Electronic device for providing physical property prediction data of composite according to composition ratio by using artificial intelligence, and operating method thereof

Abstract

According to one aspect of the present disclosure, a method for providing physical property prediction data of a composite according to a composition ratio by using artificial intelligence comprises the steps of: checking data related to at least one composition that constitutes the composite; encoding, into first physical property data, physical property data of the composite synthesized on the basis of the data related to the at least one composition; calculating an embedding result corresponding to the first physical property data through a first model corresponding to an artificial intelligence learning model based on a neural network; and providing the physical property prediction data of the composite for second physical property data on the basis of the calculated embedding result.

Description

Electronic device and its operating method for providing prediction data on the properties of a composite according to the composition ratio using artificial intelligence

[0001] Embodiments of the present disclosure relate to an electronic device and an operating method thereof that provide data predicting the properties of a composite according to a composition ratio using artificial intelligence, and more specifically, to providing data predicting the properties of a composite produced according to a composition ratio by learning basic experimental data on the properties of the composite.

[0002] Chemical complexes possess specific properties based on the content and ratio of their constituent components. For chemical complexes with specific chemical properties, formulation data for the compositions is limited at the experimental level. Considering recent advances in artificial intelligence, it may be possible to develop a trained model by augmenting and synthesizing data using limited training data.

[0003] By synthesizing data similar to the limited experimental data available and increasing the similarity of the synthesized data to an AI learning model, the model can be updated to confirm the user's desired results. This raises the issue of learning data based on the composition of chemical complexes and generating predictive data based on the learned data.

[0004] However, the electronic device and its operating method that provide prediction data on the properties of a composite according to the composition ratio using conventional artificial intelligence had a problem in that the properties of the composite could not be accurately predicted based on the learned content.

[0005] Embodiments of the present disclosure aim to address various issues, including the aforementioned ones, by providing an electronic device and its operating method capable of sampling similar data synthesized based on limited learning data, using it as input along with previously provided experimental data, and predicting the physical properties of a composite using a specific activation function. However, these tasks are exemplary and are not intended to limit the scope of the present disclosure.

[0006] According to one aspect of the present disclosure, a method for providing property prediction data of a composite according to a composition ratio using artificial intelligence may include a step of identifying data related to at least one composition constituting the composite, a step of encoding property data of the composite synthesized based on the data related to the at least one composition into first property data, a step of calculating an embedding result corresponding to the first property data through a first model corresponding to an artificial intelligence learning model based on a neural network, and a step of providing property prediction data of the composite for second property data based on the calculated embedding result.

[0007] According to the present embodiment, the step of identifying data related to the at least one composition may include the step of identifying experimental data for the complex and synthetic data for the complex, and the step of sampling data related to the at least one composition based on the experimental data and the synthetic data, wherein the data related to the at least one composition may include type, content, and formulation data of at least one composition constituting the complex.

[0008] According to the present embodiment, a method for providing prediction data of properties of a composite according to a composition ratio using artificial intelligence may further include a step of confirming data related to at least one sampled composition as an input data set, wherein the input data set is characterized in that it is preprocessed as an input data set for learning of the first model, and the synthetic data may be characterized in that it is synthesized through data augmentation from the experimental data.

[0009] According to the present embodiment, the step of encoding with the first property data may include the step of setting the property information of the complex to be predicted through the first model and the step of performing batch normalization by expressing data related to the composition of the complex to confirm the set property information as individual features.

[0010] According to the present embodiment, the step of producing an embedding result corresponding to the first material property data may include a step of generating a masked feature through masking of individual features of the first material property data.

[0011] According to the present embodiment, a method for providing prediction data of a composite material property according to a composition ratio using artificial intelligence may include a step of calculating a scale term based on the masked feature to determine importance for the individual feature, a step of confirming decision outputs of individual domains through an activation function layer based on the importance, and a step of outputting the final decision outputs of the individual domains as an embedding result corresponding to the first material property data through a fully connected layer.

[0012] According to the present embodiment, the step of confirming the decision output of the individual domains may include the step of confirming the first decision output in a first domain among the individual domains and the step of confirming the second decision output in a second domain among the individual domains based on the first decision output, and the individual domains including the first domain and the second domain may be characterized in that they are configured with the fully connected layer.

[0013] According to the present embodiment, the step of providing the property prediction data of the complex may include the step of providing the second property data, at least part of which is different from the first property data learned through the first model.

[0014] According to the present embodiment, the step of providing the property prediction data of the complex may include the step of confirming the ground truth for the second property data and the step of evaluating the accuracy of the second property data based on the ground truth.

[0015] According to the present embodiment, an electronic device for providing property prediction data of a composite according to a composition ratio using artificial intelligence includes a communication unit, a memory, and at least one processor electrically connected to the communication unit and the memory, wherein the at least one processor is configured to check experimental data for a composition formulation of a composite having specific properties based on data related to at least one composition constituting the composite, encode the property data of the composite synthesized based on the data related to the at least one composition into first property data, and calculate an embedding result corresponding to the first property data through a first model corresponding to an artificial intelligence learning model based on a neural network, and provide property prediction data of the composite for second property data based on the calculated embedding result.

[0016] Other aspects, features and advantages other than those described above will become apparent from the following detailed description, claims and drawings for carrying out the invention.

[0017] Additionally, these general and specific aspects may be implemented using a system, a computer program, or a combination of any systems and computer programs.

[0018] According to the exemplary embodiments of the present disclosure, as described above, to more accurately predict the physical properties of a composite, an artificial intelligence learning model can be trained using multiple activation functions from a limited training data set, and prediction data for specific physical properties desired by the user can be generated based on the performance evaluation results. Of course, the scope of the present disclosure is not limited by these effects.

[0019] FIG. 1 is a block diagram schematically illustrating internal components of an electronic device that provides prediction data on the properties of a composite according to a composition ratio using artificial intelligence according to an exemplary embodiment of the present disclosure.

[0020] FIG. 2 is a block diagram schematically illustrating a system that provides prediction data on the properties of a composite according to a composition ratio according to an exemplary embodiment of the present disclosure.

[0021] FIG. 3 is a schematic diagram illustrating a method for providing prediction data on the properties of a composite according to a composition ratio according to an exemplary embodiment of the present disclosure.

[0022] FIG. 4 is a flowchart schematically illustrating a method for providing prediction data on the properties of a composite according to a composition ratio according to an exemplary embodiment of the present disclosure.

[0023] FIG. 5 is a flowchart schematically illustrating a process for verifying data related to a composite composition according to an exemplary embodiment of the present disclosure.

[0024] FIG. 6 is a block diagram schematically illustrating a process for encoding and embedding data for predicting properties of a composite according to a composition ratio according to an exemplary embodiment of the present disclosure.

[0025] FIG. 7 is an exemplary diagram schematically illustrating a process for encoding data for predicting physical properties of a composite according to a composition ratio according to an exemplary embodiment of the present disclosure.

[0026] FIG. 8 is a flowchart schematically illustrating a process for embedding data for predicting properties of a composite according to a composition ratio according to an exemplary embodiment of the present disclosure.

[0027] FIG. 9a and FIG. 9b are exemplary diagrams of prediction data of physical properties of a composite according to a composition ratio corresponding to an activation function according to an exemplary embodiment of the present disclosure.

[0028] FIG. 10 is an exemplary diagram schematically illustrating a performance evaluation result screen of the physical property prediction data of a composite according to a composition ratio according to an exemplary embodiment of the present disclosure.

[0029] The present disclosure is capable of various modifications and embodiments. Specific embodiments are illustrated in the drawings and described in detail in the detailed description. The effects and features of the present disclosure, as well as methods for achieving them, will become clearer with reference to the embodiments described in detail below, along with the drawings. However, the present disclosure is not limited to the embodiments disclosed below and may be implemented in various forms.

[0030] In the examples below, the terms first, second, etc. are not used in a limiting sense, but are used for the purpose of distinguishing one component from another.

[0031] In the examples below, singular expressions include plural expressions unless the context clearly indicates otherwise.

[0032] In the following examples, terms such as “include” or “have” mean that a feature or component described in the specification is present, and do not preclude the possibility that one or more other features or components may be added.

[0033] In the following examples, when a part such as a layer, region, or component is said to be on or above another part, it includes not only the case where it is directly above the other part, but also the case where another region, component, or the like is interposed in between.

[0034] For convenience of explanation, the sizes of components in the drawings may be exaggerated or reduced. For example, the sizes and thicknesses of each component shown in the drawings are arbitrarily indicated for convenience of explanation, and thus the present disclosure is not necessarily limited to the figures shown.

[0035] In some embodiments, where implementations are otherwise feasible, specific sequences of operations may be performed in a different order than described. For example, two steps described in succession may be performed substantially simultaneously, or in a reverse order from the described order.

[0036] In this specification, “A and / or B” refers to the case where it is A, or B, or both A and B. And, “at least one of A and B” refers to the case where it is A, or B, or both A and B.

[0037] In the following examples, when it is said that layers, regions, components, etc. are connected, it includes cases where the layers, regions, components, etc. are directly connected, and / or cases where other layers, regions, components, etc. are interposed between the layers, regions, components, etc. and are indirectly connected. For example, when it is said in this specification that layers, regions, components, etc. are electrically connected, it refers to cases where the layers, regions, components, etc. are directly electrically connected, and / or cases where other layers, regions, components, etc. are interposed between them and are indirectly electrically connected.

[0038] The x-axis, y-axis, and z-axis are not limited to the three axes in the Cartesian coordinate system, but can be interpreted in a broader sense that includes them. For example, the x-axis, y-axis, and z-axis may be orthogonal to each other, but they can also refer to different directions that are not orthogonal to each other.

[0039] The advantages and features of the present disclosure, and methods for achieving them, will become clearer with reference to the embodiments described below in detail with the accompanying drawings. However, the present disclosure is not limited to the embodiments disclosed below and may be implemented in various different forms. These embodiments are provided solely to ensure that the disclosure is complete and to fully inform those skilled in the art of the scope of the present disclosure, and the present disclosure is defined solely by the scope of the claims.

[0040] The terminology used in this disclosure is for the purpose of describing embodiments only and is not intended to limit the present disclosure. In this disclosure, the singular may also include the plural unless specifically stated otherwise. The terms "comprises" and / or "comprising" as used herein do not exclude the presence or addition of one or more other components in addition to the mentioned components. Like reference numerals refer to like components throughout the disclosure, and "and / or" may include each and any combination of one or more of the mentioned components. Although "first", "second", etc. are used to describe various components, these components are not limited by these terms. These terms are only used to distinguish one component from another. Therefore, it should be understood that a first component mentioned below may also be a second component within the technical spirit of the present disclosure.

[0041] The word "exemplary" is used herein to mean "serving as an example or illustration." Any embodiment described herein as "exemplary" should not necessarily be construed as preferred or advantageous over other embodiments.

[0042] Embodiments of the present disclosure may be described in terms of a function or a block that performs a function. A block, which may be referred to as a "unit" or a "module" in the present disclosure, may be physically implemented by analog or digital circuits such as logic gates, integrated circuits, microprocessors, microcontrollers, memories, passive electronic components, active electronic components, optical components, hardwired circuits, etc., and may optionally be driven by firmware and software. Furthermore, the term "unit" as used in the disclosure refers to software, hardware elements such as FPGAs or ASICs, and the "unit" may perform certain roles. However, the "unit" is not limited to software or hardware. The "unit" may be configured to reside on an addressable storage medium and may be configured to execute one or more processors. Thus, as an example, a "part" may include elements such as software elements, object-oriented software elements, class elements, and task elements, processes, functions, attributes, procedures, subroutines, segments of program code, drivers, firmware, microcode, circuitry, data, databases, data structures, tables, arrays, and variables. The functionality provided within the elements and "parts" may be combined into a smaller number of elements and "parts" or further separated into additional elements and "parts."

[0043] Embodiments of the present disclosure can be implemented using at least one software program running on at least one hardware device and capable of performing network management functions to control elements.

[0044] Spatially relative terms such as "below," "beneath," "lower," "above," and "upper" may be used to readily describe the relationship between one component and other components as depicted in the drawings. Spatially relative terms may be understood to encompass different orientations of components during use or operation in addition to the orientations depicted in the drawings. For example, if a component depicted in the drawings were flipped over, a component described as "below" or "beneath" another component may end up "above" the other component. Thus, the exemplary term "below" may encompass both the above and below orientations. Components may also be oriented in other directions, and thus spatially relative terms may be interpreted accordingly.

[0045] Unless otherwise defined, all terms (including technical and scientific terms) used in this disclosure may be used with the meaning commonly understood by those skilled in the art to which this disclosure pertains. Furthermore, terms defined in commonly used dictionaries are not to be interpreted ideally or excessively unless explicitly and specifically defined otherwise.

[0046] Hereinafter, embodiments of the present disclosure will be described in detail with reference to the attached drawings. When describing with reference to the drawings, identical or corresponding components are given the same reference numerals and redundant descriptions thereof will be omitted.

[0047] FIG. 1 is a block diagram schematically illustrating internal components of an electronic device that provides prediction data on the properties of a composite according to a composition ratio using artificial intelligence according to an exemplary embodiment of the present disclosure.

[0048] An electronic device (100) (e.g., an electronic device for providing prediction data on the properties of a composite) may include a processor (110), a communication unit (120), a memory (130), and the like. The internal components that the electronic device (100) may include are not limited thereto. The electronic device (100) of the present disclosure may perform the functions of the processor (110) through a separate processing server or cloud server instead of the processor (110).

[0049] The electronic device (100) according to the embodiment may be a type of server, a central processing unit, an application providing server, etc. The server may include a device that provides data to other devices connected to the network via an application or the web. For example, the other devices may include devices such as desktops, laptops, tablets, and mobile terminals. In another example, the electronic device (100) may be a device encompassing electronic devices capable of performing the operations of the present disclosure.

[0050] Referring to FIG. 1, the processor (110) may be implemented to perform a data providing operation that predicts the properties of a composite according to the composition ratio using a memory (130) that stores data on an algorithm for controlling the operation of components within the electronic device (100) or a program that reproduces the algorithm, and the data stored in the memory (130). In this case, the processor (110) and the memory (130) may each be implemented as separate chips. Alternatively, the processor (110) and the memory (130) may also be implemented as a single chip.

[0051] The processor (110) can control one or more of the components discussed above in combination to implement various embodiments according to the present disclosure described in FIGS. 2 to 10 below in an electronic device (100).

[0052] The communication unit (120) according to the embodiment may include one or more components that enable communication with an external device, and may include, for example, at least one of a broadcast reception module, a wired communication module, a wireless communication module, a short-range communication module, and a location information module.

[0053] An input / output interface (not shown) according to an embodiment serves as a passageway for various types of external devices connected to the electronic device (100) of the present disclosure. This input / output interface may include at least one of a wired / wireless headset port, an external charger port, a wired / wireless data port, a memory card port, a port for connecting a device equipped with an identification module (SIM), an audio I / O (Input / Output) port, a video I / O (Input / Output) port, and an earphone port. The electronic device (100) of the present disclosure may perform appropriate control related to an external device connected to the input / output interface.

[0054] The memory (130) according to the embodiment can store data supporting various functions of the electronic device (100), a program for the operation of the processor (110), and input / output data (e.g., images, videos, etc.). The memory (130) can store a plurality of application programs (or applications) running on the electronic device (100), data for the operation of the electronic device (100), and commands. At least some of these application programs can be downloaded from an external server via wireless communication.

[0055] The memory (130) may include at least one type of storage medium among a flash memory type, a hard disk type, an SSD (Solid State Disk type), an SDD (Silicon Disk Drive type), a multimedia card micro type, a card type memory (e.g., SD or XD memory, etc.), a random access memory (RAM), a static random access memory (SRAM), a read-only memory (ROM), an electrically erasable programmable read-only memory (EEPROM), a programmable read-only memory (PROM), a magnetic memory, a magnetic disk, and an optical disk. In addition, the memory (130) may be a database that is separate from the electronic device (100) but is connected by wire or wirelessly.

[0056] At least one component may be added or deleted to correspond to the performance of the components illustrated in FIG. 1. Furthermore, it will be readily apparent to those skilled in the art that the relative positions of the components may be altered to correspond to the performance or structure of the device.

[0057] Meanwhile, the disclosed embodiments may be implemented in the form of a recording medium storing computer-executable instructions. The instructions may be stored in the form of program code, and when executed by a processor, may generate program modules to perform the operations of the disclosed embodiments. The recording medium may be implemented as a computer-readable recording medium.

[0058] Computer-readable storage media include all types of storage media that store instructions that can be deciphered by a computer. Examples include read-only memory (ROM), random access memory (RAM), magnetic tape, magnetic disks, flash memory, and optical data storage devices.

[0059] FIG. 2 is a block diagram schematically illustrating a system that provides prediction data on the properties of a composite according to a composition ratio according to an exemplary embodiment of the present disclosure.

[0060] Referring to FIG. 2, a system (hereinafter referred to as the "system") that provides data predicting the properties of a composite according to the composition ratio using artificial intelligence may include an electronic device (100) and an external device (200). For example, the electronic device (100) and the external device (200) may be connected to a network via a communication unit (e.g., the communication unit (120) of FIG. 1).

[0061] A system according to an embodiment may provide predicted data regarding the properties of a composite formed according to the composition ratio in an electronic device (100). For example, the electronic device (100) may augment the data by synthesizing the data using the property data of the composite formed according to the composition formulation, which is pre-stored in a memory (e.g., memory (130) of FIG. 1) or received from an external device (200). Here, the composite may have various properties even if it is the same composite depending on the ratio according to the content of the composition forming the composite. In other words, the properties of the composite may differ depending on the composition formulation. Specifically, the electronic device (100) may preprocess the property data of the composite and perform learning using artificial intelligence. More specifically, the electronic device (100) may augment input data, such as experimental data (e.g., property data), through a model that has completed learning. Through this, the system may provide predicted data regarding the properties of the composite according to the composition formulation corresponding to the result that the user desires to analyze, and data regarding the accuracy of the predicted data.

[0062] FIG. 3 is a schematic diagram illustrating a method for providing prediction data on the properties of a composite according to a composition ratio according to an exemplary embodiment of the present disclosure.

[0063] Referring to FIG. 3, an electronic device (e.g., the electronic device (100) of FIG. 1) can perform a process of learning a model and a process of augmenting data to provide prediction data on the properties of a composite formed according to the contents and ratios of compositions according to a composition formulation. Specifically, the electronic device can infer changes in the properties of a composite according to the composition formulation of the composite. More specifically, the electronic device can predict the properties of a composite according to an unlearned composition formulation by learning the property data of a composite according to a known composition formulation.

[0064] According to an embodiment, a processor (e.g., processor (110) of FIG. 1) may receive data related to at least one composition constituting a composite from an external device (e.g., external device (200) of FIG. 2) via a communication unit (e.g., communication unit (120) of FIG. 1). This may correspond to the electronic device collecting learning data (Data Collect) (310). The data related to at least one composition may be experimental data and / or synthetic data. For example, the processor may receive and confirm experimental data obtained by experimentally measuring physical property data according to a composition formulation of the composite. At this time, the experimental data may be included in the data related to at least one composition. For another example, the processor may confirm synthetic data synthesized by data augmentation of experimental data. At this time, the synthetic data may be included in the data related to at least one composition. In addition, the synthetic data may be data synthesized by the processor through learning based on the experimental data to be similar to the experimental data. That is, the processor can synthesize experimental data that is given in limited quantities to more accurately predict the properties of a composite, augment the synthetic data, and learn the experimental data and synthetic data together to use them to predict the properties of a composite.

[0065] According to an embodiment, the processor may select data having similar properties of the complex to be predicted and the type of composition constituting the complex. Thereafter, the processor may perform a preprocessing process on the selected data. The preprocessing process processes experimental data to facilitate learning of an artificial intelligence learning model (e.g., a property prediction model). The preprocessing process may be a process of selecting data having the same type and properties from a form in which the formulation information and property information of the composition are organized in separate tables.

[0066] Experimental data in an external device or memory may have a unique ID of a complex set to distinguish each formulation information. The processor may check discrete data on the composition and content of the experimental data to obtain preprocessed data as a first column and a first row. The processor may convert each composition into a first column and its content into a first row based on the unique ID of the complex. For example, the first column may be composed of PPO, HIPS, CNT, Clay, CF, GF, Antioxidant, and Coupling agent. The first row may be the first column corresponding values ​​corresponding to the unique ID.

[0067] The processor can obtain preprocessed data as a second column and a second row by checking discrete data for the properties and property values ​​of experimental data. For data in the property table, the processor can convert the property values ​​into a second row while setting the properties of the composition as the second column based on a unique ID. For example, the second column can be composed of flexural strength, flexural modulus, tensile strength, impact strength, and heat deflection temperature. The second row can be the corresponding values ​​of the second column corresponding to the unique ID.

[0068] After this, the processor can merge the first and second columns into a third column based on the unique ID of the complex, and can merge the first row and the second row into a third row. The preprocessed data can be sorted by the unique ID of the complex that the user is looking for. In other words, the processor can transform the form of the data frame by merging the first and second columns, and the first and second rows, based on the unique ID of the complex. The processor can perform learning through data related to the composition obtained by performing the preprocessing process, or synthesize and augment data through a Tabular Data Generative Model (TDGM). Here, the TDGM corresponds to a model that generates synthetic data having a distribution similar to the original data by sampling according to the distribution of experimental data corresponding to the original data, and can perform learning using the Wasserstein Distance. The TDGM can be set to use a model based on Tabular GAN or Variational Autoencoder.

[0069] According to an embodiment, a processor may perform a data representation (320) process that selects a property value of a composite to be predicted after learning through an artificial intelligence learning model and then defines a method of representing the data. For example, when predicting a property of a composite, such as flexural strength, the processor may classify data on the composition and the content of the composition according to the composition formulation of the composite into input data features, which are input data sets of the artificial intelligence learning model. At this time, features corresponding to the input data set of the artificial intelligence learning model may be encoded to perform a regression or classification process of the property to be predicted through a feature transformer. The encoded result may be generated as a decision output through an activation function layer, and the decision outputs of individual domains may be finally derived as a property value to be predicted through an FC layer (Fully Connected Layer).

[0070] According to an embodiment, a processor may perform a knowledge transfer (330) process by assigning weights to individual features using an artificial intelligence learning model. The knowledge transfer process may be a process of extracting a prediction result of a composite material property by configuring individual domains into fully connected layers through a specific activation function layer. Thereafter, the processor may evaluate the accuracy of the artificial intelligence learning model's performance in predicting composite material properties by comparing the predicted result of the composite material property with the ground truth. This may be referred to as a performance evaluation (340) process.

[0071] FIG. 4 is a flowchart schematically illustrating a method for providing prediction data on the properties of a composite according to a composition ratio according to an exemplary embodiment of the present disclosure.

[0072] Referring to FIG. 4, a processor (e.g., processor (110) of FIG. 1) may provide a user with a specific property of a complex desired to be identified based on the results learned using an artificial intelligence learning model. In other words, the processor may provide the user with a predicted property value based on the composition ratio of a specific complex through a property prediction model.

[0073] In step S100, the processor can identify data related to at least one composition constituting the composite. The data related to at least one composition may be experimental data and / or synthetic data. The synthetic data may be data synthesized by the processor through learning based on experimental data to resemble the experimental data. That is, the processor can synthesize limited experimental data to augment the synthetic data for more accurate prediction of the physical properties of the composite, and can utilize the experimental data and synthetic data together for learning to predict the physical properties of the composite.

[0074] In an embodiment, data extraction for constructing experimental data may be accomplished by extracting paragraphs related to composition formulations from text in a document and extracting data related to composition formulations from a table. In this case, if the text is extracted from a PDF document, an OCR-based model (e.g., Nougat) can be used to convert the text within the PDF document into an mmd file for Python recognition, and the text can then be extracted. The extracted text can be classified into classes related to composition formulations within the input sentences using an object name recognition model (e.g., MaterialsBERT). In this case, if a sentence contains information on the composition content or physical properties of the composition formulation, the corresponding paragraph can be stored. In addition, experimental data can be extracted by adaptively adjusting the size of the bounding box in recognizing the table structure and sharing an attention module to enable cooperation between related tasks. The processor can check the experimental data extracted in this way from a database of a memory (e.g., memory (130) of FIG. 1) or an external device (e.g., external device (200) of FIG. 2).

[0075] In step S200, the processor may encode the physical property data of the composite synthesized based on data related to at least one composition into first physical property data. At this time, the processor may set the physical property information of the composite to be predicted through an artificial intelligence learning model corresponding to the first model. That is, the process of setting the physical property information may correspond to the process of encoding the first physical property data corresponding to individual features. The processor may perform batch normalization by expressing the data related to the composition of the composite to confirm the set physical property information as individual features.

[0076] In step S300, the processor can generate an embedding result corresponding to the first material data through the first model. The processor can perform masking on individual features of the first material data and generate masked features corresponding to the masked results.

[0077] In step S400, the processor may provide property prediction data of the composite for the second property data based on the produced embedding results. The processor may provide a result predicting the property of the composite according to a composition ratio that has not been previously learned by learning previously known data (e.g., experimental data) and / or data similar to previously known data (e.g., synthetic data) through an artificial intelligence learning model. In addition, the processor may additionally perform an accuracy evaluation of the artificial intelligence learning model used in the property prediction result. Through this, the processor may update the model to improve the accuracy of the artificial intelligence learning model. In other words, the processor may perform an update by relearning data closer to the correct answer to improve the accuracy of the artificial intelligence learning model.

[0078] FIG. 5 is a flowchart schematically illustrating a process for verifying data related to a composite composition according to an exemplary embodiment of the present disclosure.

[0079] In step S110, a processor (e.g., processor (110) of FIG. 1) may identify data related to a composition constituting the complex. The data related to the composition constituting the complex may include experimental data and synthetic data. Additionally, the data related to the composition constituting the complex may include the type, content, and formulation data of at least one composition.

[0080] In step S120, the processor may perform a sampling process based on experimental data and synthetic data. Here, sampling may be a process by which the processor constructs an input data set for an artificial intelligence learning model based on the experimental data and synthetic data. In other words, the processes of extracting and preprocessing table-type data to construct an input data set for an artificial intelligence learning model and extracting the data may be referred to as a sampling process.

[0081] In step S130, the processor may generate an input data set for an artificial intelligence learning model. The artificial intelligence learning model may utilize individual features or data encoding individual features related to the properties of the composite through a sampling process as the input data set. The processor may perform learning using individual features related to the properties of the composite as the input data set and the results of predicting the properties of the composite according to the composition ratio of the composite as the output data set.

[0082] FIG. 6 is a block diagram schematically illustrating a process for encoding and embedding data for predicting properties of a composite according to a composition ratio according to an exemplary embodiment of the present disclosure.

[0083] The AI ​​learning model according to the embodiment may be a first model. For example, the AI ​​learning model may be a model for predicting the properties of a composite and may be a deep learning-based network designed for learning tabular data (e.g., tabular data). In this case, the first model may have an autoencoder structure consisting of an encoder and a decoder that reflect the characteristics of selecting variables in a tree-based model. The first model may receive tabular data containing composition formulation data and property data without separate preprocessing, perform learning, and perform classification and regression.

[0084] The first model according to the embodiment can be configured to use a method of assigning weights to individual features rather than selecting only specific features. That is, the first model can perform learning through steps for multiple domains by masking features from input tabular data. At this time, the importance of features in each domain is identified for each step, and masking is performed to select only the masked features for learning, thereby improving prediction performance. The input features are encoded through a feature transformer. A processor (e.g., processor (110) of FIG. 1) can check the decision outputs of individual domains through an activation function layer based on importance and produce the final decision output as an embedding result.

[0085] A processor according to an embodiment may perform a batch normalization process (610). The processor may perform batch normalization on features of property data input to an artificial intelligence learning model and perform encoding on individual features.

[0086] A processor according to an embodiment may generate a decision output at a step corresponding to an individual domain by passing the encoded result through an activation function layer. This may correspond to 620 and 630 in FIG. 6. For example, the processor may identify a first decision output in a first domain among the individual domains. For another example, the processor may identify a second decision output in a second domain among the individual domains based on the first decision output. For yet another example, the processor may extract second material property data from a fully connected layer that connects output terminals of individual domains including the first domain and the second domain (640). This may correspond to a process in which the processor generates an overall decision embedding by summing the results of decision outputs of individual steps of individual domains, and passes the embedding through a fully connected layer to ultimately produce a classification or regression result.

[0087] According to an embodiment, a processor may utilize a transformer layer utilizing attention as a mechanism for calculating weights in an output sequence for each position in an input sequence. An attentive transformer layer can assign importance to various positions in an input sequence, thereby helping an artificial intelligence learning model understand context and focus on relevant information. This attention mechanism can help model the relationship between the input sequence and the output sequence.

[0088] After this, the processor can generate a trainable mask through a mask layer. A trainable mask can refer to a parameterized mask containing weights that are updated during the training of an AI learning model. This mask can be adjusted during the training process of the AI ​​learning model and can be used to adjust the importance of specific inputs or outputs for specific purposes. For example, a mask used in an attention mechanism allows an AI learning model to determine how much attention to pay to a specific input. The mask can be adjusted during model training to help the model better adapt to the training data. For another example, a trainable mask can be used to dynamically adjust the weights for specific layers or features. The mask contains information on which features to utilize. Multiplying the mask and the feature can perform feature selection. Feature selection can be performed using masked features generated by multiplying the feature of the previous domain step by the mask. At this point, the processor can calculate the importance of features using a learnable mask, multiply it by the features of the previous domain, and generate a masked feature. This masked feature can then be used to calculate a prior scale term to be applied to the mask of the next domain. The prior scale term indicates the extent to which individual features were utilized in the previous domain, and can reduce the reflection rate of variables selected in the previous domain.

[0089] For example, a processor can calculate the importance of individual features by multiplying the input features of an AI learning model by a mask. This allows the processor to select important features for use when performing regression or classification. This corresponds to a feature selection process. For example, if an AI learning model receives three features as input, the processor can select important features by multiplying each feature by a mask, such as [0.1, 0.2, 0.7], in an Attitudinal Transformer and performing training. By representing data as masked features, which are obtained by multiplying features by masks, the selection of important features and knowledge transfer can be facilitated.

[0090] FIG. 7 is an exemplary diagram schematically illustrating a process for encoding data for predicting physical properties of a composite according to a composition ratio according to an exemplary embodiment of the present disclosure.

[0091] In step S210, a processor (e.g., processor (110) of FIG. 1) may set property information of a complex to be predicted through a first model. The processor sets property information for generating an input data set of the first model based on data related to the composition.

[0092] In step S220, the processor may perform batch normalization by representing data related to the composition as individual features to verify the set physical property information. The result of this batch normalization can be generated and verified by the processor as an input data set, corresponding to the encoded result. The input data set may be characterized as having been preprocessed as an input data set for training the first model.

[0093] FIG. 8 is a flowchart schematically illustrating a process for embedding data for predicting properties of a composite according to a composition ratio according to an exemplary embodiment of the present disclosure.

[0094] In step S310, a processor (e.g., processor (110) of FIG. 1 ) may generate a masked feature. The processor generates the masked feature by masking individual features of the first material data. The processor may perform learning by multiplying each feature by a mask whose sum totals to 1.

[0095] In step S320, the processor can determine the importance of individual features. The processor determines the importance by calculating a scale term based on the masked features. That is, in response to calculating the prior scale term from the masked features, the processor can determine the importance of individual features.

[0096] In step S330, the processor can verify the decision outputs of individual domains. The processor verifies the decision outputs of individual domains through an activation function layer based on importance. An activation function may refer to a function that determines the output of a neuron in an artificial intelligence network. Each neuron can multiply an input value by a weight, add a bias, and then pass this result to the activation function to generate an output value. This output value can be passed to the neurons in the next layer or the final output to enable model prediction or learning. The activation function can mainly introduce non-linearity to help the artificial intelligence learning model learn complex functions.

[0097] In step S340, the processor may produce an embedding result through a fully connected layer. The final decision output of each domain may be derived as an embedding result corresponding to the first property data. That is, the processor may provide second property data, at least partially different from the first property data, from the fully connected layer connecting the individual domains.

[0098] FIG. 9a and FIG. 9b are exemplary diagrams of prediction data of physical properties of a composite according to a composition ratio corresponding to an activation function according to an exemplary embodiment of the present disclosure.

[0099] A processor (e.g., processor (110) of FIG. 1) may utilize a Swish function (920) and / or a Mish function (930) instead of a Rectified Linear Unit (ReLU) function (910) as an activation function. The ReLU function is a nonlinear function that is set to output the input value as is when the input value is positive, and to output 0 when the input value is negative. Although it may have computational efficiency and thus a fast learning speed, there may be a problem that the gradient may become 0 when the input value is negative, resulting in the loss of neurons.

[0100] The Swish function according to the embodiment can be expressed as the mathematical formula (1) below.

[0101] Mathematical formula (1)

[0102] Swish(x)=x*σ(βx)

[0103] In the above mathematical equation (1), σ is a sigmoid function, and β is a new hyperparameter of the Swish function. The Swish function has a similar form to the ReLU function, but has the characteristic of changing into a smoother form.

[0104] The Mish function according to the embodiment can be expressed as the mathematical formula (2) below.

[0105] Mathematical formula (2)

[0106] Mish(x)=x*tanh(softplus(x))

[0107] In the above mathematical expression (2), softplus(x) represents log(1+e^x), and the Mish function can have a smooth form while having nonlinearity.

[0108] Referring to Fig. 9a, the processor can extract more accurate output values ​​from negative input values ​​by utilizing the Swish function and / or the Mish function as activation functions instead of the ReLU function in the first model. The processor can use each of the Swish function layers and the Mish function layers to predict the properties of the composite using all of the first models that utilize each, and perform knowledge transfer with a model that shows higher accuracy. That is, the processor can train the first model by setting the activation function to a two-track layer, and then select a learning model with higher accuracy among the properties predicted by the second model trained through the Swish function or the third model trained through the Mish function to perform knowledge transfer.

[0109] According to an embodiment, a processor can improve the accuracy of predicting the properties of a composite by replacing the ReLU function with a Swish function and / or a Mish function. Referring to an example diagram (940) of Fig. 9b, property prediction data (e.g., second property data) using the formulation data of mPPO can be confirmed. The example diagram (940) of Fig. 9b corresponds to the results of predicting 25 test sets using a total of 1057 data using 57 formulation data and 1000 synthetic formulation data to train a property prediction model (e.g., first model). For each activation function, the mean absolute error (mae) and the mean absolute percentage error (mae) were used as indices for evaluating the prediction performance of the model, and the f1-score can also be used. For three properties of flexural strength, flexural modulus, and heat deflection temperature, the result using the Swish function showed the lowest error rate in the example diagram (940). For the two properties of tensile strength and impact strength, the result using the Mish function in the example (940) showed the lowest error rate.

[0110] FIG. 10 is an exemplary diagram (1000) schematically illustrating a performance evaluation result screen of the physical property prediction data of a composite according to a composition ratio according to an exemplary embodiment of the present disclosure.

[0111] Referring to FIG. 10, a processor (e.g., processor (110) of FIG. 1) can derive a performance evaluation result for a final characteristic prediction model. The processor can verify the ground truth for the second property data. Based on the ground truth, the processor can evaluate the accuracy of the second property data. The second property data may be data on the properties of a composite having a ground truth, which is at least partially different from the first property data. For example, if the flexural strength value for a specific composition formulation of mPPO is known as the ground truth, and the processor has not performed learning on it, the processor can verify to what extent the second property data corresponding to the flexural strength of mPPO predicted through learning differs from the known ground truth. The verification of this difference can be regarded as a verification according to the derivation of a performance evaluation result for the final characteristic prediction model.

[0112] Referring to Fig. 10, the x-axis (1010) may be strain (%), and the y-axis (1020) may be stress (MPa). That is, the processor may compare the ground truth value (1030) and the predicted value (1040) of the property value of the composite corresponding to the composition formulation while placing the property values ​​to be predicted on the x-axis (1010) and the y-axis (1020). That is, by checking the error rate of the ground truth value (1030) and the predicted value (1040), the processor may determine which artificial intelligence learning model is better. The artificial intelligence learning model here may be a second model learned by the first model using the Swish function, or a third model learned by the first model using the Mish function. The processor may provide the user with a screen showing the prediction performance results of the artificial intelligence learning model, as shown in Fig. 10.

[0113] According to an embodiment, the processor may select or update an artificial intelligence learning model by determining which activation function layer is advantageous for learning. For example, the processor may adopt an artificial intelligence learning model utilizing a Swish function based on the performance evaluation result and provide the user with property prediction data. For another example, the processor may adopt an artificial intelligence learning model utilizing a Mish function based on the performance evaluation result and provide the user with property prediction data. For yet another example, the processor may adopt an artificial intelligence learning model utilizing a Swish function for some property predictions and an artificial intelligence learning model utilizing a Mish function for some property predictions based on the performance evaluation result and provide the user with property prediction data. The processor may be able to selectively perform learning as described above because learning is performed individually for each property of the composite during the learning process.

[0114] In one embodiment, the processor can generate synthetic data similar to published experimental data to overcome the difficulty of acquiring a minimum amount of data for training a model predicting the properties of a composite due to limitations in the amount of publicly available experimental data. This may correspond to a process of preprocessing experimental data structured in a table format through the processor, as described above, and then generating data similar to the experimental data to augment the data.

[0115] In one embodiment, the processor can perform reverse engineering to accelerate formulation optimization based on target properties for optimizing the formulation of a composite. The processor can utilize an actor-critic reinforcement learning model to generate a new composite formulation suited to the target properties from an existing composite formulation.

[0116] While the present disclosure has primarily focused on electronic devices, it is not limited thereto. For example, a method for manufacturing such electronic devices is also within the scope of the present disclosure.

[0117] While this disclosure has been described with reference to the embodiments illustrated in the drawings, these are merely exemplary, and those skilled in the art will appreciate that various modifications and equivalent alternative embodiments are possible. Therefore, the true scope of technical protection of this disclosure should be determined by the technical spirit of the appended claims.

Claims

1. A method for providing prediction data on the properties of a composite according to the composition ratio using artificial intelligence, A step of identifying data related to at least one composition constituting the complex through a processor of an electronic device; A step of encoding the physical property data of the composite synthesized based on data related to the at least one composition through the processor into first physical property data; A step of using the above processor to produce an embedding result corresponding to the first property data through a first model corresponding to an artificial intelligence learning model based on a neural network; and A step of providing property prediction data of the complex for the second property data based on the embedding result produced through the processor, The step of producing an embedding result corresponding to the above first property data is: A step of generating a masked feature by masking individual features of the first material data using the above processor; A step of calculating a scale term based on the masked feature through the processor to determine the importance for the individual feature; A step of confirming the decision output of individual domains through an activation function layer based on the importance through the above processor; and A method for providing composite material property prediction data, comprising a step of using the above processor to produce the final decision output of the individual domains as an embedding result corresponding to the first material property data through a fully connected layer.

2. In paragraph 1, The step of checking data related to at least one composition is: A step of confirming experimental data for the complex and synthetic data for the complex through the processor; and A step of sampling data related to the at least one composition based on the experimental data and the synthetic data through the processor, Data relating to at least one of the above compositions, A method for providing composite material property prediction data, characterized in that it includes data on the type, content, and formulation of at least one composition constituting the composite.

3. In paragraph 2, Further comprising a step of confirming data related to at least one composition sampled through the processor as an input data set, The above input data set is characterized in that it is preprocessed as an input data set for learning of the first model, A method for providing composite material property prediction data, characterized in that the synthetic data is synthesized through data augmentation from the experimental data.

4. In paragraph 1, The step of encoding the above first property data is: A step of setting the physical property information of the complex to be predicted through the first model using the above processor; and A method for providing composite material property prediction data, comprising a step of performing batch normalization by expressing data related to the composition of the composite to confirm the set material property information through the processor as individual features.

5. In paragraph 1, The step of checking the decision output of the above individual domains is: A step of confirming a first decision output in a first domain among the individual domains through the processor; and A step of confirming a second decision output in a second domain among the individual domains based on the first decision output through the processor, A method for providing composite material property prediction data, wherein the individual domains including the first domain and the second domain are characterized by being composed of the fully connected layer.

6. In paragraph 1, The step of providing property prediction data of the above complex is: A method for providing composite material property prediction data, comprising the step of providing second material property data, at least part of which is different from first material property data learned through the first model, using the above processor.

7. In paragraph 6, The step of providing property prediction data of the above complex is: A step of confirming the ground truth for the second property data through the processor; and A method for providing composite material property prediction data, comprising a step of evaluating the accuracy of the second material property data based on the ground truth through the processor.

8. In an electronic device that provides prediction data on the properties of a composite according to the composition ratio using artificial intelligence, Department of Communications; memory; and At least one processor electrically connected to the communication unit and the memory, At least one processor, Identify data relating to at least one composition comprising the complex, Encoding the physical property data of the composite synthesized based on data related to at least one composition as first physical property data, Generate masked features for individual features of the first property data through a first model corresponding to an artificial intelligence learning model based on a neural network, The importance of each feature is determined by calculating a scale term based on the masked feature. Based on the above importance, the decision output of each domain is checked through the activation function layer, The final decision output of the above individual domains is used to produce an embedding result corresponding to the first property data through a fully connected layer, An electronic device providing property prediction data, which is set to provide property prediction data of the complex for the second property data based on the above-mentioned produced embedding results.

Images