Obtain the original material information, online process real-time data, and environmental and equipment metadata of the new material sample to be evaluated, form a material description vector, a process state vector, and a context metadata vector, and obtain the input tensor through deterministic input encoding;
[0007] Step 2: Input the input tensor into the material branch encoder, process branch encoder, and metadata branch encoder to obtain the corresponding embedding vectors, and then use the fusion unit to obtain a unified embedding vector;
[0008] Step 3: Input the unified embedding vector into the ensemble sub-model to obtain the performance point prediction value and calculate the cognitive uncertainty; calculate the application domain distance based on the application domain statistics to obtain the reliability vector;
[0009] Step 4: Compare the cognitive uncertainty quantity with the cognitive uncertainty threshold, and compare the applicable domain distance with the applicable domain distance threshold to obtain the gating state and determine the gating failure type;
[0010] Step 5: Within the statistical sliding window, calculate the gating failure rate based on the samples whose gating status is failed, and compare it with the drift threshold to obtain the drift determination result;
[0011] Step 6: When the drift determination result is that drift has occurred, calculate the sampling priority value based on the cognitive uncertainty, cognitive uncertainty threshold, applicable domain distance and applicable domain distance threshold, form a supplementary testing queue within the supplementary testing budget, and perform supplementary testing on the supplementary testing queue to obtain a newly labeled sample set;
[0012] Step 7: Construct an incremental training set from the newly labeled sample set and the replay buffer sample set, update the ensemble sub-model parameters and the applicable domain statistics, and then send the updated ensemble sub-model parameters and the updated applicable domain statistics back to Step 3.
[0013] This invention provides a novel online evaluation system for materials based on artificial intelligence, comprising:
[0014] The input tensor module is used to obtain the original material information, online process real-time data and environmental and equipment metadata of the new material sample to be evaluated, form a material description vector, a process state vector and a context metadata vector, and obtain the input tensor through deterministic input encoding;
[0015] The unified embedding module is used to input the input tensor into the material branch encoder, process branch encoder and metadata branch encoder to obtain the corresponding embedding vector, and then obtain the unified embedding vector through the fusion unit;
[0016] The reliability assessment module is used to input the unified embedding vector into the ensemble sub-model to obtain the performance point prediction value and calculate the cognitive uncertainty; it also calculates the application domain distance based on the application domain statistics to obtain the reliability vector.
[0017] The gating determination module is used to compare the cognitive uncertainty quantity with the cognitive uncertainty threshold, and the applicable domain distance with the applicable domain distance threshold to obtain the gating status and determine the gating failure type.
[0018] The drift determination module is used to calculate the gating failure rate of samples with failed gating states within a statistical sliding window and compare it with the drift threshold to obtain the drift determination result.
[0019] The supplementary sampling module is used to calculate the sampling priority value based on the cognitive uncertainty, cognitive uncertainty threshold, applicable domain distance, and applicable domain distance threshold when the drift determination result is that drift has occurred. It forms a supplementary testing queue within the supplementary testing budget and performs supplementary testing on the supplementary testing queue to obtain a newly labeled sample set.
[0020] The incremental update module is used to construct an incremental training set by adding the newly labeled sample set and the replay buffer sample set, update the parameters of the ensemble sub-model and update the applicable domain statistics, and then send the updated ensemble sub-model parameters and updated applicable domain statistics back to step 3.
[0021] The beneficial effects of this invention are as follows: By uniformly encoding and fusing raw material information, real-time online process data, and environmental and equipment metadata, a unified embedded representation is constructed. This representation is then used to predict material properties using an integrated sub-model, thereby enabling online evaluation of new materials during production or experimentation. Compared to related technologies that rely on offline detection or single-model prediction, this invention introduces uncertainty measurement and applicability domain determination mechanisms during the evaluation process. By jointly analyzing the stability of the prediction results and the sample distribution location, the reliability of the evaluation results is quantitatively judged, improving the stability and credibility of the evaluation results. Simultaneously, this invention continuously monitors the gating failure rate through a statistical sliding window, automatically identifying changes in data distribution. When drift is detected, samples are selected for supplementary testing and labeling based on a priority strategy. Combined with an incremental training mechanism, model parameters and applicability domain statistics are updated, allowing the evaluation model to gradually adapt to new data distribution environments. Through these technical solutions, the online evaluation system for new materials can maintain stable evaluation capabilities even under complex production conditions, improving the continuity and reliability of the new material evaluation process. Attached Figure Description
[0022] Figure 1 This is a flowchart of a novel online evaluation method for materials based on artificial intelligence, according to the present invention. Detailed Implementation
[0023] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.
[0024] like Figure 1 As shown, a novel online evaluation method for materials based on artificial intelligence includes the following steps:
[0025] Step 1: Obtain the original material information, online process real-time data, and environmental and equipment metadata of the new material sample to be evaluated, form a material description vector, a process state vector, and a context metadata vector, and obtain the input tensor through deterministic input encoding;
[0026] Step 2: Input the input tensor into the material branch encoder, process branch encoder, and metadata branch encoder to obtain the corresponding embedding vectors, and then use the fusion unit to obtain a unified embedding vector;
[0027] Step 3: Input the unified embedding vector into the ensemble sub-model to obtain the performance point prediction value and calculate the cognitive uncertainty; calculate the application domain distance based on the application domain statistics to obtain the reliability vector;
[0028] Step 4: Compare the cognitive uncertainty quantity with the cognitive uncertainty threshold, and compare the applicable domain distance with the applicable domain distance threshold to obtain the gating state and determine the gating failure type;
[0029] Step 5: Within the statistical sliding window, calculate the gating failure rate based on the samples whose gating status is failed, and compare it with the drift threshold to obtain the drift determination result;
[0030] Step 6: When the drift determination result is that drift has occurred, calculate the sampling priority value based on the cognitive uncertainty, cognitive uncertainty threshold, applicable domain distance and applicable domain distance threshold, form a supplementary testing queue within the supplementary testing budget, and perform supplementary testing on the supplementary testing queue to obtain a newly labeled sample set;
[0031] Step 7: Construct an incremental training set from the newly labeled sample set and the replay buffer sample set, update the ensemble sub-model parameters and the applicable domain statistics, and then send the updated ensemble sub-model parameters and the updated applicable domain statistics back to Step 3.
[0032] In one embodiment of the present invention, the original material information, online process real-time data, and environmental and equipment metadata of the new material sample to be evaluated are obtained to form a material description vector, a process state vector, and a context metadata vector, and an input tensor is obtained through deterministic input encoding, including:
[0033] Step 11: Obtain the original material information, online process real-time data, and environmental and equipment metadata of the new material sample to be evaluated. The original material information describes the basic composition and properties of the material, such as material composition parameters, structural properties, or proportioning information. The online process real-time data describes the real-time operating status of the material during production or experimentation, such as equipment operating status, process parameters, or process monitoring data. The environmental and equipment metadata describes the environmental conditions and equipment information during material production or testing, such as equipment identification, operating environment, or equipment configuration parameters. To ensure consistency of multi-source data across time, this embodiment determines the online process real-time data sampling timestamp that triggers the online evaluation as the evaluation time. When multiple records exist for the online process real-time data, the system calculates the time difference between the timestamp of each record and the evaluation time, and selects the record with the smallest time difference as the online process real-time data record corresponding to the evaluation time. Subsequently, using the same new material sample to be evaluated and the evaluation time as a unified index, the original material information, online process real-time data, and environmental and equipment metadata are aligned. For fields still missing after alignment, a missing marker is written in the corresponding position. The missing markers mentioned above are used to indicate that the field does not provide valid data at the current evaluation time. As an explicit data placeholder, it can be identified by the model and processed uniformly in the subsequent encoding stage.
[0034] Step 12: Based on the pre-fixed field order during the training phase, the aligned raw material information is converted into a material description vector, the real-time online process data is converted into a process state vector, and the environmental and equipment metadata is converted into a context metadata vector. These vectors represent material property features, process operation features, and environmental and equipment features, respectively. Specifically, the material description vector is the vector representation of the raw material information converted into a numerical sequence according to the preset field order; the process state vector is the vector representation of the real-time online process data converted in the same way; and the context metadata vector is the vector representation of the environmental and equipment metadata. To ensure consistency in the model input dimensions, this embodiment further performs dimension alignment processing on the three vectors based on the target dimensions fixed during the training phase. Specifically, when a vector contains missing fields, missing markers are filled in at the corresponding positions; when the vector length exceeds the target dimension, it is truncated from the end to the target dimension according to the field order. This method ensures that data from different sources have a unified vector dimension before entering the model, thereby avoiding abnormal model input due to inconsistent data dimensions.
[0035] Step 13: Perform deterministic input encoding on the material description vector, process state vector, and context metadata vector. This deterministic input encoding refers to using data encoding rules identical to those used in the training phase during the online evaluation phase to numerically transform the input data, ensuring consistent encoding results for the same input at different times. Specifically, scalar fields are standardized using the mean and standard deviation fixed during the training phase, converting them into standardized values with a uniform scale. Category fields are mapped using a mapping table fixed during the training phase, converting category identifiers into corresponding numerical codes. This mapping table is a data dictionary established and fixed during the model training phase based on historical data, used to ensure that the encoding rules used in the online evaluation phase are completely consistent with those used in the training phase. After encoding, the material description vector, process state vector, and context metadata vector are integrated according to a pre-fixed field order to form a unified input tensor. The input tensor is a multi-dimensional data structure that can be directly processed by the artificial intelligence model, containing comprehensive information from material properties, process states, and environmental context, and serves as input for subsequent embedded computation and performance prediction models. The artificial intelligence model is a data-driven computational model for realizing online evaluation of new materials. It includes, but is not limited to, a material branch encoder, a process branch encoder, a metadata branch encoder, a fusion unit, and an integrated sub-model. It is used to perform representation learning, fusion calculation, performance prediction, and reliability determination on material property information, process state information, and environmental context information.
[0036] Through the steps described above, this embodiment achieves a unified representation and structured encoding of multi-source heterogeneous data for new materials, enabling material property information, process operation information, and environmental equipment information to be input into the artificial intelligence computing model in a unified tensor form. This method ensures consistency in data representation between the online evaluation and model training phases through deterministic field order, dimensional control, and encoding rules, thereby improving the stability and repeatability of the model's input data. In AI-based online evaluation scenarios for new materials, this method effectively addresses the problems of large differences in multi-source data structures, difficulties in time alignment, and inconsistent input data dimensions.
[0037] In one embodiment of the present invention, the input tensor is used to input the material branch encoder, process branch encoder, and metadata branch encoder to obtain corresponding embedding vectors, and then a unified embedding vector is obtained through a fusion processor, including:
[0038] Step 21: Read the input tensor and perform a consistency check on the input tensor based on the field order and target dimension fixed in the same version as the material branch encoder, process branch encoder, and metadata branch encoder. The field order refers to the arrangement of data fields determined and fixed during the training phase, ensuring that the data arrangement in the online evaluation phase is completely consistent with the model training phase; the target dimension refers to the fixed length of the model input vector set during the training phase. Since artificial intelligence models typically rely on stable input dimensions and field arrangements for computation, consistency verification of the input tensor is necessary during the online evaluation phase. When the verification result shows that the field order and target dimension of the input tensor are consistent with the input interfaces of the three types of branch encoders, the system determines that the input tensor is the input data object for subsequent calculations; if any condition is inconsistent, subsequent calculations are terminated and an input consistency error flag is returned to avoid inconsistent inputs causing distortion of the model calculation results.
[0039] Step 22: Input the input tensor into the material branch encoder, process branch encoder, and metadata branch encoder respectively to obtain the material branch embedding vector, process branch embedding vector, and metadata branch embedding vector. The branch encoder refers to an artificial intelligence feature extraction model constructed for different types of data features. It transforms the original input vector into a more semantically expressive embedding vector by performing multi-layer feature mapping on the input data. In this embodiment, the branch encoder adopts a neural network model structure, consisting of a feature encoding network, an autoencoder network, or other deep neural network structures used for feature representation learning, composed of multi-layer fully connected networks. The above model extracts latent features from the input data step by step through multi-layer nonlinear mapping, thereby generating embedding vectors that can represent material property information, process state information, or environmental context information. Specifically, the material branch encoder is used to extract the feature representation corresponding to the original material information, the process branch encoder is used to extract dynamic state features from real-time online process data, and the metadata branch encoder is used to extract context information features from environmental and equipment metadata. The encoders have been trained and their parameters fixed during the model training phase, and their output is the corresponding embedding vector. Subsequently, the system verifies the fixed dimensions of the material branch embedding vector, process branch embedding vector, and metadata branch embedding vector based on the input interface fixed in the same version as the fusion processor. When the dimensions of all three types of embedding vectors are consistent with the requirements of the fusion processor input interface, the fusion calculation stage is allowed; if the dimensions of any embedding vector are inconsistent, the fusion processor calculation is terminated, thereby ensuring the consistency of the model structure.
[0040] Step 23: Following the input order fixed during the training phase, the material branch embedding vector, process branch embedding vector, and metadata branch embedding vector are sequentially concatenated to obtain the fused input vector. This concatenation refers to connecting multiple embedding vectors in a fixed order along the vector dimension into a longer vector to preserve the feature information of each branch. Subsequently, the fused input vector is input into the fusion processor, and a multi-layer nonlinear transformation is performed on the fused input vector through the pre-trained computation graph of the fusion processor to obtain a unified embedding vector. The fusion processor is an artificial intelligence computational model for multi-source feature fusion, which contains a multi-layer nonlinear transformation structure to learn the correlation between different information sources in a high-dimensional feature space. In this embodiment, the fusion processor adopts a neural network fusion model, such as a feature fusion network composed of multi-layer fully connected networks, a feature fusion network based on an attention mechanism, or a feature fusion network based on a gating structure. Through the above fusion structure, the material branch embedding vector, process branch embedding vector, and metadata branch embedding vector are jointly mapped, enabling features from different sources to form a unified representation in the same feature space, thereby obtaining a unified embedding vector. A pre-trained computational graph refers to a computational structure that has been determined and solidified during the model training phase. It includes computational nodes at each layer and their connections, and remains unchanged during the online evaluation phase. Through multi-layer nonlinear transformations, the fusion processor can comprehensively represent features from different branches, thereby obtaining a unified embedding vector. The unified embedding vector is a high-level semantic vector that comprehensively represents the attribute features, process state features, and environmental context features of a new material sample. It can be used as a unified input in subsequent artificial intelligence models.
[0041] Through the above steps, this embodiment constructs a multi-branch feature extraction and fusion mechanism based on an artificial intelligence computing model, enabling the representation and fusion of original material information, real-time online process data, and environmental and equipment metadata in a unified feature space. By using a branch encoder to learn independent features from different data sources and a fusion processor to perform multi-source feature fusion, the artificial intelligence model can simultaneously model the relationships between material property information, production process status information, and environmental context information in a unified feature space. This improves the accuracy and stability of online evaluation of new materials and allows the artificial intelligence model to operate stably in an online environment.
[0042] In one embodiment of the present invention, a unified embedding vector is input into an ensemble sub-model to obtain performance point prediction values, and a cognitive uncertainty quantity is calculated; based on the application domain statistics, the application domain distance is calculated to obtain a reliability vector, including:
[0043] Step 31: Read the unified embedding vector obtained in step 23 and perform a dimension consistency check. Since the ensemble sub-model has fixed the dimensional structure of the input interface during the training phase, it is necessary to check the fixed dimensions of the unified embedding vector during the online evaluation phase. When the dimensions of the unified embedding vector are consistent with the input interface of the ensemble sub-model, the system determines that the unified embedding vector as the input data object for subsequent calculations; when the dimensions are inconsistent, a preset verification failure flag is written and subsequent calculations are terminated to avoid abnormal inputs causing distortion of the model calculation results.
[0044] Step 32: Input the unified embedding vector into the ensemble sub-model to obtain the performance prediction result. The ensemble sub-model refers to an artificial intelligence prediction model set composed of multiple sub-models. It improves the stability and robustness of the prediction results by performing parallel predictions on the same input data and statistically integrating the prediction results. In this embodiment, the ensemble sub-model contains a preset number of sub-models, each of which performs prediction calculations based on model parameters obtained during the training phase. In this embodiment, the sub-models adopt a neural network prediction model structure, such as a regression prediction network composed of multi-layer fully connected networks, a deep neural network model, or other artificial intelligence model structures used for numerical prediction. Different sub-models can use the same model structure but are trained with different training initialization parameters or different data subsets, or they can use prediction models with different structures, thus forming model diversity. In this way, multiple sub-models can predict the performance of the unified embedding vector from different perspectives and generate multiple performance prediction values.
[0045] The system inputs a unified embedding vector into each sub-model to obtain multiple performance prediction values. Then, it sums these prediction values to obtain a total prediction, and divides this total prediction by the number of sub-models, i.e., the preset number, to obtain the performance point prediction value. The performance point prediction value refers to the comprehensive prediction result of the integrated sub-model on the performance of the new material sample to be evaluated. It is obtained by averaging the outputs of multiple sub-models and is used to characterize the material's performance estimate under current process conditions.
[0046] Step 33: After obtaining multiple performance prediction values and performance point prediction values, the system further calculates the cognitive uncertainty. The cognitive uncertainty is used to characterize the degree of uncertainty in the model's prediction results, reflecting the dispersion of prediction results among different sub-models. Specifically, the difference between each performance prediction value and the performance point prediction value is calculated, and the difference is squared to obtain the squared difference; then, all squared differences are summed, and the summation result is divided by a preset quantity to obtain the cognitive uncertainty. The larger this value, the greater the difference between the prediction results of different sub-models, and the higher the uncertainty of the model's prediction for that sample.
[0047] After obtaining the cognitive uncertainty, the system further calculates the applicability domain distance based on the applicability domain statistics. The applicability domain statistics are a set of statistical parameters obtained during the model training phase based on the unified embedding vectors of the training samples. They are used to describe the distribution characteristics of the model training data in the embedding space. In this embodiment, the applicability domain statistics include the embedding space mean and the covariance with a regularization term. The embedding space mean represents the average position of the unified embedding vectors of the training samples in the feature space, and the covariance with a regularization term describes the correlation between each feature dimension and ensures the stability of the covariance matrix during calculation. The system first reads the applicability domain statistics and verifies that its version is consistent with the ensemble submodel version to ensure that the statistical benchmark is consistent with the model training data. Subsequently, by subtracting the embedding space mean from the unified embedding vector to obtain the difference vector, and then calculating its inverse matrix based on the covariance with a regularization term, the difference vector and the inverse matrix are multiplied to obtain the intermediate vector. Next, the intermediate vector and the difference vector are multiplied to obtain the squared distance value, and the square root operation is performed on the squared distance value to obtain the applicability domain distance. The calculation process of the applicable domain distance is equivalent to the Mahalanobis distance calculation method, which is used to characterize the degree of deviation of the current sample from the center of the training data distribution in the embedding space. The larger the value, the more likely the sample is to be outside the model training data distribution.
[0048] After obtaining the cognitive uncertainty and the applicable domain distance, the system integrates them in a pre-defined order to generate a reliability vector. This reliability vector is a vector structure composed of multiple reliability metrics, used to comprehensively characterize the credibility of the AI model's predictions for the current new material samples. The cognitive uncertainty reflects the consistency of predictions within the model, while the applicable domain distance reflects the degree of difference between the sample and training data distributions; together, they constitute a multi-dimensional description of the model's predictive reliability.
[0049] Through the above steps, this embodiment constructs a reliability assessment mechanism that combines performance prediction, uncertainty estimation, and applicability domain determination. This enables the online evaluation of new materials based on artificial intelligence models to not only output material performance prediction results but also simultaneously output the reliability information of the model predictions. Compared with traditional evaluation methods that rely solely on a single prediction value, this invention calculates cognitive uncertainty through integrated sub-models and calculates the applicability domain distance through applicability domain statistics. This quantifies the reliability of the prediction results from two different dimensions, thereby improving the ability of the AI-based online evaluation method for new materials to identify anomalous and out-of-distribution samples, and enabling the system to maintain stable evaluation performance even in complex industrial environments.
[0050] In one embodiment of the present invention, the cognitive uncertainty quantity is compared with the cognitive uncertainty threshold, and the applicable domain distance is compared with the applicable domain distance threshold to obtain the gating state, and the gating failure type is determined, including:
[0051] Step 41: Read the cognitive uncertainty, applicable domain distance, cognitive uncertainty threshold, and applicable domain distance threshold, and verify the validity of these parameters. The cognitive uncertainty threshold and applicable domain distance threshold are reference thresholds determined during the model training or system configuration phases, used to determine whether the model prediction results are within an acceptable range. The system first verifies that all four parameters exist and are comparable values. If any parameter is missing or cannot be compared numerically, the subsequent calculation process is terminated, gating is not performed, and no gating state or gating failure type is generated, thereby avoiding erroneous evaluation results under abnormal parameter conditions.
[0052] Step 42: After parameter verification, the system compares the cognitive uncertainty quantity with the cognitive uncertainty threshold, and also compares the applicable domain distance with the applicable domain distance threshold. If the cognitive uncertainty quantity does not exceed the cognitive uncertainty threshold, and the applicable domain distance does not exceed the applicable domain distance threshold, the gating state is determined to be passed; otherwise, the gating state is determined to be failed. The gating state refers to the judgment result of whether the prediction result of the artificial intelligence model meets the reliability conditions, which is obtained by jointly judging the uncertainty index and the applicable domain index. When the gating state is passed, it means that the current new material sample is within the distribution range of the model training data, and the model's prediction of the sample has low uncertainty. Therefore, the prediction result can be used as a valid evaluation result to continue to participate in subsequent processing.
[0053] Step 43: When the gating state fails, the system further determines the gating failure type. Specifically, when the applicable domain distance exceeds the applicable domain distance threshold, only the gating failure type is determined to be "out of applicable domain". "Out of applicable domain" means that the current sample's position in the embedded feature space significantly deviates from the training data distribution range, indicating that the sample may belong to a material state or process condition not covered during model training. When the applicable domain distance does not exceed the applicable domain distance threshold and the cognitive uncertainty exceeds the cognitive uncertainty threshold, only the gating failure type is determined to be "cognitive uncertainty". "Cognitive uncertainty" means that although the current sample is within the model training data distribution range, different sub-models show significant differences in their prediction results for the sample, indicating that the model's prediction stability in this region is low.
[0054] Through the above steps, this embodiment constructs a gating decision mechanism based on the prediction results of an artificial intelligence model, enabling the system to perform dual reliability judgments on the prediction results based on uncertainty indicators and data distribution indicators. Compared with traditional methods that rely solely on a single predicted value for evaluation, this invention simultaneously introduces cognitive uncertainty and the applicable domain distance to conduct multi-dimensional evaluations of the credibility of the model's prediction results, thereby effectively identifying out-of-distribution samples and unstable prediction samples. This gating decision mechanism can prevent unreliable prediction results from directly participating in the evaluation process, thereby improving the stability and security of the online evaluation system in complex industrial environments, and enabling the artificial intelligence model to continuously maintain effective evaluation capabilities in dynamic data environments.
[0055] In one embodiment of the present invention, the gating failure rate is calculated based on samples with a failed gating state within a statistical sliding window, and compared with a drift threshold to obtain a drift determination result, including:
[0056] Step 51: Read the gating status and statistical sliding window, and use the current evaluation time as the window endpoint to determine the sample set within the window based on the statistical sliding window. The statistical sliding window refers to a continuous time interval used for statistical analysis in time series data, used to limit the range of samples participating in the statistical calculation; the current evaluation time refers to the time point corresponding to the online evaluation of the new material sample to be evaluated. After determining the window range, the system performs statistics on the sample set within the window to obtain the total number of samples within the window, and further counts the number of samples with a failed gating status within this set, thus obtaining the number of failed samples. A sample with a failed gating status indicates that the prediction result of that sample did not pass the reliability judgment.
[0057] Step 52: After obtaining the total number of samples within the window and the number of failed samples, the system verifies the validity of the statistical results. Specifically, the system first verifies that the total number of samples within the window is greater than zero and that the drift threshold is a comparable value. The drift threshold is a reference threshold used to determine whether the model's running state has changed, and it is preset during the system configuration or model training phase. When the above conditions are met, the number of failed samples is used as the numerator, and the total number of samples within the window is used as the denominator. The ratio between the numerator and denominator is calculated to obtain the gating failure rate. The gating failure rate represents the proportion of samples that fail the gating judgment within the statistical sliding window, and its value reflects the degree to which the reliability of the prediction results of the artificial intelligence model decreases within this time interval. When the total number of samples within the window is not greater than zero or the drift threshold is not comparable, the subsequent drift judgment process is terminated to avoid abnormal statistical data affecting the judgment results.
[0058] Step 53: After obtaining the gating failure rate, the system compares the gating failure rate with the drift threshold to determine the drift judgment result. When the gating failure rate is greater than or equal to the drift threshold, the drift judgment result is determined to be drift; when the gating failure rate is less than the drift threshold, the drift judgment result is determined to be no drift. Drift refers to the phenomenon that the distribution of input data or the applicable conditions of the model change relative to the training phase during the online evaluation of new materials. This change may be caused by factors such as changes in material formulation, adjustments to production processes, changes in equipment status, or changes in environmental conditions. When drift is detected, it indicates that the current artificial intelligence model's adaptability to the latest data may have decreased, requiring further subsequent retesting or model updates.
[0059] Through the above steps, this embodiment constructs a drift detection mechanism based on gating statistical results, enabling the online evaluation system for new materials to monitor the predictive stability of the artificial intelligence model in real time during continuous operation. By calculating the gating failure rate through a statistical sliding window and combining it with a drift threshold for automatic judgment, the system can promptly identify situations where the model's applicability declines when data distribution changes. Continuous monitoring of the trend in model predictive reliability ensures that the artificial intelligence model maintains stable evaluation capabilities in dynamic production environments.
[0060] In one embodiment of the present invention, when the drift determination result indicates that drift has occurred, a sampling priority value is calculated based on the cognitive uncertainty quantity, the cognitive uncertainty threshold, the applicable domain distance, and the applicable domain distance threshold. A supplementary testing queue is formed within the supplementary testing budget, and supplementary testing is performed on the supplementary testing queue to obtain a newly labeled sample set, including:
[0061] Step 61: Read the drift determination result and verify that drift has occurred. When the determination result is that drift has occurred, it indicates that the distribution of input data or the applicable conditions of the model in the current online evaluation system for new materials has changed compared to the training phase. Therefore, it is necessary to supplement new labeled data to improve the model's adaptability to the current data distribution. Subsequently, the system reads the restricted evaluation sample set, cognitive uncertainty, cognitive uncertainty threshold, applicable domain distance, applicable domain distance threshold, and supplementary testing budget. The restricted evaluation sample set refers to the set of samples that failed the gating determination during the online evaluation process. These samples are temporarily stored due to insufficient reliability of the prediction results and are used for subsequent supplementary testing selection. The supplementary testing budget refers to the maximum number of samples that the system allows to perform supplementary testing operations within one supplementary testing cycle, which is usually preset during the system configuration phase based on the detection equipment capability or experimental cost. The system then verifies that the restricted evaluation sample set is a non-empty set, and simultaneously verifies that the cognitive uncertainty threshold and applicable domain distance threshold are comparable values and that the supplementary testing budget is a positive integer. If any condition is not met, the subsequent sampling priority calculation and supplementary testing process are terminated to avoid performing supplementary testing operations when data conditions are insufficient.
[0062] Step 62: After completing the condition verification, the system calculates the sampling priority value based on the cognitive uncertainty and the distance to the applicable domain. Specifically, the system first reads the pre-set weighting coefficients. These weighting coefficients are used to adjust the contribution of the cognitive uncertainty index and the applicable domain index in the sampling priority calculation, and their values are usually predetermined during the system configuration phase. For each sample in the restricted evaluation sample set, the system first calculates the ratio between the cognitive uncertainty and the cognitive uncertainty threshold, and multiplies this ratio by the weighting coefficient to obtain the first weighted value; then it calculates the ratio between the applicable domain distance and the applicable domain distance threshold, and multiplies this ratio by the complementary value of the weighting coefficient to obtain the second weighted value; finally, it sums the first weighted value and the second weighted value to obtain the sampling priority value of the sample. In this way, the system can simultaneously consider two factors: model prediction uncertainty and data distribution deviation, thereby comprehensively evaluating the importance of the sample to model updates. Subsequently, the system sorts the samples in the restricted evaluation sample set from high to low according to the sampling priority value, and selects samples whose ranking does not exceed the supplementary testing budget to form a supplementary testing queue. When the number of samples in the restricted assessment sample set is less than the supplementary assessment budget, the supplementary assessment queue is directly formed from the restricted assessment sample set.
[0063] Step 63: After generating the supplementary testing queue, the system iterates through each sample in the queue and performs supplementary testing on each sample to obtain the corresponding annotation results. Supplementary testing refers to re-acquiring the sample's true performance data through experimental testing equipment or material testing procedures to obtain reliable annotation results. After obtaining the supplementary testing results, the system establishes a one-to-one correspondence between each sample in the supplementary testing queue and its corresponding annotation results, and aggregates the supplementary testing queue and its corresponding annotation results to form a newly annotated sample set. This newly annotated sample set is used for incremental training or model updates of the artificial intelligence model in subsequent steps, thereby enabling the model to gradually adapt to the new data distribution environment.
[0064] Through the above steps, this embodiment constructs an active sampling mechanism based on the reliability information of artificial intelligence prediction results. In the online evaluation method and system for new materials based on artificial intelligence, when a data distribution drift is detected, the system can calculate the sampling priority based on the cognitive uncertainty and the distance to the applicable domain, and prioritize the selection of the most informational samples from the limited evaluation sample set for supplementary testing. By guiding the sampling process with the prediction uncertainty information and data distribution information of the artificial intelligence model, the limited supplementary testing resources can be concentrated on the most critical samples, thereby improving the effectiveness of new training data for model updates.
[0065] In one embodiment of the present invention, an incremental training set is constructed by combining the newly labeled sample set and the replay buffer sample set, the ensemble sub-model parameters are updated, and the applicable domain statistics are updated. The updated ensemble sub-model parameters and the updated applicable domain statistics are then fed back to step 3, including:
[0066] Step 71: Read the newly labeled sample set and the replay buffer sample set, and perform a consistency check on the field structure of the two types of sample data. The replay buffer sample set refers to a portion of the historical training sample set retained during system operation. It is used to provide stable data references during model updates to avoid overfitting or forgetting historical knowledge when the model relies solely on the latest samples for training. The system checks the field structure of the two sample sets to ensure that they contain the same data fields and a consistent data arrangement structure. If the field structure of either set is inconsistent, the subsequent model update process is terminated; if the field structure is consistent, the newly labeled sample set and the replay buffer sample set are merged, and duplicate samples are deduplicated to obtain the incremental training set. The incremental training set is the training data set used for model parameter updates, which includes both newly acquired sample data and some historical sample data.
[0067] Step 72: Read the current ensemble sub-model parameters and incremental training configuration, and perform version consistency verification. The ensemble sub-model parameters refer to the set of model parameters constituting each sub-model in the ensemble sub-model, such as the weight parameters and bias parameters in a neural network. The incremental training configuration refers to a set of training parameters used to control the incremental training process of the model, which may include the number of training epochs, learning rate, or other training control parameters. After completing the version consistency verification, the system performs incremental training iterations on the ensemble sub-model parameters based on the incremental training set and according to the number of training epochs set in the incremental training configuration, thereby obtaining the updated ensemble sub-model parameters. Incremental training refers to continuing to train the model using newly added training data while retaining the original model parameters, enabling the model to gradually learn new data distribution characteristics.
[0068] Step 73: After updating the ensemble sub-model parameters, the system further updates the applicability domain statistics. The applicability domain statistics are a set of parameters describing the statistical distribution characteristics of training samples in the embedding feature space, typically including the embedding space mean and covariance with a regularization term. The system first reads the current applicability domain statistics and the unified embedding vector corresponding to the newly labeled sample set, and performs a consistency check on their dimensions. Then, based on the unified embedding vector corresponding to the newly labeled sample set, the system calculates a new mean vector and a covariance with a regularization term, and merges and updates them with the original applicability domain statistics to obtain the updated applicability domain statistics. After completing the above update, the updated ensemble sub-model parameters and the updated applicability domain statistics are sent back to step 3, serving as the basis for new model calculations in subsequent online evaluation processes.
[0069] Through the above steps, this embodiment constructs a model incremental update mechanism based on newly labeled data, enabling the online evaluation system for new materials to continuously absorb new sample information and dynamically update the parameters of the artificial intelligence model and the statistics of the applicable domain during continuous operation. By updating the parameters of the ensemble sub-model through incremental training and updating the statistics of the applicable domain through the embedding features of newly added samples, the artificial intelligence model can gradually adapt to changes in data distribution caused by changes in material formulation, process conditions, and environment during production, allowing the system to maintain accurate and reliable evaluation capabilities continuously in dynamic industrial environments.
[0070] This invention provides a novel online evaluation system for materials based on artificial intelligence, comprising:
[0071] The input tensor module is used to obtain the original material information, online process real-time data and environmental and equipment metadata of the new material sample to be evaluated, form a material description vector, a process state vector and a context metadata vector, and obtain the input tensor through deterministic input encoding;
[0072] The unified embedding module is used to input the input tensor into the material branch encoder, process branch encoder and metadata branch encoder to obtain the corresponding embedding vector, and then obtain the unified embedding vector through the fusion unit;
[0073] The reliability assessment module is used to input the unified embedding vector into the ensemble sub-model to obtain the performance point prediction value and calculate the cognitive uncertainty; it also calculates the application domain distance based on the application domain statistics to obtain the reliability vector.
[0074] The gating determination module is used to compare the cognitive uncertainty quantity with the cognitive uncertainty threshold, and the applicable domain distance with the applicable domain distance threshold to obtain the gating status and determine the gating failure type.
[0075] The drift determination module is used to calculate the gating failure rate of samples with failed gating states within a statistical sliding window and compare it with the drift threshold to obtain the drift determination result.
[0076] The supplementary sampling module is used to calculate the sampling priority value based on the cognitive uncertainty, cognitive uncertainty threshold, applicable domain distance, and applicable domain distance threshold when the drift determination result is that drift has occurred. It forms a supplementary testing queue within the supplementary testing budget and performs supplementary testing on the supplementary testing queue to obtain a newly labeled sample set.
[0077] The incremental update module is used to construct an incremental training set by adding the newly labeled sample set and the replay buffer sample set, update the parameters of the ensemble sub-model and update the applicable domain statistics, and then send the updated ensemble sub-model parameters and updated applicable domain statistics back to step 3.
[0078] It should be noted that the range and threshold size are set for ease of comparison. The size of the threshold depends on the amount of sample data and the number of bases set by those skilled in the art for each set of sample data, as long as it does not affect the ratio between the parameter and the quantized value.
[0079] The embodiments of the present invention have been described above, but the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of the present embodiments, all of which are within the protection scope of the present embodiments.
Claims
1. A novel online evaluation method for materials based on artificial intelligence, characterized in that, Includes the following steps: Step 1: Obtain the original material information, online process real-time data, and environmental and equipment metadata of the new material sample to be evaluated, form a material description vector, a process state vector, and a context metadata vector, and obtain the input tensor through deterministic input encoding; Step 2: Input the input tensor into the material branch encoder, process branch encoder, and metadata branch encoder to obtain the corresponding embedding vectors, and then use the fusion unit to obtain a unified embedding vector; Step 3: Input the unified embedding vector into the ensemble sub-model to obtain the performance point prediction value and calculate the cognitive uncertainty; calculate the application domain distance based on the application domain statistics to obtain the reliability vector; The cognitive uncertainty quantity is used to characterize the degree of uncertainty of the model prediction results, which reflects the degree of dispersion of the prediction results among different sub-models; the applicable domain statistics are a set of statistical parameters obtained by statistical analysis of the unified embedding vector of the training samples during the model training phase, which are used to describe the distribution characteristics of the model training data in the embedding space. Step 4: Compare the cognitive uncertainty quantity with the cognitive uncertainty threshold, and compare the applicable domain distance with the applicable domain distance threshold to obtain the gating state and determine the gating failure type; wherein, the gating state refers to the judgment result of whether the prediction result of the artificial intelligence model meets the reliability conditions; Step 5: Within the statistical sliding window, calculate the gating failure rate based on the samples whose gating status is failed, and compare it with the drift threshold to obtain the drift determination result; Step 6: When the drift determination result is that drift has occurred, calculate the sampling priority value based on the cognitive uncertainty, cognitive uncertainty threshold, applicable domain distance and applicable domain distance threshold, form a supplementary testing queue within the supplementary testing budget, and perform supplementary testing on the supplementary testing queue to obtain a newly labeled sample set; Step 7: Construct an incremental training set from the newly labeled sample set and the replay buffer sample set, update the ensemble sub-model parameters and the applicable domain statistics, and then send the updated ensemble sub-model parameters and the updated applicable domain statistics back to Step 3.
2. The online evaluation method for new materials based on artificial intelligence according to claim 1, characterized in that, The process involves acquiring raw material information, real-time online process data, and environmental and equipment metadata from the sample of the new material to be evaluated. This data forms a material description vector, a process state vector, and a context metadata vector. The input tensor is then obtained through deterministic input encoding, including: Step 11: Obtain the original material information, online process real-time data, and environmental and equipment metadata of the new material sample to be evaluated; determine the evaluation time as the online process real-time data sampling timestamp that triggers the online evaluation; when there are multiple records of online process real-time data, select the record with the smallest time difference from the evaluation time; align the three data with the same new material sample to be evaluated and the evaluation time, and write missing markers for missing fields. Step 12: Based on the field order fixed during the training phase, convert the original material information into a material description vector, convert the real-time online process data into a process state vector, and convert the environmental and equipment metadata into a context metadata vector; based on the target dimension fixed during the training phase, perform missing marker filling on the material description vector, process state vector, and context metadata vector, and truncate them from the end to the target dimension according to the field order. Step 13: Perform deterministic input encoding on the material description vector, process state vector, and context metadata vector; standardize the scalar fields according to the mean and standard deviation fixed during the training phase; map the category fields according to the mapping table fixed during the training phase; and integrate the material description vector, process state vector, and context metadata vector in a fixed order to obtain the input tensor.
3. The online evaluation method for new materials based on artificial intelligence according to claim 1, characterized in that, The input tensors are input to the material branch encoder, process branch encoder, and metadata branch encoder to obtain corresponding embedding vectors, which are then processed by a fusion unit to obtain a unified embedding vector, including: Step 21: Read the input tensor. Based on the field order and target dimension fixed in the same version as the material branch encoder, process branch encoder, and metadata branch encoder, verify that the input tensor is consistent with the input interface of the three. If they are consistent, determine that the input tensor is the input data object. If they are inconsistent, terminate the subsequent calculation. Step 22: Input the input tensors into the material branch encoder, process branch encoder, and metadata branch encoder respectively to obtain the material branch embedding vector, process branch embedding vector, and metadata branch embedding vector; based on the input interface fixed in the same version as the fusion machine, verify that the fixed dimensions of the material branch embedding vector, process branch embedding vector, and metadata branch embedding vector are consistent with the fusion machine input interface. If they are inconsistent, terminate the fusion machine calculation. Step 23: Concatenate the material branch embedding vector, process branch embedding vector, and metadata branch embedding vector in sequence according to the input order fixed in the training phase to obtain the fusion input vector. Input the fusion input vector into the fusion unit and perform multi-layer nonlinear transformation on the fusion input vector through the pre-trained computation graph of the fusion unit to obtain the unified embedding vector.
4. The online evaluation method for new materials based on artificial intelligence according to claim 1, characterized in that, The unified embedding vector is input into the ensemble sub-model to obtain the performance point prediction value, and the cognitive uncertainty is calculated. The applicability domain distance is calculated based on the applicability domain statistics, resulting in a reliability vector, including: Step 31: Read the unified embedding vector and verify that the fixed dimension of the unified embedding vector is consistent with the input interface of the ensemble sub-model; if consistent, determine the unified embedding vector as the unique input data object for subsequent calculations; if inconsistent, write a preset verification failure flag and terminate steps 32 and 33. Step 32: Input the unified embedding vector into a preset number of sub-models in the ensemble sub-model to obtain multiple performance prediction values, sum the multiple performance prediction values to obtain a total, and divide the total by the preset number to obtain the performance point prediction value; Step 33: Calculate the difference between each performance prediction value and the performance point prediction value, square the difference to obtain the squared difference, sum the squared differences and divide by the preset quantity to obtain the cognitive uncertainty; read the applicable domain statistics and verify that its version is consistent with the integrated sub-model version, calculate the difference vector based on the unified embedding vector and the embedding space mean in the applicable domain statistics, obtain the inverse matrix based on the covariance with regularization term in the applicable domain statistics, multiply the difference vector with the inverse matrix to obtain the intermediate vector, perform the inner product of the intermediate vector and the difference vector to obtain the squared distance value, perform the square root operation on the squared distance value to obtain the applicable domain distance; integrate them in the order of cognitive uncertainty first and applicable domain distance to obtain the reliability vector.
5. The online evaluation method for new materials based on artificial intelligence according to claim 1, characterized in that, By comparing the cognitive uncertainty quantity with the cognitive uncertainty threshold, and by comparing the applicable domain distance with the applicable domain distance threshold, the gating state is obtained, and the gating failure type is determined, including: Step 41: Read the cognitive uncertainty, applicable domain distance, cognitive uncertainty threshold, and applicable domain distance threshold, and verify that the cognitive uncertainty, applicable domain distance, cognitive uncertainty threshold, and applicable domain distance threshold all exist and are comparable values; if any one of them is not satisfied, terminate steps 42 and 43 and do not generate a gating state or gating failure type. Step 42: Compare the cognitive uncertainty quantity with the cognitive uncertainty threshold, and compare the applicable domain distance with the applicable domain distance threshold; if the cognitive uncertainty quantity does not exceed the cognitive uncertainty threshold and the applicable domain distance does not exceed the applicable domain distance threshold, the gate control state is determined to be passed; otherwise, the gate control state is determined to be failed. Step 43: When the gating state is in failure, if the applicable domain distance exceeds the applicable domain distance threshold, only the gating failure type is determined to be the type that exceeds the applicable domain; if the applicable domain distance does not exceed the applicable domain distance threshold and the cognitive uncertainty exceeds the cognitive uncertainty threshold, only the gating failure type is determined to be the type that is cognitively uncertain.
6. The online evaluation method for new materials based on artificial intelligence according to claim 1, characterized in that, Within a statistical sliding window, the gating failure rate is calculated for samples with a failed gating state, and compared with a drift threshold to obtain the drift determination result, including: Step 51: Read the gating status and statistical sliding window, and determine the sample set within the window based on the statistical sliding window with the current evaluation time as the window endpoint; count the total number of samples within the window sample set, and count the number of failed samples for the samples whose gating status is failed. Step 52: Verify that the total number of samples in the verification window is greater than zero and the drift threshold is a comparable value; if the condition is met, use the number of failed samples as the numerator and the total number of samples in the window as the denominator, and calculate the ratio of the numerator to the denominator to obtain the gate failure rate; if the condition is not met, terminate step 53. Step 53: Compare the gating failure rate with the drift threshold. When the gating failure rate is greater than or equal to the drift threshold, the drift determination result is determined to be drift. When the gating failure rate is less than the drift threshold, the drift determination result is determined to be no drift.
7. The online evaluation method for new materials based on artificial intelligence according to claim 1, characterized in that, When the drift determination result indicates that drift has occurred, a sampling priority value is calculated based on the cognitive uncertainty, cognitive uncertainty threshold, applicable domain distance, and applicable domain distance threshold. A supplementary testing queue is formed within the supplementary testing budget, and supplementary testing is performed on the supplementary testing queue to obtain a newly labeled sample set, including: Step 61: Read the drift determination result and verify that the drift determination result indicates that drift has occurred. Read the restricted evaluation sample set, cognitive uncertainty quantity, cognitive uncertainty threshold, applicable domain distance, applicable domain distance threshold, and supplementary test budget. Verify that the restricted evaluation sample set is not empty, the cognitive uncertainty threshold and applicable domain distance threshold are comparable values, and the supplementary test budget is a positive integer. If these conditions are not met, terminate steps 62 and 63. Step 62: Read the weight coefficients. For each sample in the restricted assessment sample set, multiply the ratio of cognitive uncertainty to cognitive uncertainty threshold by the weight coefficient to obtain the first weighted value. Multiply the ratio of applicable domain distance to applicable domain distance threshold by the complementary value of the weight coefficient to obtain the second weighted value. Sum the first weighted value and the second weighted value to obtain the sampling priority value. Sort the samples according to the sampling priority value from high to low and select samples whose ranking does not exceed the supplementary testing budget to form a supplementary testing queue. When the number of restricted assessment sample sets is less than the supplementary testing budget, the supplementary testing queue is formed using the restricted assessment sample set. Step 63: Traverse the supplementary testing queue and perform supplementary testing on each sample to obtain the annotation results. Establish a one-to-one correspondence between each sample in the supplementary testing queue and the corresponding annotation results, and combine the supplementary testing queue and the corresponding annotation results to obtain the newly annotated sample set.
8. The online evaluation method for new materials based on artificial intelligence according to claim 1, characterized in that, Construct an incremental training set using the newly labeled sample set and the replay buffer sample set, update the ensemble sub-model parameters and the applicable domain statistics, and then feed the updated ensemble sub-model parameters and the updated applicable domain statistics back to step 3, including: Step 71: Read the newly labeled sample set and the replay buffer sample set and verify that the sample field structure is consistent. If either condition is not met, terminate steps 72 and 73. If the condition is met, merge the newly labeled sample set and the replay buffer sample set and remove duplicates to obtain the incremental training set. Step 72: Read the ensemble sub-model parameters and incremental training configuration and verify that the versions are consistent. Perform incremental training iterations on the ensemble sub-model parameters based on the incremental training set and the number of training rounds configured in the incremental training configuration to obtain the updated ensemble sub-model parameters. Step 73: Read the applicable domain statistics and the unified embedding vector corresponding to the newly labeled sample set and verify that the dimensions are consistent. Calculate the mean vector and the covariance with regularization term of the unified embedding vector corresponding to the newly labeled sample set and merge them with the applicable domain statistics to obtain the updated applicable domain statistics. Then, send the updated ensemble sub-model parameters and the updated applicable domain statistics back to step 3.
9. A novel online evaluation system for materials based on artificial intelligence, characterized in that, The method for online evaluation of novel materials based on artificial intelligence as described in any one of claims 1-8 includes: The input tensor module is used to obtain the original material information, online process real-time data and environmental and equipment metadata of the new material sample to be evaluated, form a material description vector, a process state vector and a context metadata vector, and obtain the input tensor through deterministic input encoding; The unified embedding module is used to input the input tensor into the material branch encoder, process branch encoder and metadata branch encoder to obtain the corresponding embedding vector, and then obtain the unified embedding vector through the fusion unit; The reliability assessment module is used to input the unified embedding vector into the ensemble sub-model to obtain the performance point prediction value and calculate the cognitive uncertainty; it also calculates the application domain distance based on the application domain statistics to obtain the reliability vector. The gating determination module is used to compare the cognitive uncertainty quantity with the cognitive uncertainty threshold, and the applicable domain distance with the applicable domain distance threshold to obtain the gating status and determine the gating failure type. The drift determination module is used to calculate the gating failure rate of samples with failed gating states within a statistical sliding window and compare it with the drift threshold to obtain the drift determination result. The supplementary sampling module is used to calculate the sampling priority value based on the cognitive uncertainty, cognitive uncertainty threshold, applicable domain distance, and applicable domain distance threshold when the drift determination result is that drift has occurred. It forms a supplementary testing queue within the supplementary testing budget and performs supplementary testing on the supplementary testing queue to obtain a newly labeled sample set. The incremental update module is used to construct an incremental training set by adding the newly labeled sample set and the replay buffer sample set, update the parameters of the ensemble sub-model and update the applicable domain statistics, and then send the updated ensemble sub-model parameters and updated applicable domain statistics back to step 3.