Systems and methods for determining similarity between structured simulation datasets

The zero-shot learning technique addresses the inefficiencies in determining similarity between structured simulation datasets by using AI embedding models and user feedback to compute key and value scores, facilitating accurate and efficient retrieval of relevant simulations across diverse domains.

WO2026235685A1PCT designated stage Publication Date: 2026-11-12SIEMENS INDUSTRY SOFTWARE INC +2
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/US2025/027838
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Filing Date
2025-05-05
Publication Date
2026-11-12

AI Technical Summary

Technical Problem

Existing techniques fail to efficiently determine similarity between structured simulation datasets, leading to duplication of work, increased research cycles, and inefficiencies in model validation, verification, and synthetic data generation, particularly in domains like aerospace engineering and computational biology, due to the lack of an end-to-end pipeline for computing trustworthy similarity metrics and utilizing user feedback.

Method used

A zero-shot learning technique is employed to determine similarity between structured simulation datasets using structured metadata representation, which includes generating key and value scores based on hierarchical key-value structures, leveraging AI embedding models to compute structural and semantic similarity without requiring task-specific training data, and incorporating user feedback for continuous learning.

Benefits of technology

This approach enables accurate and efficient identification of relevant past simulations, reducing redundant efforts and enhancing model reuse, validation, and synthetic data generation across diverse domains by providing a scalable and adaptable similarity assessment.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure US2025027838_12112026_PF_FP_ABST
    Figure US2025027838_12112026_PF_FP_ABST
Patent Text Reader

Abstract

Disclosed is a computer-implemented method for determining similarity between structured simulation datasets. The computer-implemented method includes receiving a query simulation dataset and a plurality of target simulation datasets, each comprising structured metadata in a layered key-value format. The method further includes extracting the structured metadata representation from the datasets and generating a key score based on structural similarity between the query and target datasets. Additionally, the method includes generating a value score based on string and numerical similarity for matched metadata keys. The method further includes determining a final similarity score for each target simulation dataset by correlating the key score and value score. Thus, based on the final similarity score, the method includes recommending the plurality of target simulation datasets by ranking the plurality of target simulation datasets in order of relevance to the query dataset. The similarity determination is performed using a zero-shot learning technique and is refined through user feedback for continuous learning.
Need to check novelty before this filing date? Find Prior Art

Description

202503029SYSTEMS AND METHODS FOR DETERMINING SIMILARITY BETWEEN STRUCTURED SIMULATION DATASETSBACKGROUND

[0001] The present embodiments relate to the field of machine learning and, more particularly, relate to systems and methods for determining similarity between structured simulation datasets using a zero-shot learning technique.

[0002] Simulation has become an indispensable analytical instrument in data-driven domains such as aerospace engineering, computational biology, finance, and autonomous system designs. For example, high fidelity simulations that are computationally intensive models embed various parameters along with nested boundary conditions and are run to forecast behavior, test “what-if” scenarios, or generate synthetic data for machine-learning. As the volume of simulation output has increased, so too has the number of accessible internal repositories that may store pre-stored simulation datasets, alternatively referred to as target simulation datasets in structured formats (e.g., JSON). Although a database including the target simulation datasets may include an immense reservoir of prior knowledge, users still struggle to identify which previously executed simulations (e.g., from the target simulation datasets), among thousands of candidates, are most pertinent to a query simulation dataset (e.g., the scenario the users are about to study). Consequently, a disadvantage of the existing technique leads to duplication of work, propagation of potentially inconsistent models, and increased research cycles.

[0003] In a technical challenge, particularly in model validation and verification, users may compare the query simulation dataset against “trusted” reference simulations (e.g., the target simulation datasets) to provide that the governing parameters and settings in the query simulation dataset behave as anticipated. Further, search functions provided by the databases are largely keyword-centric and thus fail to consider the multilayered key-value metadata that encodes solver hierarchies, material property tables, random-seed initialization settings, and other structurally significant attributes while comparing the database with the query simulation dataset.

[0004] Another technical challenge arises during knowledge transfer from the pre-stored simulations in the database. The target simulation datasets may include hidden failure cases such as divergence, non-convergence, or impossible mass-balance results that often emerge only after extensive computing. Thus, identifying and learning from the failure202503029cases may save high-performance computing resources, yet the indicators are difficult to locate within the target simulation datasets. Without an automated mechanism to compare heterogeneous metadata (e.g., in the target simulation datasets), it remains difficult to identify the pre-stored simulations that reveal pitfalls relevant to a new model (e.g., the query simulation dataset).

[0005] In another technical challenge, particularly in real-world projects, thousands of slightly different simulations may often be run to explore different possibilities. Thus, analysis of results from previous, similar simulations may help reduce the number of new runs or simulation comparisons needed. However, this is only possible if similar simulations are quickly identified ideally as soon as a new query is submitted. Traditional search techniques, such as keyword or tag-based search, are slow and require a lot of manual work to set up and maintain, consequently rendering them impractical and less efficient.

[0006] In another technical challenge, particularly in understanding complex systems, some systems are too large, too small, or too intricate to study directly in real-world conditions. In such cases, simulations are used to explore behavior of the systems.However, based on analyzing existing simulations (e.g., the target simulation datasets) that are similar, it becomes possible to understand the key dynamics and interactions within these systems. In some cases, important details such as parameter settings and boundary conditions from past simulations may help improve the accuracy and setup of new simulations. However, finding the most relevant existing simulations remains difficult without an effective way to measure similarity.

[0007] Another technical challenge particularly involves the generation of synthetic data, which is often used to train machine learning models by providing a wide range of example cases. The synthetic data as training samples may either be newly generated or retrieved from existing simulation datasets (e.g., the target simulation datasets). In many cases, past simulations already contain data that represents specific conditions or scenarios needed for training. However, without a reliable technique to identify the most relevant simulations out of the target simulation datasets, valuable data may be overlooked or underutilized.SUMMARY AND DESCRIPTION

[0008] The above-discussed technical challenges reveal a lack of an end-to-end pipeline202503029capable of computing a trustworthy similarity metric between structured simulation datasets (e.g., the target simulation datasets and the query simulation dataset). Some existing approaches such as cosine similarity on flattened parameter vectors, or ad-hoc regular-expression filters on file headers fail to generalize across domains. Consequently, existing techniques do not make use of new advances that may combine text and numbers into a single, meaningful format for comparison between the target simulation datasets and the query simulation dataset. The existing techniques also fail to use user feedback to improve the accuracy of results over time as more simulations are viewed and rated.

[0009] Consequently, the existing techniques cause organizations to expend significant manual effort to rediscover simulations that are already available, duplicating storage footprints and consuming compute hours that may otherwise be directed toward novel inquiry. In critical applications, such as autonomous vehicle safety or pharmaceutical discovery, the inability to rapidly identify the closest prior simulation not only inflates costs but also introduces verification risk and time-to-market delays.

[0010] In light of the above, there exists a need for an improved system and method for determining similarity between the query simulation dataset and the plurality of target simulation datasets.

[0011] Therefore, it is an object of the present invention to provide a system, apparatus, and method for determining similarity between structured simulation datasets using a zeroshot learning technique, with human feedback, and careful consideration of costs, coverage, and staffing.

[0012] Throughout the present disclosure, the term “structured simulation datasets ” refers to simulation-related data that is organized in a predefined, machine-readable format, where the data is associated with the execution, configuration, or outcome of one or more simulation runs. The structured simulation datasets include raw simulation outputs (e.g., time-series values, state variables, or computed results) and metadata that describe the simulation context. The metadata is arranged in a layered or hierarchical key-value structure or format (e.g., JSON or XML). The structured simulation datasets may originate from a variety of domains including, but not limited to, fluid dynamics, structural analysis, thermal simulations, electrical circuit modelling, computational biology, or industrial process modelling. Each of the structured simulation datasets may typically contains multiple layers of descriptive information that represent nested components of a simulation environment. For example, in a multi-physics simulation, a structured simulation dataset202503029may include simulation modules for thermal behavior, structural stress, and fluid flow with own sub-set of parameters and results organized. The structured simulation datasets may serve as both input and reference data for similarity-based recommendation techniques. In some embodiments, the structured simulation datasets may be stored in centralized databases, simulation repositories, or version-controlled platforms that may be accessed via Application Programming Interfaces (APIs) or file transfer protocols. Advantageously, the structured simulation datasets facilitate the standardization and interoperability of simulation data across various tools, software, or environments, thereby supporting efficient model reuse, validation, and synthetic data generation as disclosed herein.

[0013] Throughout the present disclosure, the term “query simulation dataset” refers to the structured simulation dataset acting as an input for the purpose of identifying and retrieving other simulation datasets that exhibit a high degree of similarity. A query simulation dataset represents a simulation run either newly executed, under consideration, or in-progress for which the user seeks to find relevant historical simulations stored in the database or repository. The query simulation dataset enables comparison-based tasks such as model validation, benchmarking, transfer learning, or synthetic data augmentation. The query simulation dataset includes a structured metadata representation formatted in a hierarchical or a layered key-value structure or format. The structured metadata representation may define the simulation context, including boundary conditions, physical parameters, solver settings, model configurations, and domain- specific descriptors. The query simulation dataset serves as the reference against which a plurality of target simulation datasets is compared. The structured metadata representation is extracted from the query simulation dataset and generates feature representations, such as text embeddings of keys and values, to compute a structural and semantic similarity score with each target simulation dataset of the plurality of target simulation datasets in the repository or the database. Consequently, based on the comparison, the most similar past simulations (e.g., from the plurality of target simulation datasets) are ranked and recommended. The query simulation dataset may be submitted via a user interface, received through the API, or automatically generated as part of a continuous simulation workflow. The query simulation dataset may originate from scientific studies or from product development stages, and the query simulation dataset may be provided in any machine-readable format compatible with the system’s metadata extraction engine (e.g., JSON or XML).

[0014] Throughout the present disclosure, the term “plurality of target simulation datasets” or target simulation datasets refers to a collection of structured simulation202503029datasets that are stored in the one or more simulation repositories or the databases and are subject to comparison with the query simulation dataset to identify similarity. Each target simulation dataset of the plurality target simulation datasets represents a previously executed simulation run and includes the structured metadata representation describing the configuration, execution parameters, and contextual attributes of the simulation. The target simulation datasets may be associated with different domains such as scientific research or product developments. The structured metadata representation of each of the target simulation datasets may be formatted in the layered key-value structure or format that captures simulation- specific information such as model type, input parameters, solver settings, mesh configuration, boundary and initial conditions, material properties, output variables, and computational resource usage. The plurality of target simulation datasets serves as the reference corpus against which the query simulation dataset is evaluated. The structured metadata representation is extracted from each of the target simulation datasets and processed in conjunction with the query simulation dataset to compute a key score and a value score representing structural and contextual similarity. Consequently, the key score and the value score are used to determine a final similarity score for each of the target simulation datasets, thereby enabling a ranked recommendation of the most relevant historical simulations.

[0015] Throughout the present disclosure, the term “structured metadata representation ” refers to a machine-readable, hierarchically organized data structure that captures descriptive information associated with the simulation dataset in the form of the layered key- value structure. The structured metadata representation indicates a normalized and interpretable abstraction of the configuration, parameters, and contextual descriptors of the structured simulation datasets. In an example, the structured metadata representation is expressed in the layered key- value pair or structure, where the “key” refers to an attribute name or identifier (e.g., temperature-range), and the “value” refers to the corresponding assigned data (e.g., a temperature value). The term “layered” denotes that the metadata is organized into multiple levels or layers of nesting referring to subsystems, modules, or components within the overall simulation. For example, the structured metadata representation may encode keys such as:relationship or nesting between layers or hierarchical layers of keys.202503029

[0016] The structured metadata representation may be typically found in formats such as JSON or XML, advantageously allowing the data to be parsed, indexed, and compared systematically. The structured metadata representation is extracted from both the query simulation dataset and the plurality of target simulation datasets. Thus, the structured metadata representation serves as the primary input for determining the key score (based on the structural and semantic similarity of keys) and the value score (e.g., based on the comparison of corresponding values). The layered nature of the key- value structure allows the key-value structure to account for both the presence and hierarchical placement of parameters, advantageously enabling a more meaningful similarity evaluation than keyword-based matching. Further, advantageously, the structured metadata representation facilitates vectorization and embedding techniques by providing a consistent framework for extracting keys and values across heterogeneous simulation formats.

[0017] Throughout the present disclosure, the term “key score” refers to a numerical measure indicating the degree of structural similarity between two structured metadata representations specifically, between the structured metadata representation of the query simulation dataset and that of each of the target simulation datasets. The key score quantifies the alignment or correspondence between sets of hierarchical layers of keys-value extracted from each structured metadata representation.

[0018] Throughout the present disclosure, the term “value score” refers to a numerical metric that represents the degree of similarity between the values associated with one or more matched keys in the structured metadata representations of the query simulation dataset and each target simulation dataset of the plurality of target simulation datasets. While the key score captures the structural alignment of metadata, the value score specifically measures the similarity of the actual content both textual and numerical contained within the matched keys.

[0019] Throughout the present disclosure, the term “final similarity score” refers to a scalar metric that quantifies the overall similarity between the query simulation dataset and each target simulation dataset of the plurality of target simulation datasets, based on both structural and contextual comparisons of their respective structured metadata representations. The final similarity score integrates two distinct components (e.g., the key score, which captures structural and semantic alignment between the hierarchical keys, and the value score, which reflects the similarity of the corresponding values associated with the matched keys).202503029

[0020] Throughout the present disclosure, the term “zero-shot learning " refers to a machine learning capability that enables similarity evaluation and makes predictions or recommendations without requiring task-specific training data or explicit prior exposure to the exact simulation datasets being compared. In an example scenario, zero-shot learning allows to generalize across heterogeneous simulation domains and formats by leveraging pre-trained one or more Al embedding models and inference strategies that do not depend on labelled examples or predefined taxonomies. The zero-shot learning is applied in the determination of both the key score and the value score. Advantageously, the use of zeroshot learning eliminates the need for costly and time-consuming preparation of labelled training data for every new simulation domain. Further, zero-shot learning enhances the scalability and adaptability, thus allowing immediate support for new simulation datasets, tools, and formats as they are introduced, without retraining or manual reconfiguration.

[0021] Through the present disclosure, the term “one or more Al embedding models” refers to a plurality of Artificial Intelligence (Al) neural networks. Examples of neural networks include, but are not limited to, Convolutional Neural network (CNN), Deep Neural Network (DNN), Recurrent Neural Network (RNN), and Restricted Boltzmann Machine (RBM). The learning technique for training each Al model uses a plurality of learning data to cause, allow, or control the system to make a determination or analysis. Examples of learning techniques include, but are not limited to, supervised learning, unsupervised learning, semi-supervised learning, or reinforcement learning. At least one of a plurality of CNN, DNN, RNN, RMB models, and the like may be implemented to thereby achieve execution of the present subject matter’s mechanism through the Al models.

[0022] The object of the present invention is achieved by a computer-implemented method for determining similarity between the structured simulation datasets. The method includes receiving, by one or more processors, the query simulation dataset and the target simulation datasets. The query simulation dataset and each target simulation dataset of the plurality of target simulation datasets include the structured simulation datasets.Advantageously, receiving the query simulation dataset and the target simulation datasets from various sources, such as simulation repositories, cloud storage, or APIs, enables flexible integration and supports large-scale, domain-independent similarity evaluation.

[0023] The method further includes extracting, by the one or more processors, the structured metadata representation from each of the query simulation dataset and the202503029plurality of target simulation datasets. The structured metadata representation indicates the layered key- value structure. In an advantageous aspect, extracting the structured metadata representation in the layered key-value format allows for standardized processing of complex simulation data, enabling precise structural, and semantic comparison across diverse datasets.

[0024] In an embodiment, the method includes serializing content associated with each of the query simulation dataset and the plurality of target simulation datasets into the hierarchical key-value format. Further, the method includes generating a structured data file for each of the query simulation dataset and the plurality of target simulation datasets based on serializing. Further, the method includes identifying a plurality of hierarchical layers of keys corresponding to at least one of nested objects and properties from the respective structured data file. The plurality of hierarchical layers indicates levels of nesting within the structured data file. In an advantageous aspect, serializing the content or the simulation content into the hierarchical key-value format and identifying the nested layers enables accurate capture of complex data relationships, thereby facilitating detailed layer- wise similarity analysis.

[0025] In an embodiment, the method includes grouping each key and a corresponding value within the respective structured data file based on the plurality of hierarchical layers. Further, the method includes extracting the structured metadata representation based on the grouping. The structured metadata representation extracted for the query simulation dataset and the plurality of target simulation datasets are formatted in the layered key-value structure. In an advantageous aspect, grouping keys and values based on hierarchical layers provides that the structured metadata accurately reflects the simulation’s internal organization, enabling consistent, and meaningful comparisons across the datasets (e.g., the query simulation dataset and the plurality of target simulation datasets).

[0026] The method further includes generating, by the one or more processors, the key score based on the structural similarity between the respective structured metadata representation corresponding to the query simulation dataset and the plurality of target simulation datasets. In an advantageous aspect, generating the key score based on structural similarity allows quantifying alignment of the structured metadata representation, thereby enabling more accurate identification of relevant past simulations.

[0027] In an embodiment, the method includes generating text embeddings based on encoding a plurality of keys from each hierarchical layer of the plurality of hierarchical202503029layers of the structured metadata representation corresponding to the query simulation dataset and each target simulation dataset of the plurality of target simulation datasets, using a text embedding model. The text embeddings indicate a numerical vector representing a semantic meaning of the key. In an advantageous aspect, using the text embeddings to encode the keys captures their semantic meaning, thereby identifying similar concepts even when different terminology is used across the datasets. Further, in an advantageous aspect, the text embedding model enables domain independent representation of the keys, allowing scalable and language-independent similarity analysis without requiring manual keyword mapping or domain-specific training.

[0028] In an embodiment, the method includes generating a dot product matrix between the text embeddings of the plurality of keys from the structured metadata representation of the query simulation dataset and the text embeddings of the plurality of keys from the structured metadata representation of each target simulation dataset of the plurality of target simulation datasets. The dot product matrix indicates pairwise similarity scores between the plurality of keys. In an advantageous aspect, the dot product matrix provides an efficient tool to compute the pairwise similarity scores between the keys, thereby enabling fast and scalable comparison across large sets of simulation metadata.

[0029] In an embodiment, the method includes selecting one or more pairs of keys having a similarity score from the dot product matrix iteratively. The similarity score indicates a highest value in the dot product matrix. In an advantageous aspect, the highest similarity scores provide that the most semantically relevant matches are prioritized, improving the accuracy of the structural alignment between the datasets.

[0030] In an embodiment, the method includes determining a mean pool score of the pairwise similarity scores associated with the selected one or more pairs of keys for each hierarchical layer of the plurality of hierarchical layers. The mean pool score indicates an average structural similarity within each of the plurality of hierarchical layers. In an advantageous aspect, the mean pool score for each hierarchical layer of the plurality of hierarchical layers enables layer-wise evaluation of the structural similarity, thereby preserving the contextual depth of metadata during comparison.

[0031] In an embodiment, the method includes determining an exponential weighted average of the mean pooled score from each hierarchical layer of the plurality of hierarchical layers using predefined weights. Further, the method includes generating the key score based on the exponentially weighted average of the pairwise similarity scores for202503029each hierarchical layer of the plurality of hierarchical layers.

[0032] In an embodiment, the method includes marking one or more values in a row and a column of the dot product matrix as invalid in response to the selected one or more pairs of keys from the dot product matrix. The marking prevents the reuse of the selected one or more pairs of keys in subsequent iterations. Further, the method includes repeating the selection and marking iteratively until all values in the dot product matrix are marked as invalid. In an advantageous aspect, marking the selected key pairs as invalid prevents duplicate matching and provides one-to-one correspondence, thereby leading to a more accurate and unbiased similarity assessment.

[0033] The method further includes generating, by the one or more processors, the value score based on a string similarity and a numerical similarity associated with the one or more matched keys in the structured metadata representation. In an advantageous aspect, the value score provides a comprehensive measure of content- level alignment, enhancing the precision of similarity evaluation between the datasets.

[0034] In an embodiment, the method includes identifying the one or more matched keys between the respective structured metadata representation of the query simulation dataset and the plurality of target simulation datasets. Further, the method includes separating values associated with the one or more matched keys into string values and numerical values for each hierarchical layer of the plurality of hierarchical layers of the structured metadata representation. In an advantageous aspect, the identification of the one or more matched keys enables tailored similarity computations for the string values and the numerical values, thereby improving the accuracy of value-level comparisons across the plurality of hierarchical layers.

[0035] In an embodiment, the method includes generating the text embeddings of the string values using the text embedding model and determining cosine similarity scores between the text embeddings of corresponding string values. Further, the method includes generating a string value similarity score based on averaging the cosine similarity scores. In an advantageous aspect, the text embeddings with the cosine similarity score allow for efficient comparison of the string values, thereby capturing semantic relationships even when exact wording differs, and improving the relevance of the string value similarity score.

[0036] In an embodiment, the method includes generating a numerical value similarity score based on the cosine similarity scores between the numerical values associated with202503029the one or more matched keys.

[0037] In an embodiment, the method includes determining a layer value score based on correlating the string value similarity score and the numerical value similarity score and computing an exponentially weighted average of the layer value score using predefined weights associated with each of the plurality of hierarchical layers. Consequently, the value score is generated based on the exponentially weighted average of the layer value scores. In an advantageous aspect, computing the exponentially weighted average enables a balanced evaluation of the content across the layers of the structured metadata representation, thereby enhancing the overall sensitivity and accuracy of the value score.

[0038] The method further includes, determining, by the one or more processors, the final similarity score for the query simulation dataset with respect to each target simulation dataset of the plurality of target simulation datasets based on the key score and the value score. The final similarity score is determined based on correlating the key score and the value score for each target simulation dataset of the plurality of target simulation datasets corresponding to the query simulation dataset. In an advantageous aspect, the final similarity score provides a universal measure capturing both structural and content-level similarity, thereby enabling more accurate and meaningful recommendations.

[0039] In an embodiment, the method includes generating a ranking list for the plurality of target simulation datasets based on the final similarity score. In an advantageous aspect, the ranking list thus allows for efficient identification of the most relevant simulation datasets (e.g., out of the target simulation datasets), thereby supporting faster decisionmaking and reuse of past simulations.

[0040] The method further includes recommending, by the one or more processors, the target simulation datasets based on ranking each of the target simulation datasets using the final similarity score.

[0041] The method further includes detecting a user interaction with the recommended target simulation datasets. The user interaction indicates a quality of the recommendation. Further, the method includes regenerating the key score and the value score based on a weighted key importance derived from the detected user interaction. In an advantageous aspect, the user interaction enables adjustment in the key score and the value score, thus leading to continuous learning and personalization and improving the accuracy and relevance of future similarity recommendations.

[0042] In an embodiment, the user interaction includes a user rating the recommendation202503029based on a qualitative assessment of the similarity between the query simulation dataset and one or more target simulation datasets of the plurality of target simulation datasets.

[0043] In an embodiment, the user interaction includes identifying a set of positive keys in the respective structured metadata representation of the query simulation dataset and the plurality of target simulation datasets, as rated by the user. Further, the user interaction includes identifying a set of negative keys in the respective structured metadata representation of the query simulation dataset and the plurality of target simulation datasets, as rated by the user. Further, the method includes determining a set of common keys indicating a union of intersections between the set of positive keys and a set of keys from the plurality of target simulation datasets rated as the neutral match, and between the set of negative keys and the set of keys from the plurality of target simulation datasets rated based on the user interaction. In an advantageous aspect, the identification of the set of positive keys, the set of negative keys, and the set of common keys based on the user interaction enables learning which metadata characterizes the highest perceived similarity in the structured metadata representation, thereby allowing for refined and user-aligned similarity scoring.

[0044] In an embodiment, the method includes deriving a weighted key-importance based on the set of common keys determined from the user interaction. The weighted keyimportance indicates a relative contribution of the key to a similarity determination.Further, the method includes determining a normalized weight for each of the set of common keys based on a frequency of occurrence across one or more user interactions. Consequently, the method includes regenerating the key score and the value score based on the normalized weight. In an advantageous aspect, the weighted key-importance based on the user interaction allows dynamic prioritization of influential metadata, thereby leading to more accurate and adaptive similarity evaluations over time.

[0045] Thus, the similarity determination between structured simulation datasets is performed using the zero-shot technique, which is further enhanced by incorporating human feedback to enable continuous learning and adaptive refinement.

[0046] The object of the present invention is also achieved by an apparatus for determining the similarity between the structured simulation datasets. The apparatus includes a memory and one or more processors communicatively coupled to the memory. The memory includes programmable instructions executable by the one or more processors. The programmable instructions, when executed by the one or more processors,202503029cause the one or more processors to perform one or more methods, as discussed throughout the present disclosure.

[0047] The object of the present invention is also achieved by a system for determining the similarity between the structured simulation datasets. In an embodiment, the system includes the apparatus for recommending the plurality of target simulation datasets.

[0048] The object of the present invention is also achieved by a computer-program product being disclosed. The computer-program product has machine-readable instructions stored therein, that when executed by one or more processors, cause the one or more processors to perform one or more methods, as discussed throughout the present invention.

[0049] To further clarify the advantages and features of the present invention, a more particular description of the invention will be rendered by reference to specific embodiments thereof, which are illustrated in the appended drawings. It is appreciated that these drawings depict only typical embodiments of the invention and are therefore not to be considered limiting of its scope. The invention will be described and explained with additional specificity and detail with the accompanying drawings.BREIF DESCRIPTION OF THE DRAWINGS

[0050] These and other features, aspects, and advantages of the present invention will become better understood when the following detailed description is read with reference to the accompanying drawings in which like characters represent like parts throughout the drawings, where:

[0051] Figure 1 illustrates a block diagram of a system for determining the similarity between the structured simulation datasets, in accordance with an embodiment;

[0052] Figure 2 illustrates a block diagram of an apparatus for determining the similarity between the structured simulation datasets, in accordance with an embodiment;

[0053] Figure 3 illustrates a functional block diagram of the system for determining the similarity between the structured simulation datasets, in accordance with an embodiment; and

[0054] Figure 4 illustrates a flow chart of a method for determining the similarity between the structured simulation datasets, in accordance with an embodiment.DETAILED DESCRIPTION

[0055] Skilled artisans will appreciate that elements in the drawings are illustrated for202503029simplicity and may not have necessarily been drawn to scale. For example, the flow charts illustrate the method in terms of the most prominent steps involved to help improve understanding of aspects of the present invention. Further, in terms of the construction of the device, one or more components of the device may have been represented in the drawings by conventional symbols, and the drawings may show only those specific details that are pertinent to understanding the embodiments of the present invention so as not to obscure the drawings with details that will be readily apparent to those of ordinary skill in the art having the benefit of the description herein.

[0056] For the purpose of promoting an understanding of the principles of the invention, reference will now be made to the various embodiments, and specific language will be used to describe the same. It will nevertheless be understood that no limitation of the scope of the invention is thereby intended, such alterations and further modifications in the illustrated system, and such further applications of the principles of the invention as illustrated therein being contemplated as would normally occur to one skilled in the art to which the invention relates.

[0057] It will be understood by those skilled in the art that the foregoing general description and the following detailed description are explanatory of the invention and are not intended to be restrictive thereof.

[0058] Reference throughout this specification to “an aspect,” “another aspect,” or similar language provides that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the present invention. Thus, appearances of the phrase “in an embodiment,” “in another embodiment,” and similar language throughout this specification may, but do not necessarily, all refer to the same embodiment.

[0059] The terms “comprises,” “comprising,” or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process or method that includes a list of steps does not include only those steps but may include other steps not expressly listed or inherent to such process or method. Similarly, one or more devices or sub-systems or elements or structures or components proceeded by “comprises... a” does not, without more constraints, preclude the existence of other devices or other sub-systems or other elements or other structures or other components or additional devices or additional subsystems or additional elements or additional structures or additional components.

[0060] Figure 1 illustrates a block diagram of a system 100 for determining the similarity202503029between the structured simulation datasets, in accordance with an embodiment of the present invention. The system 100 includes an apparatus 102, a query simulation dataset 104, and a database 106a storing a plurality of target simulation datasets 106b. The apparatus 102 is further configured to generate and output a ranked list of recommended target simulation datasets 110 (e.g., from the plurality of target simulation datasets 106b) based on a final similarity score. Additionally, the apparatus 102 is configured to incorporate user interaction or feedback 108 to refine and improve recommended target simulation datasets 110 process over time.

[0061] In an embodiment, the apparatus 102 is configured to receive the query simulation dataset 104 and the plurality of target simulation datasets 106b from the database 106a. Each of these datasets includes the structured simulation data represented in a standardized, machine-readable format. In an advantageous aspect, the structured format of the structured simulation data is to be provided for providing consistency in processing and may include hierarchical serialization formats such as JSON or XML, which support the layered key- value structures.

[0062] In an embodiment, the apparatus 102, upon receiving the query simulation dataset 104 and the plurality of target simulation datasets 106b, is configured to extract the structured metadata representation from both the query simulation dataset 104 and each of the target simulation datasets 106b. Advantageously, the structured metadata representation provides the internal configuration, parameter settings, boundary conditions, and contextual attributes of each simulation in the layered key-value format. Each of the keys in the structured metadata representation may correspond to a simulation parameter or model attribute, while its associated value may be numeric or textual, indicating a specific configuration used in the simulation run.

[0063] In an embodiment, the apparatus 102 is configured to generate the key score for each of the target simulation datasets 106b relative to the query simulation dataset 104. The key score is generated based on the structural similarity between the hierarchical keys in the structured metadata representation. In an example, the structural similarity corresponds to the presence of matching keys, semantic alignment, and hierarchical position within the structured metadata representation. Further, the apparatus 102 is configured to utilize embedding techniques to encode keys into numerical vectors and compare the numerical vectors using similarity metrics such as the dot product matrix.

[0064] In an embodiment, the apparatus 102 is configured to generate the value score for202503029each of the target simulation datasets 106b based on comparing the values associated with matched keys. In an example, the comparison includes both string-based comparisons (e.g., solver types, model names) and numeric comparisons (e.g., step size, mesh density), thus capturing both semantic and quantitative differences. Consequently, the comparison generates the value score as discussed in forthcoming paragraphs, thereby reflecting the contextual similarity between the query simulation dataset 104 and the target simulation datasets 106b.

[0065] In an embodiment, the apparatus 102 is configured to determine the final similarity score for each of the target simulation datasets 106b based on correlating the key score and the value score. The correlation may involve a weighted summation technique correlating the importance of structural and value-level similarity. Therefore, the final similarity score serves as the primary metric for determining relevance between the query simulation dataset 104 and each of the target simulation datasets 106b in the repository or database 106a.

[0066] In an embodiment, the apparatus 102 is configured to generate the ranked list of the recommended target simulation datasets 110, ordering the target simulation datasets 106b such as in a highest to lowest final similarity score. Advantageously, the ranked list enables users to quickly identify the most relevant historical simulations (e.g., out of the target simulation datasets 106b) corresponding to a new or in-progress simulation task.

[0067] As illustrated in Figure 1, the system 100 further incorporates user interaction or feedback 108, and thus, the apparatus 102 is configured to learn from user preferences and continuously refine the similarity computation. Advantageously, the user interaction or the human feedback may influence a future key score and a future value score to align better with domain- specific relevance, enhancing the adaptability and precision over time of the system 100.

[0068] In an advantageous aspect, the apparatus 102 stores the one or more Al models with the zero-shot learning technique. The apparatus 102 performs the similarity determination without requiring prior domain-specific training or labelled examples for each type of the structured simulation datasets. In an advantageous aspect, the zero-shot learning technique is particularly beneficial in environments where the structured simulation datasets vary significantly across domains, formats, or parameter naming conventions. Therefore, utilizing pre-trained embedding models capable of encoding both textual keys and values into a unified semantic space, the apparatus 102 is configured to202503029generalize across unseen simulation structures and structured metadata representations.

[0069] Further, the zero-shot learning technique allows the apparatus 102 to identify structurally and contextually relevant target simulation datasets 106b even when the apparatus 102 has not been explicitly trained on those specific simulation types.Advantageously, the zero-shot learning technique enhances the flexibility, scalability, and domain-agnostic nature of the similarity recommendation, enabling the system 100 to operate effectively across diverse simulation repositories without manual intervention or retraining.

[0070] In an example scenario, an engineering team working on the thermal analysis of a battery module system for electric vehicles generates the query simulation dataset 104 including metadata such as detailed in Table 1 below and a nested structure specifying boundary conditions and mesh parameters.Table 1:

[0071] In the example scenario, the query simulation dataset 104 is submitted to the apparatus 102, which is configured to recommend similar past simulations stored in the database 106a.

[0072] In the example scenario, the engineering team’s current simulation uses a different modelling tool and different parameter naming conventions compared to prior simulations (e.g., env_temp vs. ambient_temp), and the zero-shot learning technique in the one or more Al models within the apparatus 102 recognizes the semantic equivalence by embedding the keys and values into a shared vector space. The apparatus 102 generates both the key score (e.g., based on structural alignment) and the value score (e.g., based on parameter- level similarity) between the query simulation dataset 104 and each of the target simulation datasets 106b in the database 106a.’

[0073] In the example scenario, the apparatus 102 then outputs the ranked list of the recommended target simulation datasets 110, highlighting a number of (e.g., several) past simulations of thermal runaway scenarios in the battery module system for the electric202503029vehicles under high ambient temperatures. The recommended target simulation datasets 110 help the engineering team validate their setup, avoid redundant modelling efforts, and gain insights into effective configurations without requiring manual tagging or retraining the system 100 on battery-specific simulations.

[0074] In a further aspect in the example scenario, the apparatus 102 receives the user interaction 108 or the human feedback on the recommended simulations (e.g., the recommended target simulation datasets 110). For example, the engineering team may rate certain recommendations as highly relevant and mark others as irrelevant. Therefore, based on the human feedback, the apparatus 102 identifies the set of positive keys (e.g., material, ambient_temp) and the set of negative keys (e.g., unused solver settings or outdated model types) that influenced the engineering team’s decision. Further, the apparatus 102 derives the weighted key-importance values and dynamically adjusts future similarity computations. As a result, subsequent recommendations are more aligned with user preferences, allowing the system 100 to continuously leam and refine its scoring mechanism or recommendation over time.

[0075] Figure 2 illustrates a block diagram 200 of the apparatus 102 for determining the similarity between the structured simulation datasets, in accordance with an embodiment of the present invention.

[0076] As shown in Figure 2, the apparatus 102 includes a bus 201, a processing unit 202, a memory unit 204, a communication unit 206, an I / O interface 208, and an output unit 210. The apparatus 102 may be a computer, a workstation, or a virtual machine running on host hardware. As an alternative, the apparatus 102 may be a real or a virtual group of computers (e.g., the technical term for a real group of computers is “cluster,” the technical term for a virtual group of computers is “cloud”).

[0077] The processing unit 202 may include one or more processors as a single processing unit or a number of (e.g., several) units. The processing unit 202 may be implemented as one or more microprocessors, microcomputers, microcontrollers, digital signal processors, central processing units, state machines, logic circuitries, and / or any devices that manipulate signals based on operational instructions. Among other capabilities, the one or more processors are configured to fetch and execute computer-readable instructions and data stored in the memory unit 204.

[0078] The memory unit 204 includes one or more computer-readable storage media that may communicate via the bus 201. The memory unit 204 may include non-volatile storage202503029elements. Examples of such non-volatile storage elements may include magnetic hard discs, optical discs, floppy discs, flash memories, or forms of electrically programmable memories (EPROM) or electrically erasable and programmable (EEPROM) memories. In addition, the memory may, in some examples, be considered a non-transitory storage medium. The term “non-transitory” may indicate that the storage medium is not embodied in a carrier wave or a propagated signal. However, the term “non-transitory” should not be interpreted to mean that the memory is non-movable. In certain examples, a non-transitory storage medium may store data that may, over time, change (e.g., in Random Access Memory (RAM) or cache).

[0079] The memory unit 204 may further include any non-transitory computer-readable medium known in the art including, for example, volatile memory, such as static randomaccess memory (SRAM) and dynamic random-access memory (DRAM), and / or nonvolatile memory, such as read-only memory (ROM), erasable programmable ROM, flash memories, hard disks, optical disks, and magnetic tapes.

[0080] The memory unit 204 includes a database 212 and modules 214. The database 212 is configured to be accessed by the processing unit 202 and stores information as required by processing unit 202 to perform the one or more functions. The database 212 may store the target simulation datasets 106b. The database 212 may also store historical data including information of past key scores, value scores, the user interactions. Additionally, the database 212 may serve dual functions in storing different types of information.

[0081] The modules 214, amongst other things, include routines, programs, objects, components, data structures, etc. that perform particular tasks or implement data types. The modules 214 may also be implemented as signal processor(s), state machine(s), logic circuitries, and / or any other device or component that manipulates signals based on operational instructions.

[0082] Further, the modules 214 may be implemented in hardware, instructions executed by the processing unit 202, or by a combination thereof. The processing unit 202 may include a computer, a processor, a state machine, a logic array, or any other suitable devices capable of processing instructions. The processing unit 202 may be a general-purpose processor that executes instructions to cause the general-purpose processor to perform the required tasks or, the processing unit may be dedicated to performing the required functions. In another embodiment of the present invention, the modules 214 may be machine-readable instructions (software) that, when executed by the processing unit202503029202, perform any of the described functionalities / methods, as discussed throughout the present invention.

[0083] Further, the modules 214 may be implemented through the Al model. A function associated with Al may be performed through the non-volatile memory, the volatile memory, and the processing unit 202.

[0084] The modules 214 may store the one or more AI / ML models with zero-shot learning capabilities. A function associated with the AI / ML module may be performed through the non-volatile memory, the volatile memory, and the processing unit 202.

[0085] The communication unit 206 is configured to communicate sensor data or any other content over a communication network via a communication port or interface or using the bus 201. Further, the communication unit 206 may include a communication port or a communication interface for sending and receiving simulation datasets via the communication network. The communication port or the communication interface may be a part of the processing unit 202 or may be a separate component. The communication port may be created in software or may be a physical connection in hardware. The communication port may be configured to connect with the communication network, external media, the display, or any other components in the system 100, or combinations thereof. The connection with the communication network may be a physical connection, such as a wired Ethernet connection, or may be established wirelessly, as discussed above. Likewise, the additional connections with other components of the system 100 may be physical or may be established wirelessly. The communication unit 206 may include the Wi-Fi module or Bluetooth module for enabling wireless communication capability and data exchange capability between various modules of the system 100.

[0086] The I / O interface 208 refers to hardware or software components that enable communication between various modules of the system 100. The I / O interface 208 serves as a communication medium for exchanging information, commands, signals, or query responses with other devices or systems. The I / O interface 208 may be a part of the processing unit 202 or maybe a separate component. The I / O interface 208 may be created in software or may be a physical connection in hardware. The VO interface 208 may be configured to connect with an external network, external media, the display, or any other components, or combinations thereof. The external network may be a physical connection, such as a wired Ethernet connection, or may be established wirelessly.

[0087] The output device 210 that may be an output unit includes a display device. The202503029display device may be an Augmented Reality / Virtual Reality (AR / VR) device to display a virtual environment to the user. The display device may include a display screen. As a non-limiting example, the display screen may be Light Emitting Diode (LED), Liquid Crystal Display (LCD), Organic Light Emitting Diode (OLED), Active Matrix Organic Light Emitting Diode (AMOLED), or Super Active Matrix Organic Light Emitting Diode (AMOLED) screen. The display screen may be of varied resolutions. The remote device, while operating as the output unit, is further configured for presenting the selected visual frames.

[0088] In an embodiment, the present embodiments also contemplate a computerprogram product, having machine-readable instructions stored therein, when executed by the one or more processors, to cause the one or more processors to perform the method for determining the similarity between the structured simulation datasets. The details on the method(s) performed by the one or more processors have been elaborated in subsequent paragraphs at least with reference to Figure. 3.

[0089] Further, the present invention also contemplates a non-transitory computer-readable medium encoded with executable instructions. The executable instructions, when executed by the one or more processors, cause the one or more processors to perform a method for determining the similarity between the structured simulation datasets. The details on the method(s) performed by the one or more processors have been elaborated in subsequent paragraphs at least with reference to Figure. 3.

[0090] Figure 3 illustrates a functional block diagram of the system 100 for determining the similarity between the structured simulation datasets, in accordance with an embodiment of the present invention.

[0091] The functional blocks 300 of the system 100 are implemented as the apparatus 102. The functional blocks 300 includes an extraction module 302, a key-score generation module 304, a value-score generation module 306, a determining module 308, and a recommendation module 310. The extraction module 302, the key-score generation module 304, the value-score generation module 306, the determining module 308, and the recommendation module 310 are in communication with each other. In an example, the user interaction 108 may be received by the recommendation module 310 to recommend the ranked list of the recommended target simulation datasets 110.

[0092] In an embodiment, the processing unit 202 via the extraction module 302 is configured to receive the query simulation dataset 104 and the target simulation datasets202503029106b from the database 106a. The query simulation dataset 104 and the target simulation datasets 106b include the structured simulation datasets.

[0093] In an embodiment, the processing unit 202 via the extraction module 302 is configured to serialize the content associated with each of the query simulation dataset 104 and each of the target simulation datasets 106b. In an example, the serialization corresponds to converting the native simulation configuration or result files typically in diverse proprietary or semi-structured formats into the hierarchical key-value format, such as JSON, or XML, thereby supporting nesting and metadata hierarchy.

[0094] In an embodiment, the processing unit 202 via the extraction module 302 is configured to generate the structured data file for each of the simulation datasets (e.g., the query simulation dataset 104 and each of the target simulation datasets 106b). In an example, the structured data file serves as a normalized version of the metadata present in each simulation dataset, advantageously preserving the essential simulation descriptors such as input parameters, boundary conditions, material properties, and domain- specific configuration values in an accessible and layered format.

[0095] In an embodiment, the processing unit 202 via the extraction module 302 is configured to identify the plurality of hierarchical layers of keys from each of the simulation datasets. The plurality of hierarchical layers corresponds to the levels of nesting within the datasets and may include both nested objects (e.g., subsystem-specific parameter groups) and nested properties (e.g., component-specific conditions or thresholds). Advantageously, the identification of the plurality of hierarchical layers is to be provided to reflect the complex structure of simulation models accurately, particularly in multi-physics or component-based simulations.

[0096] In an embodiment, the processing unit 202 via the extraction module 302 is configured to group each key and the corresponding value of the key based on the plurality of hierarchical layers of the datasets. Advantageously, the grouping provides that each key-value pair is associated with the correct contextual level (e.g., top-level simulation configuration versus deeply nested component-level parameters), thereby preserving the semantic relationships between data elements.

[0097] In an embodiment, the processing unit 202 via the extraction module 302 is configured to consequently extract the structured metadata representation for each of the query simulation dataset 104 and the target simulation datasets 106b based on the hierarchical grouping.202503029

[0098] In an example, the structured metadata representation is formatted in the layered key- value structure, which provides a unified and consistent foundation for downstream comparison operations. The structured metadata representation is optimized for both structural similarity analysis (e.g., through key alignment and embedding) and contextual value matching (e.g., comparing associated string or numeric values), as performed by the subsequent modules of the apparatus 102 in the forthcoming paragraphs.

[0099] In an advantageous aspect, the structured metadata representation serves as the primary input to the key-score generation module 304 and the value-score generation module 306, providing that both structural and semantic features of the simulation datasets (e.g., the query simulation dataset 104 and each of the target simulation datasets 106b) are retained and made accessible for similarity computation. Therefore, based on enabling standardized extraction across heterogeneous simulation inputs, the extraction module 302 advantageously supports domain-independent, scalable, and accurate simulation similarity determination.

[0100] In an embodiment, the processing unit 202 via the key-score generation module 304 is configured to generate the key score. The key score quantifies the structural similarity between the structured metadata representation of the query simulation dataset 104 and that of each of the target simulation datasets 106b. The structured metadata representations, as previously extracted by the extraction module 302, are formatted in the layered key- value structure. In an example, the keys represent simulation parameters or descriptors, and the values represent associated data or settings.

[0101] In an embodiment, the processing unit 202 via the key-score generation module 304 is configured to generate the text embeddings for the plurality of keys present in each hierarchical layer of the plurality of hierarchical layers of the structured metadata representation. The generation of the text embeddings is performed separately for the query simulation dataset 104 and for each target simulation dataset 106b. In a non-limiting example, the processing unit 202 via the key-score generation module 304 utilizes a sentence transformer model, a neural network architecture derived from Bidirectional Encoder Representations from Transformers (BERT). The sentence transformer model is adapted using Siamese and triplet network structures, which are specifically designed to generate the text embeddings with semantically similar phrases or terms placed closer together in the vector space.

[0102] Advantageously, the text embedding model converts each metadata key (e.g., a202503029textual identifier such as ambient_temperature) into a high-dimensional numerical vector that captures the semantic meaning of the term beyond exact string matches. Therefore, the text embedding model allows for robust similarity computation even when different datasets use domain-specific variations in naming (e.g., env_temp vs. ambient_temp), as long as the underlying semantics are equivalent or similar.

[0103] In an embodiment, the processing unit 202 via the key-score generation module 304 is configured to generate the dot product matrix between the text embeddings of the keys from the query simulation dataset 104 and the text embeddings of the keys from the target simulation datasets 106b. The dot matrix as illustrated in Table 2 below provides the pairwise similarity scores between each key in the query simulation dataset 104 and each key in the target simulation datasets 106b. Further, each value in the dot matrix represents the cosine-aligned semantic similarity between two key embeddings, thereby effectively measuring alignment of the metadata terms in meaning and context.Table 2:wherein the pairwise similarity scores are computed based on equation (1) below:

[0104] In an embodiment, the processing unit 202 via the key-score generation module 304 is configured to isolate the most relevant key matches and proceeds to select the one or more pairs of keys (e.g., selecting the highest similarity scores from the dot product matrix). In an example, the selection is performed iteratively, in which, in each iteration, for example, the key pair corresponding to the maximum value in the dot product matrix is selected. The key pairs are treated as semantically equivalent or structurally aligned for the purpose of similarity scoring.

[0105] In an embodiment, the processing unit 202 via the key-score generation module 304 is configured to mark the corresponding row and column of the selected key pair in the dot product matrix as invalid after each selection. In an advantageous aspect, the marking prevents the reuse of the selected keys in subsequent iterations, thereby enforcing a one-to-one correspondence and thereby eliminating bias that arises from repeatedly matching the same key to multiple other keys. Further, the processing unit 202 via the key-202503029score generation module 304 is configured to iteratively perform the selection and the invalidation process continuously until all values in the dot product matrix are marked invalid. Therefore, this indicates that no further unique matches are possible between the two sets of keys.

[0106] In an embodiment, the processing unit 202 via the key-score generation module 304 is configured to determine the mean pool score for each hierarchical layer of the plurality of hierarchical layers within the structured metadata representation. The mean pool score is determined based on averaging the pairwise similarity scores of all selected key pairs that fall within a given hierarchical level. In an example, the mean pool score corresponds to a localized structural similarity indicator, reflecting the alignment of the key structures at a specific metadata depth (e.g., at the top-level parameters or deeply nested component attributes).

[0107] In an embodiment, the processing unit 202 via the value-score generation module 306 is configured to combine the layer-specific scores (e.g., for each hierarchical layer of the plurality of hierarchical layers) into a single similarity metric. The key-score generation module 304 determines the exponential weighted average of the mean pool scores from each of the plurality of hierarchical layers. The predefined weights assigned to each hierarchical layer of the plurality of hierarchical layers may be based on the perceived importance or sensitivity of parameters at that depth, allowing domain- specific tuning, if desired. Consequently, the exponentially weighted average aggregation leads to the determination of the key score. Therefore, the key score encapsulates the structural similarity between the query simulation dataset 104 and each target simulation dataset 106b.

[0108] Further, the key score is then forwarded to the determining module 308, where the key score is used in conjunction with the value score to determine the final similarity score used for ranking recommendations. In an advantageous aspect, using semantically rich text embeddings, iterative matching logic, and hierarchical weighting, the key-score generation module 304 enables robust, scalable, and domain-agnostic structural similarity assessment between complex simulation datasets.

[0109] In an embodiment, the processing unit 202 via the value-score generation module 306 is configured to generate the value score. The value score represents the content-level similarity between the query simulation dataset 104 and each of the target simulation datasets 106b. The value score complements the key score by focusing not on the202503029structural arrangement of metadata keys, but on the semantic and the numerical similarity of the values associated with those keys.

[0110] In an embodiment, the processing unit 202 via the value-score generation module 306 is configured to identify the one or more matched keys between the structured metadata representations of the query simulation dataset 104 and each of the target simulation datasets 106b. The one or more matched keys refer to the keys that were identified as structurally similar or identical during the computation of the key score. The identification of the one or more matched keys advantageously provides that value comparisons are made only for semantically aligned parameters across the datasets (e.g., the query simulation dataset 104 and each of the target simulation datasets 106b).

[0111] In an embodiment, the processing unit 202 via the value-score generation module 306 is configured to separate the associated values for each of the one or more matched keys into the string values and the numerical values, based on their data type, for each of the plurality of hierarchical layers within the structured metadata representation. For example, a parameter such as solver_datatype would be treated as a string, while a parameter such as ambient_temperature would be classified as numerical. Advantageously, the separation is essential for applying different comparison methods suited to each data type.

[0112] In an embodiment, the processing unit 202 via the value-score generation module 306 is configured to utilize the text embedding model (e.g., the sentence transformer model) to generate the text embeddings of the string values that numerically encode the semantic meaning of each string. In an example, the text embeddings indicate highdimensional vectors that capture contextual understanding of the textual values, such as identifying similarity between terms such as “temperature” and “ambient temperature,” thereby enabling effective comparisons even when different terminology is used across the datasets.

[0113] In an embodiment, the processing unit 202 via the value-score generation module 306 is configured to determine the cosine similarity scores between the corresponding string value embeddings of the query simulation dataset 104 and the target simulation datasets 106b. In an example, the cosine similarity scores provide a normalized measure of angular distance between the embedding vectors, with values ranging from completely dissimilar to semantically identical.

[0114] Further, the processing unit 202 via the value-score generation module 306 is202503029configured to generate the string value similarity score based on averaging the cosine similarity scores across all matched string values within a hierarchical layer.

[0115] In an embodiment, the processing unit 202 via the value-score generation module 306 is configured to generate the numerical value similarity score for the numerical values. The value-score generation module 306 is configured to determine the cosine similarity scores between the numerical values (e.g., scalar numerical values) associated with the one or more matched keys. Advantageously, the numerical value similarity score accommodates domain-specific variability in scale and units while still enabling effective comparison. Consequently, the numerical value similarity score corresponds to numeric value alignment between the query simulation dataset 104 and the target simulation datasets 106b.

[0116] In an embodiment, the processing unit 202 via the value-score generation module 306 is configured to correlate the string value similarity score and the numerical value similarity score two measures to determine the layer value score for each hierarchical layer of the plurality of hierarchical layers. In an example, the correlation includes weighted averaging or other mathematical operations to balance the influence of semantic (e.g., textual) and quantitative (e.g., numeric) dimensions. The layer value score represents the degree of value-level similarity between the query simulation dataset 104 and the target simulation datasets 106b for that specific layer (e.g., out of the plurality of hierarchical layers) of metadata.

[0117] In an embodiment, the processing unit 202 via the value-score generation module 306 is configured to compute the exponentially weighted average of the layer value scores using predefined weights associated with each hierarchical layer of the plurality of hierarchical layers. The predefined weights allow higher importance to specific layers based on either domain knowledge or learned prioritization from the user interaction.

[0118] In an embodiment, the processing unit 202, via the value-score generation module 306, is configured to consequently generate the value score based on the exponentially weighted average of the layer value scores. In an example, the value score represents a scalar metric that quantifies the contextual and semantic similarity of parameter values between the query simulation dataset 104 and each of the target simulation datasets 106b. The value score is then forwarded to the determining module 308, where the value score is used in conjunction with the key score to compute the final similarity score for ranking and recommendation.202503029

[0119] In an advantageous aspect, through the layered approach, and semantically informed comparison, the value-score generation module 306 enables accurate, scalable, and domain-independent value similarity assessment across heterogeneous simulation datasets.

[0120] In an embodiment, the processing unit 202 via the determining module 308 is configured to determine the final similarity score for the query simulation dataset 104 with respect to each target simulation dataset of the plurality of target simulation datasets 106b. In an example, the final similarity score integrates the key score (e.g., the structural similarity) and the value score (e.g., the content similarity) to provide a single, unified measure of similarity that captures both the structuring of the simulation metadata and values assigned to key parameters.

[0121] In an example, the key score, as generated by the key-score generation module 304, represents the structural alignment between the layered key-value structure of the query simulation dataset 104 and the target simulation datasets 106b. The value score, generated by the value-score generation module 306, represents the semantic and numerical similarity of the matched values across the metadata.

[0122] The determining module 308 is configured to correlate the key score and the value score to derive the final similarity score. The correlation may include various mathematical functions. In some embodiments, weights applied to the key score and the value score may be static (e.g., predefined) or dynamic (e.g., adjusted based on system learning or the user interaction).

[0123] The final similarity score represents a scalar value associated with each of the target simulation datasets 106b relative to the query simulation dataset 104.Advantageously, the final similarity score allows the system 100 to evaluate and compare all candidate target datasets in a quantitative, ranked format.

[0124] In an embodiment, the processing unit 202 via the recommendation module 310 is configured to generate the ranking list of the target simulation datasets 106b. In an example, the recommendation module 310 is configured to sort the target simulation datasets 106b in descending order of their respective final similarity scores, such that datasets with the highest degree of relevance appear at the top of the list. This ranked output forms the recommended target simulation datasets 110, as illustrated in Figure 1 and Figure 3.

[0125] In an advantageous aspect, the ranking enables users to efficiently identify which202503029historical simulations most closely resemble the current simulation context represented by the query dataset 104. Thus, the need for manual searching or tagging is eliminated, and faster model validation, reuse of simulation components, or generation of synthetic data is supported. Further, because the ranking is based on the final similarity score that considers both structure and content, ranking provides a more meaningful and accurate recommendation compared to systems that use only text matching or metadata filters.

[0126] In some embodiments, the recommendation module 310 is further configured to present the ranked list, or the ranked results may be delivered via a user interface, application programming interface (API), or integrated into a downstream simulation pipeline via the output device 210.

[0127] In an embodiment, the processing unit 202 via the recommendation module 310 is configured to detect the user interaction 108 to enhance the quality and relevance of the recommended target simulation datasets 110. In an example, the user interaction indicates a quality of the recommendation as denoted by the user. In an advantageous aspect, the user interaction 108 serves as a feedback signal, playing a pivotal role in enabling the system 100 to continuously learn from user preferences and refine its recommendations. The user interaction 108 may be collected through an interface that allows the user to rate or classify each of the recommended target simulation datasets 110 based on their perceived similarity to the query simulation dataset 104. The user interaction includes the user rating the recommended target simulation datasets 110 based on a qualitative assessment of the similarity between the query simulation dataset 104 and one or more of the target simulation datasets 106.

[0128] In an embodiment, the user interaction 108 leads to regeneration of the key score and the value score in a more informed manner, based on observed user choices. In a nonlimiting example, the feedback provided by the user is categorized into qualitative labels, which may typically include a good recommendation. The good recommendation indicates that the system’s 100 output closely aligns with the user’s intent. Similarly, a neutral recommendation indicates that the similarity is slightly acceptable or unclear, and a bad recommendation indicates that the result is irrelevant or misleading. In an advantageous aspect, the classification of the user interaction 108 into the qualitative labels helps the system 100 isolate which metadata elements (e.g., the keys) contributed most significantly to the user’s perception of similarity.

[0129] In an embodiment, the processing unit 202 via the recommendation module 310 is202503029configured to classify the metadata keys into multiple sets based on the structured metadata representations associated with the query dataset 104 and the target simulation datasets 106b.

[0130] In an example, the recommendation module 310 is configured to identify the set of positive keys in (set- A) that exhibit semantic and value-level alignment between the query simulation dataset 104 and those target simulation datasets marked as the good recommendations. Thus, the set-A is inferred to have contributed positively to the perceived similarity.

[0131] In an example, the recommendation module 310 is configured to identify the set of neutral positive keys (e.g., set-A’) representing similar matches identified between the query simulation dataset 104 and the neutral recommendations.

[0132] In an example, the recommendation module 310 is configured to identify the set of negative Keys (e.g., set-B) referring to keys that are marked as the bad recommendations. Thus, the set-B indicates a mismatch or irrelevance in the recommendation.

[0133] In an example, the recommendation module 310 is configured to identify the set of neutral negative keys (e.g., set-B’) referring to dissimilarity in the query-neutral comparison, and serve as a secondary filter to validate the truly negative attributes present in the bad recommendations.

[0134] In an example, the recommendation module 310 is configured to identify the set of common keys (set-H) referring to intersections between A and A’, and between B and B’, and the recommendation module 310 is configured to derive from their union to form a consolidated set-H.

[0135] Thus, the set-H represents the keys that most consistently influence the quality of recommendations either based on enhancing or degrading perceived similarity and is used for deriving the weighted key-importance.

[0136] In an embodiment, the processing unit 202 via the recommendation module 310 is configured to update the final similarity score based on the user feedback. Thus, the recommendation module 310 is configured to count a frequency of occurrence of each key in the respective set-H over time.

[0137] Further, the recommendation module 310 is configured to determine a normalization score based on dividing the individual key count by the total number of keys observed, resulting in a normalized importance score ranging from 0 to 1 for each of the202503029key. The normalization score is computed based on equation (2):KeyWeightNormalizedscoreNumber of times the key is part of H for a set Q, P, N, Neu Number of times all the keys are part of H for a set Q, P, N, NEu ... (2)

[0138]

[0139] The normalization score indicates the relative influence of the key in contributing to high-quality similarity matches, as inferred directly from the user interaction 108.

[0140] Thus, the normalized weights (e.g., the key importance weights) are integrated into both the key score and the value score generations in subsequent iterations.

[0141] In an example, during the key score generation (e.g., via the key-score generation module 304), the mean pooling is modified to apply the normalized weight (e.g., weighted averaging). Advantageously, this provides that the keys with greater inferred importance contribute more significantly to the overall structural similarity score.

[0142] In an advantageous aspect, the user feedback enhances the final similarity scoring, thereby allowing the system 100 to deliver adaptive and user-centric recommendations. The apparatus 102 via the recommendation module 310, correlates user perception with internal metadata structure and the content, thus evolving over time to reflect domainspecific preferences without the need for manually defined key weighting schemes.Therefore, alignment between human feedback and inference of the apparatus 102 enables a more intelligent, explainable, and practically effective recommendation system, significantly enhancing the simulation dataset retrieval and reuse experience in high-volume simulation environments.

[0143] Figure 4 illustrates a flow chart of a method 400 for determining the similarity between the structured simulation datasets, in accordance with an embodiment of the present invention. The method 400 includes a series of operation steps 402 through 416 performed by the processing unit 202 of the system 100.

[0144] At step 402, the processing unit 202 receives the query simulation dataset 104 and the target simulation datasets 106b. The query simulation dataset 104 and each of the202503029target simulation datasets 106b include the structured simulation datasets.

[0145] Advantageously, receiving the query simulation dataset and the target simulation datasets from various sources such as simulation repositories, cloud storage, or APIs enables flexible integration and supports large-scale, domain-independent similarity evaluation.

[0146] At step 404, the processing unit 202 extracts the structured metadata representation from each of the query simulation dataset 104 and the target simulation datasets 106b. The structured metadata representation indicates the layered key-value structure. In an advantageous aspect, extracting the structured metadata representation in the layered key-value format allows for standardized processing of complex simulation data, enabling precise structural and semantic comparison across diverse datasets.

[0147] In an embodiment, the method 400 further includes serializing content associated with each of the query simulation dataset 104 and the plurality of target simulation datasets 106b into the hierarchical key-value format. Further, the method 400 includes generating the structured data file for each of the query simulation dataset 104 and the plurality of target simulation datasets 106b based on serializing. Further, the method 400 includes identifying the plurality of hierarchical layers of keys corresponding to at least one of the nested objects and the properties from the respective structured data file. The plurality of hierarchical layers indicates levels of nesting within the structured data file. In an advantageous aspect, serializing the content or the simulation content into the hierarchical key- value format and identifying the nested layers enables accurate capture of complex data relationships, thereby facilitating detailed layer-wise similarity analysis.

[0148] In an embodiment, the method 400 includes grouping each of the key and the corresponding value within the respective structured data file based on the plurality of hierarchical layers. Further, the method 400 includes consequently extracting the structured metadata representation based on the grouping. The structured metadata representation extracted for the query simulation dataset 104 and the target simulation datasets 106b are formatted in the layered key -value structure. In an advantageous aspect, grouping keys and values based on hierarchical layers provides that the structured metadata accurately reflects the simulation’s internal organization, enabling consistent and meaningful comparisons across the datasets (e.g., the query simulation dataset and the plurality of target simulation datasets).

[0149] At step 406, the processing unit 202 generates the key score based on the202503029structural similarity between the respective structured metadata representation corresponding to the query simulation dataset 104 and the target simulation datasets 106b. In an advantageous aspect, generating the key score based on structural similarity allows quantifying alignment of the structured metadata representation, thereby enabling more accurate identification of relevant past simulations.

[0150] In an embodiment, the method 400 includes generating text embeddings based on encoding the plurality of keys from each hierarchical layer of the plurality of hierarchical layers of the structured metadata representation corresponding to the query simulation dataset 104 and each of the target simulation datasets 106b, using the text embedding model. The text embeddings indicate the numerical vector representing the semantic meaning of the key. In an advantageous aspect, using the text embeddings to encode the keys captures their semantic meaning, thereby identifying similar concepts even when different terminology is used across the datasets. Further, in an advantageous aspect, the text embedding model enables domain-independent representation of the keys, allowing scalable and language-independent similarity analysis without requiring manual keyword mapping or domain-specific training.

[0151] In an embodiment, the method 400 includes generating the dot product matrix between the text embeddings of the plurality of keys from the structured metadata representation of the query simulation dataset 104 and the text embeddings of the plurality of keys from the structured metadata representation of each of the target simulation datasets 106b. The dot product matrix indicates the pairwise similarity scores between the plurality of keys. In an advantageous aspect, the dot product matrix provides the efficient tool to compute the pairwise similarity scores between the keys, thereby enabling fast and scalable comparison across large sets of simulation metadata.

[0152] In an embodiment, the method 400 includes selecting one or more pairs of keys having a similarity score from the dot product matrix iteratively. The similarity score indicates a highest value in the dot product matrix. In an advantageous aspect, the highest similarity scores provide that the most semantically relevant matches are prioritized, improving the accuracy of the structural alignment between the datasets.

[0153] In an embodiment, the method 400 includes determining the mean pool score of the pairwise similarity scores associated with the selected one or more pairs of keys for each hierarchical layer of the plurality of hierarchical layers. The mean pool score indicates an average structural similarity within each hierarchical layer of the plurality of202503029hierarchical layers. In an advantageous aspect, the mean pool score for each hierarchical layer of the plurality of hierarchical layers enables layer- wise evaluation of the structural similarity, thereby preserving the contextual depth of metadata during comparison.

[0154] In an embodiment, the method 400 includes determining the exponential weighted average of the mean pooled score from each hierarchical layer of the plurality of hierarchical layers using predefined weights. Further, the method 400 includes generating the key score based on the exponentially weighted average of the pairwise similarity scores for each hierarchical layer of the plurality of hierarchical layers.

[0155] In an embodiment, the method 400 includes marking one or more values in the row and the column of the dot product matrix as invalid in response to the selected one or more pairs of keys from the dot product matrix. The marking prevents the reuse of the selected one or more pairs of keys in subsequent iterations. Further, the method 400 includes repeating the selection and marking iteratively until all values in the dot product matrix are marked as invalid. In an advantageous aspect, marking the selected key pairs as invalid thus prevents duplicate matching and provides one-to-one correspondence, thereby leading to a more accurate and unbiased similarity assessment.

[0156] At step 408, the processing unit 202 generates the value score based on the string similarity and the numerical similarity associated with the one or more matched keys in the structured metadata representation. In an advantageous aspect, the value score provides a comprehensive measure of content-level alignment, enhancing the precision of similarity evaluation between the datasets.

[0157] In an embodiment, the method 400 includes identifying the one or more matched keys between the respective structured metadata representation of the query simulation dataset 104 and the target simulation datasets 106b. Further, the method 400 includes separating values associated with the one or more matched keys into string values and numerical values for each hierarchical layer of the plurality of hierarchical layers of the structured metadata representation. In an advantageous aspect, the identification of the one or more matched keys enables tailored similarity computations for the string values and the numerical values, thereby improving the accuracy of value-level comparisons across the plurality of hierarchical layers.

[0158] In an embodiment, the method 400 includes generating the text embeddings of the string values using the text embedding model, and determining cosine similarity scores between the text embeddings of corresponding string values. Further, the method 400202503029includes generating a string value similarity score based on averaging the cosine similarity scores. In an advantageous aspect, the text embeddings with the cosine similarity score allow for efficient comparison of the string values, thereby capturing semantic relationships even when exact wording differs, and improving the relevance of the string value similarity score.

[0159] In an embodiment, the method 400 includes generating the numerical value similarity score based on the cosine similarity scores between the numerical values associated with the one or more matched keys.

[0160] In an embodiment, the method 400 includes determining the layer value score based on correlating the string value similarity score and the numerical value similarity score, and computing the exponentially weighted average of the layer value score using predefined weights associated with each of the plurality of hierarchical layers.Consequently, the value score is generated based on the exponentially weighted average of the layer value scores. In an advantageous aspect, computing the exponentially weighted average enables a balanced evaluation of the content across the layers of the structured metadata representation, thereby enhancing the overall sensitivity and accuracy of the value score.

[0161] At step 410, the processing unit 202 determines the final similarity score for the query simulation dataset 104 with respect to each of the target simulation datasets 106b based on the key score and the value score. The final similarity score is determined based on the correlation of the key score and the value score for each of the target simulation datasets 106b corresponding to the query simulation dataset 104. In an advantageous aspect, the final similarity score provides a quantification capturing both structural and content-level similarity, thereby enabling more accurate and meaningful recommendations.

[0162] At step 412, the processing unit 202 generates the ranking list for the plurality of target simulation datasets 106b based on the final similarity score. In an advantageous aspect, the ranking list thus allows for efficient identification of the most relevant simulation datasets (e.g., out of the target simulation datasets), thereby supporting faster decision-making and reuse of past simulations.

[0163] At step 414, the processing unit 202 recommends the target simulation datasets 110 based on the ranking for each of the target simulation datasets 106b using the final similarity score.

[0164] At step 416, the processing unit 202 detects the user interaction 108 with the202503029recommended target simulation datasets 110. The user interaction 108 indicates the quality of the recommendation. Further, the method 400 includes regenerating the key score and the value score based on the weighted key importance derived from the detected user interaction. In an advantageous aspect, the user interaction enables adjustment in the key score and the value score, thus leading to continuous learning and personalization and improving the accuracy and relevance of future similarity recommendations.

[0165] In an embodiment, the user interaction 108 includes the user rating the recommendation based on the qualitative assessment of the similarity between the query simulation dataset 104 and one or more of the target simulation datasets 106b.

[0166] In an embodiment, the user interaction 108 includes identifying the set of positive keys in the respective structured metadata representation of the query simulation dataset 104 and the target simulation datasets 106b rated. Further, the user interaction 108 includes identifying the set of negative keys in the respective structured metadata representation of the query simulation dataset 104 and the target simulation datasets 106b rated.

[0167] Further, the method 400 includes determining the set of common keys indicating the union of intersections between the set of positive keys and the set of keys from the plurality of target simulation datasets 106b rated as the neutral match, and between the set of negative keys and the set of keys from the plurality of target simulation datasets 106b rated based on the user interaction 108. In an advantageous aspect, the identification of the set of positive keys, the set of negative keys, and the set of common keys based on the user interaction enables learning which metadata characterizes the highest perceived similarity in the structured metadata representation, thereby allowing for refined and user-aligned similarity scoring.

[0168] In an embodiment, the method 400 includes deriving the weighted key-importance based on the set of common keys determined from the user interaction. The weighted keyimportance indicates a relative contribution of the key to the similarity determination. Further, the method 400 includes determining the normalized weight for each of the set of common keys based on the frequency of occurrence across one or more user interactions. Consequently, the method 400 includes regenerating the key score and the value score based on the normalized weight. In an advantageous aspect, the weighted key-importance based on the user interaction 108 allows dynamic prioritization of influential metadata, thereby leading to more accurate and adaptive similarity evaluations over time.

[0169] The various actions, acts, blocks, steps, or the like in the flow diagrams may be202503029performed in the order presented, in a different order, or simultaneously. Further, in some embodiments, some of the actions, acts, blocks, steps, or the like may be omitted, added, modified, skipped, or the like without departing from the scope of the invention.

[0170] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one ordinary skilled in the art to which this invention belongs. The system, methods, and examples provided herein are illustrative only and not intended to be limiting.

[0171] While specific language has been used to describe the present subject matter, any limitations arising on account thereto are not intended. As would be apparent to a person in the art, various working modifications may be made to the method in order to implement the inventive concept as taught herein. The drawings and the foregoing description give examples of embodiments. Those skilled in the art will appreciate that one or more of the described elements may well be combined into a single functional element. Alternatively, certain elements may be split into multiple functional elements. Elements from one embodiment may be added to another embodiment.

[0172] The foregoing description of the specific embodiments will so fully reveal the general nature of the embodiments herein that others can, by applying current knowledge, readily modify and / or adapt for various applications such specific embodiments without departing from the generic concept, and, therefore, such adaptations and modifications should and are intended to be comprehended within the meaning and range of equivalents of the disclosed embodiments. It is to be understood that the phraseology or terminology employed herein is for the purpose of description and not of limitation. Therefore, while the embodiments herein have been described in terms of preferred embodiments, those skilled in the art will recognize that the embodiments herein can be practiced with modification within the scope of the embodiments as described herein.

Claims

202503029WE CLAIM:

1. A computer-implemented method for determining similarity between structured simulation datasets, the computer-implemented method comprising:receiving, by one or more processors, a query simulation dataset and a plurality of target simulation datasets, wherein each dataset of the query simulation dataset and the plurality of target simulation datasets comprises structured simulation datasets;extracting, by the one or more processors, a structured metadata representation from each dataset of the query simulation dataset and the plurality of target simulation datasets, wherein the structured metadata representation indicates a layered key- value structure;generating, by the one or more processors, a key score based on structural similarity between the respective structured metadata representation corresponding to the query simulation dataset and the plurality of target simulation datasets;generating, by the one or more processors, a value score based on a string similarity and a numerical similarity associated with one or more matched keys in the structured metadata representation;determining, by the one or more processors, a final similarity score for the query simulation dataset with respect to each target simulation dataset of the plurality of target simulation datasets based on the key score and the value score; andrecommending, by the one or more processors, the plurality of target simulation datasets based on ranking each target simulation dataset of the plurality of target simulation datasets using the final similarity score.

2. The computer-implemented method according to claim 1, further comprising: detecting a user interaction with the recommended plurality of target simulation datasets, wherein the user interaction indicates a quality of the recommendation; and regenerating the key score and the value score based on a weighted key-importance derived from the detected user interaction.

3. The computer-implemented method according to claim 1, wherein extracting the structured metadata representation comprises:serializing content associated with each dataset of the query simulation dataset and the plurality of target simulation datasets into a hierarchical key-value format;202503029generating a structured data file for each dataset of the query simulation dataset and the plurality of target simulation datasets based on serializing;identifying a plurality of hierarchical layers of keys corresponding to nested objects, properties, or the nested objects and the properties from the respective structured data file, wherein the plurality of hierarchical layers indicates levels of nesting within the structured data file;grouping each key and a corresponding value within the respective structured data file based on the plurality of hierarchical layers; andextracting the structured metadata representation based on the grouping, wherein the structured metadata representation extracted for the query simulation dataset and the plurality of target simulation datasets are formatted in the layered key-value structure.

4. The computer-implemented method according to any of the preceding claims, wherein generating the key score (304) comprises:generating text embeddings based on encoding a plurality of keys from each hierarchical layer of the plurality of hierarchical layers of the structured metadata representation corresponding to the query simulation dataset and each target simulation dataset of the plurality of target simulation datasets, using a text embedding model, wherein the text embeddings indicate a numerical vector representing semantic meaning of a key;generating a dot product matrix between the text embeddings of the plurality of keys from the structured metadata representation of the query simulation dataset and the text embeddings of the plurality of keys from the structured metadata representation of each target simulation dataset of the plurality of target simulation datasets, wherein the dot product matrix indicates pairwise similarity scores between the plurality of keys;selecting one or more pairs of keys having a similarity score from the dot product matrix iteratively, wherein the similarity score indicates a highest value in the dot product matrix;determining a mean pool score of the pairwise similarity scores associated with the selected one or more pairs of keys for each hierarchical layer of the plurality of hierarchical layers, wherein the mean pool score indicates an average structural similarity within each hierarchical layer of the plurality of hierarchical layers;determining an exponential weighted average of the mean pooled score from each hierarchical layer of the plurality of hierarchical layers using predefined weights; and202503029generating the key score based on the exponentially weighted average of the pairwise similarity scores for each hierarchical layer of the plurality of hierarchical layers.

5. The computer-implemented method according to claim 4, further comprising: marking one or more values in a row and a column of the dot product matrix as invalid in response to the selected one or more pairs of keys from the dot product matrix, wherein the marking prevents reuse of the selected one or more pairs of keys in subsequent iterations; andrepeating the selection and marking iteratively until all values in the dot product matrix are marked as invalid.

6. The computer-implemented method according to any of claims 1 to 3, wherein generating the value score comprises:identifying the one or more matched keys between the respective structured metadata representation of the query simulation dataset and the plurality of target simulation datasets;separating values associated with the one or more matched keys into string values and numerical values for each hierarchical layer of the plurality of hierarchical layers of the structured metadata representation;generating the text embeddings of the string values using the text embedding model;determining cosine similarity scores between the text embeddings of corresponding string values;generating a string value similarity score based on averaging the cosine similarity scores;generating a numerical value similarity score based on the cosine similarity scores between the numerical values associated with the one or more matched keys;determining a layer value score based on correlating the string value similarity score and the numerical value similarity score;computing an exponentially weighted average of the layer value score using predefined weights associated with each hierarchical layer of the plurality of hierarchical layers; andgenerating the value score based on the exponentially weighted average of the layer value scores.2025030297. The computer-implemented method according to any of claims 1 to 3, wherein determining the final similarity score comprises determining the final similarity score based on correlating the key score and the value score for each target simulation dataset of the plurality of target simulation dataset corresponding to the query simulation datasets.

8. The computer-implemented method according to any of claims 1 to 3, further comprising:generating a ranking list for the plurality of target simulation datasets based on the final similarity score.

9. The computer-implemented method according to any of claims 1 to 3, wherein the user interaction comprises a user rating the recommendation based on a qualitative assessment of the similarity between the query simulation dataset and one or more target simulation datasets of the plurality of target simulation datasets.

10. The computer-implemented method according to any of claims 1 to 3, further comprising:identifying a set of positive keys in the respective structured metadata representation of the query simulation dataset and the plurality of target simulation datasets rated based on the user interaction;identifying a set of negative keys in the respective structured metadata representation of the query simulation dataset and the plurality of target simulation datasets rated based on the user interaction; anddetermining a set of common keys indicating a union of intersections between the set of positive keys and a set of keys from the plurality of target simulation datasets rated as a neutral match, and between the set of negative keys and the set of keys from the plurality of target simulation datasets rated based on the user interaction.

11. The computer-implemented method according to claim 10, further comprising: deriving a weighted key-importance based on the set of common keys determined from the user interaction, wherein the weighted key-importance indicates a relative contribution of the key to a similarity determination;202503029determining a normalized weight for each of the set of common keys based on a frequency of occurrence across one or more user interactions; andregenerating the key score and the value score based on the normalized weight.

12. The computer-implemented method according to any of claims 1 to 3, wherein the determining of the similarity between the structured simulation datasets is performed using a zero-shot technique further incorporating human feedback for continuous learning.

13. An apparatus for determining similarity between structured simulation datasets, the apparatus comprising:one or more processing units; anda memory unit communicatively coupled to the one or more processing units, wherein the memory unit comprises a recommendation module stored in the form of machine-readable instructions executable by the one or more processing units, wherein the recommendation module is configured to perform method steps according to any of the claims 1 to 12.

14. A system for determining similarity between structured simulation datasets, the system comprising:an apparatus according to claim 13 for recommending the plurality of target simulation datasets according to any of the method claims 1 to 12.

15. A computer-program product, having machine-readable instructions stored therein, which when executed by one or more processors, cause the one or more processors to perform a method according to any of the claims 1 to 12.