Artificial intelligence-based engineering prediction method, device, and system

Abstract

An embodiment of the present disclosure may provide an electronic device comprising: a memory implemented to store a plurality of instructions; and at least one processor electronically connected to the memory and configured to execute at least some of the plurality of instructions stored in the memory, wherein the at least one processor is configured to perform the operations of: obtaining an input for a scene including at least one object; extracting a condition vector corresponding to the scene on the basis of the input for the scene; and predicting a behavior of the at least one object at a first time point on the basis of the condition vector by using a pre-trained artificial intelligence model.

Description

AI-based engineering prediction methods, devices, and systems

[0001] The present disclosure relates to an artificial intelligence-based engineering prediction method, device and system, and more particularly, to an apparatus and method for efficiently predicting a physical field for a three-dimensional (3D) shape for a scene including at least one object.

[0002] CAE simulation has become an essential element in predicting physical performance during product design and development. However, problems such as fluid flow analysis and structural analysis of complex 3D models require enormous computational costs and time.

[0003] To address these issues, active research is underway on surrogate models that leverage AI technology to rapidly predict simulation results. Early AI models focused on predicting single scalar values, such as the drag coefficient or maximum displacement, but these models provided limited information for understanding the overall physical phenomena of a product.

[0004] Recently, attempts have been made to directly predict 3D fields, such as pressure or displacement distribution across a product. However, existing methods utilizing models such as 3D convolutional neural networks (3D CNNs) or graph neural networks (GNNs) require massive amounts of high-quality training data and high computing resources, making practical applications difficult.

[0005] Furthermore, existing simulation analysis techniques suffer from the significant time and cost involved in predicting object dynamics over time in complex multi-object situations. For example, simulations such as crash tests require high-performance computing resources to accurately model and analyze interactions between multiple objects, limiting real-time or rapid predictions. These issues can reduce efficiency and increase costs, particularly during the design and testing processes.

[0006] Furthermore, engineering data is becoming increasingly complex and is increasingly composed of 3D data rather than 2D data. However, existing field prediction methods based on 2D data have limitations in extending to 3D field prediction. Because 3D data contains higher-dimensional information, more sophisticated and scalable approaches are required to effectively process and analyze it. These limitations are a major factor limiting the applicability of existing technologies in the modern engineering environment, where 3D data is essential.

[0007] Prior art literature information related to this is as follows.

[0008] (Patent Document 0001) Japanese Patent Publication No. 2011-221896 (November 4, 2011)

[0009] One object of the present disclosure is to provide a method and device for predicting a 3D field using implicit neural representations to predict a 3D field against a 3D shape.

[0010] In addition, one task of the present disclosure is to solve the practical problem of difficulty in securing a large amount of high-quality CAE simulation data for 3D shapes, and to provide an artificial intelligence-based device and method that can predict 3D physical fields quickly and accurately at low cost.

[0011] In addition, one object of the present disclosure is to provide a dynamic prediction method based on implicit neural representation, which extracts a condition vector for a scene including at least one object and predicts the dynamic behavior of the object based on the condition vector, and an electronic device implementing such a method.

[0012] The tasks of the present disclosure are not limited to the tasks mentioned above, and other tasks not mentioned will be clearly understood by those skilled in the art from the description below.

[0013] An electronic device according to one embodiment of the present disclosure includes a memory configured to store a plurality of instructions, at least one processor electronically connected to the memory and configured to execute at least some of the plurality of instructions stored in the memory, wherein the at least one processor may be configured to perform an operation of receiving three-dimensional shape information and extracting a latent vector representing a characteristic of the shape information, an operation of predicting a first field value at the first coordinate through a pre-trained low-fidelity neural network using the latent vector and the first coordinate as inputs, an operation of calculating a correction term for the first field value through at least one high-fidelity neural network using the latent vector, the first coordinate, and the predicted first field value as inputs, and an operation of predicting a second field value based on the first field value and the correction term.

[0014] In one embodiment of the present disclosure, the electronic device may include an operation of extracting the latent vector, an operation of converting the three-dimensional shape information into a three-dimensional grid in the form of an encoded distance function, and an operation of extracting the latent vector from the three-dimensional grid using a three-dimensional convolutional neural network.

[0015] In one embodiment of the present disclosure, the electronic device may be configured such that the low-fidelity neural network is trained based on a low-fidelity data set and the at least one high-fidelity neural network is trained based on a high-fidelity data set. The low-fidelity data set may be obtained from a model discretized into a grid having a first resolution, and the high-fidelity data may be obtained from a model discretized into a grid having a second resolution higher than the first resolution.

[0016] In one embodiment of the present disclosure, the electronic device may be further configured to perform an operation of generating a learning data set for training the low-fidelity neural network or the at least one high-fidelity neural network. The learning data set may include a plurality of coordinate values ​​and a plurality of label values ​​corresponding to each of the coordinate values. Each of the label values ​​may be determined by identifying a preset number of nearest nodes among a plurality of nodes included in the low-fidelity data set or the high-fidelity data set based on the corresponding coordinate value and calculating an average of field values ​​of the identified nearest nodes.

[0017] In one embodiment of the present disclosure, the electronic device may include a first high-fidelity neural network that models a linear relationship for an input and a second high-fidelity neural network that models a nonlinear relationship. The operation of calculating the correction term may include the operation of calculating the correction term based on a first output value output from the first high-fidelity neural network and a second output value output from the second high-fidelity neural network.

[0018] In one embodiment of the present disclosure, the electronic device may be trained based on an integrated loss function that includes both a prediction error for low-fidelity data and a prediction error for high-fidelity data, wherein the low-fidelity neural network and the at least one high-fidelity neural network are trained.

[0019] In one embodiment of the present disclosure, the electronic device may include an operation of predicting the second field value by adding the correction term to the first field value to obtain the second field value.

[0020] In one embodiment of the present disclosure, the electronic device may further include a mapping neural network that converts input variables of high-fidelity data used for training the high-fidelity neural network into input variables of low-fidelity data used for training the low-fidelity neural network. The variables converted by the mapping neural network may be provided as inputs to the low-fidelity neural network.

[0021] In one embodiment of the present disclosure, the electronic device may be configured such that the at least one high-fidelity neural network learns a residual, which is a difference between the first field value and the second field value. The learned residual may be output as the correction term.

[0022] In addition, according to one embodiment of the present disclosure, an electronic device may be provided, including a memory configured to store a plurality of instructions and at least one processor electronically connected to the memory and configured to execute at least some of the plurality of instructions stored in the memory, wherein the at least one processor is configured to perform an operation of obtaining an input for a scene including at least one object, an operation of extracting a condition vector corresponding to the scene based on the input for the scene, and an operation of predicting a behavior of the at least one object at a first time point based on the condition vector using a pre-learned artificial intelligence model.

[0023] The input to the scene may include at least one of an input associated with a shape of the at least one object, an input associated with a motion of the at least one object, an input associated with a physical property of the at least one object, or an input associated with a physical quantity of the at least one object.

[0024] The operation of extracting a condition vector may include an operation of classifying an input for the scene into a first input and a second input, inputting the first input into a first extractor, extracting a first vector corresponding to a first characteristic associated with a shape, extracting a second vector corresponding to a second characteristic associated with a movement, and extracting the condition vector based on the first vector and the second vector.

[0025] The first extractor extracts a first vector by mapping a first input corresponding to a shape of the at least one object to a latent space, and the second extractor extracts a second vector by adjusting a second input corresponding to a kinematic parameter of the at least one object to a predetermined scale, and can extract the condition vector by merging the second vector with the first vector.

[0026] The operation of extracting a condition vector includes an operation of obtaining a latent vector based on a 3D shape of at least one object and an operation of extracting a condition vector corresponding to the scene based on the latent vector, and the at least one processor may be set to input the condition vector and position coordinates of nodes corresponding to the 3D shape at the first point in time into the pre-learned artificial intelligence model.

[0027] The predicting action may include an action of predicting a first physical quantity of a first point of the at least one object at the first time point based on the condition vector, and an action of predicting a behavior of the at least one object at the first time point based on the first physical quantity.

[0028] The predicting action may include an action of predicting field values ​​in the global region of the at least one object at the first time point based on the condition vector and an action of predicting a behavior of the at least one object at the first time point based on the field values.

[0029] At least one processor may perform an operation of predicting a behavior of the at least one object at a second time based on a behavior of the at least one object at the first time.

[0030] The pre-trained artificial intelligence model is trained based on simulation data, and the interval between the first time point and the second time point may correspond to a time stamp of the simulation data.

[0031] At least one processor may be further configured to perform an operation of selecting at least one key coordinate on a global coordinate included in the scene based on a predetermined criterion, and may be configured to change a time interval for time-series prediction when inputting the key coordinate into the pre-trained artificial intelligence model.

[0032] The solutions to the problems of the present invention are not limited to the solutions described above, and solutions not mentioned can be clearly understood by a person having ordinary skill in the art to which the present invention pertains from this specification and the attached drawings.

[0033] According to one embodiment of the present disclosure, the learning cost of the entire model can be drastically reduced by using a large amount of low-fidelity data with low computational cost together with a small amount of high-fidelity data with high cost.

[0034] In addition, according to one embodiment of the present disclosure, high prediction accuracy can be achieved by effectively learning complex correlations between two data through a residual learning structure that corrects errors based on low-fidelity prediction values.

[0035] Additionally, according to one embodiment of the present disclosure, a coordinate-based implicit neural network can be utilized to predict field values ​​at arbitrary coordinates without being dependent on a specific grid, thereby enabling generation of a high-resolution continuous field map.

[0036] According to one embodiment of the present disclosure, by predicting the behavior of an object using a pre-learned artificial intelligence model, the time and cost required for predicting dynamics over time in a situation involving multiple objects can be reduced.

[0037] In addition, according to one embodiment of the present disclosure, by enabling field prediction of engineering data composed of 3D data, the problem of insufficient scalability of existing 2D data field prediction methods can be solved.

[0038] The effects of the present invention are not limited to the effects described above, and effects not mentioned can be clearly understood by a person having ordinary skill in the art to which the present invention pertains from this specification and the attached drawings.

[0039] FIG. 1 is a diagram illustrating a configuration of an electronic device according to various embodiments.

[0040] FIG. 2 is a diagram for explaining a field prediction method using a field prediction system according to various embodiments.

[0041] FIG. 3 is a flowchart illustrating a field prediction method using a field prediction system according to various embodiments.

[0042] FIG. 4 is a flowchart illustrating a method for extracting a latent vector according to various embodiments.

[0043] Figure 5 is a diagram illustrating an example of low-fidelity data and high-fidelity data.

[0044] Figure 6 is a diagram illustrating an example of a low-fidelity prediction result and a high-fidelity prediction result.

[0045] FIG. 7 is a diagram illustrating an example of a multi-fidelity neural network-based field prediction system according to various embodiments.

[0046] FIG. 8 is a diagram illustrating an example of a field prediction system including a mapping neural network according to various embodiments.

[0047] FIG. 9 is a diagram for explaining a specific implementation method of a multi-fidelity neural network according to various embodiments.

[0048] FIG. 10 is a diagram illustrating a method for an electronic device to perform dynamic field prediction using an artificial intelligence model according to various embodiments.

[0049] FIG. 11 is a diagram illustrating a method for an electronic device to extract a condition vector according to various embodiments.

[0050] FIG. 12 is a diagram illustrating an example of an electronic device extracting a condition vector according to various embodiments.

[0051] FIG. 13 is a diagram illustrating an example of an electronic device predicting the behavior of an object according to various embodiments.

[0052] FIG. 14 is a diagram illustrating another example of an electronic device predicting the behavior of an object according to various embodiments.

[0053] FIG. 15 is a diagram illustrating an exemplary method for an electronic device to perform prediction according to various embodiments.

[0054] FIG. 16 is a diagram illustrating another exemplary method for an electronic device to perform prediction according to various embodiments.

[0055] Figure 17 shows various examples of 3D data.

[0056] Figure 18 shows various examples of sampling coordinates of 3D data.

[0057] Hereinafter, embodiments of the present disclosure will be described in detail with reference to the accompanying drawings. In describing the embodiments, descriptions of technical details that are well-known in the technical field to which the present disclosure pertains and are not directly related to the present disclosure will be omitted. This is to avoid obscuring the gist of the present disclosure by omitting unnecessary explanations and to convey the gist more clearly.

[0058] Since the embodiments described in this specification are intended to clearly explain the idea of ​​the present invention to a person having ordinary skill in the art to which the present invention pertains, the present invention is not limited to the embodiments described in this specification, and the scope of the present invention should be interpreted to include modified or altered examples that do not depart from the idea of ​​the present invention.

[0059] The terms used in this specification have been selected from widely used terms, taking into account the functions of the present invention. However, these terms may vary depending on the intentions of those skilled in the art, precedents, or the emergence of new technologies. However, if a specific term is defined and used with an arbitrary meaning, the meaning of that term will be described separately. Therefore, the terms used in this specification should be interpreted based on the actual meaning of the term and the overall content of this specification, rather than simply the name of the term.

[0060] The drawings attached to this specification are intended to facilitate explanation of the present invention, and the shapes depicted in the drawings may be exaggerated as necessary to help understanding of the present invention, and therefore the present invention is not limited by the drawings.

[0061] In this specification, each of the phrases "A or B", "at least one of A and B", "at least one of A or B", "A, B, or C", "at least one of A, B, and C", and "at least one of A, B, or C" may include any one of the items listed together in that phrase, or all possible combinations thereof.

[0062] If a detailed description of the composition or function of a known disclosure related to the present invention in this specification is deemed to obscure the gist of the present invention, a detailed description thereof will be omitted as necessary. Furthermore, the numbers (e.g., "first," "second," etc.) used throughout the description of this specification are merely identifiers used to distinguish one component from another.

[0063] In addition, the suffixes "part" and "part" for components used in the following description are given or used interchangeably only for the convenience of writing the specification, and do not have distinct meanings or roles in themselves.

[0064] That is, the embodiments of the present disclosure are provided to make the present disclosure complete and to inform those skilled in the art of the scope of the present disclosure, and the invention of the present disclosure is defined solely by the scope of the claims. Like reference numerals refer to like elements throughout the specification.

[0065] Terms such as “first” and / or “second” may be used to describe various components, but the components should not be limited by the terms. The terms are only for the purpose of distinguishing one component from another, for example, a first component may be referred to as a second component, and similarly, a second component may also be referred to as a first component, without departing from the scope of the present disclosure.

[0066] When a component is referred to as being "connected" or "connected" to another component, it should be understood that it may be directly connected or connected to that other component, but that there may be other components in between. Conversely, when a component is referred to as being "directly connected" or "directly connected" to another component, it should be understood that there are no other components in between. Other expressions that describe the relationship between components, such as "between" and "directly between" or "adjacent to" and "directly adjacent to", should be interpreted similarly.

[0067] Each block of the processing flow diagrams in the drawings and combinations of the flow diagrams can be performed by computer program instructions. These computer program instructions can be installed in a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing equipment, so that the instructions executed by the processor of the computer or other programmable data processing equipment create a means for performing the functions described in the flow diagram block(s). These computer program instructions can also be stored in a computer-available or computer-readable memory that can direct a computer or other programmable data processing equipment to implement the functions in a specific manner, so that the instructions stored in the computer-available or computer-readable memory can also produce an article of manufacture that includes an instruction means for performing the functions described in the flow diagram block(s). Since the computer program instructions may be installed on a computer or other programmable data processing device, a series of operational steps may be performed on the computer or other programmable data processing device to create a computer-executable process, so that the instructions that cause the computer or other programmable data processing device to perform the steps for performing the functions described in the flowchart block(s) may also be able to provide steps for performing the functions described in the flowchart block(s).

[0068] Additionally, a device-readable storage medium may be provided in the form of a non-transitory storage medium. Here, "non-transitory" simply means that the storage medium is a tangible device and does not contain signals (e.g., electromagnetic waves). This term does not distinguish between cases where data is stored semi-permanently or temporarily on the storage medium.

[0069] Additionally, each block may represent a module, segment, or portion of code that contains one or more executable instructions for performing a specified logical function(s). It should also be noted that in some alternative implementation examples, the functions mentioned in the blocks may occur out of order. For example, two blocks shown in succession may in fact be performed substantially concurrently, or the blocks may sometimes be performed in reverse order, depending on their respective functions. For example, the operations performed by a module, program, or other component may be performed sequentially, in parallel, iteratively, or heuristically, or one or more of the operations may be performed in a different order, omitted, or one or more additional operations may be added.

[0070] The term "unit" as used herein refers to a software or hardware component such as a Field Programmable Gate Array (FPGA) or an Application Specific Integrated Circuit (ASIC). The "unit" performs specific roles but 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. Accordingly, according to some embodiments, the "unit" includes components such as software components, object-oriented software components, class components, and task components, 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 components and "units" may be combined into a smaller number of components and "units" or further separated into additional components and "units." Additionally, the components and '~parts' may be implemented to activate one or more CPUs within a device or secure multimedia card. Furthermore, according to various embodiments of the present disclosure, the '~parts' may include one or more processors.

[0071] The operating principles of the present disclosure are described in detail below with reference to the attached drawings. In the following description of the present disclosure, detailed descriptions of related known functions or configurations will be omitted if they are deemed to unnecessarily obscure the gist of the present disclosure. Furthermore, the terms described below are defined based on the functions of the present disclosure and may vary depending on the intent or custom of the user or operator. Therefore, their definitions should be based on the overall content of this specification.

[0072] In this disclosure, "3D shape information" may refer to data representing the shape or structure of an object in three-dimensional space. The three-dimensional shape information may be generated through a CAD model, 3D scan data, or 3D modeling software. The three-dimensional shape information may be used as input data for extracting latent vectors. The three-dimensional shape information may be converted into a three-dimensional grid in the form of a coded distance function.

[0073] A "latent vector" can refer to a vector that represents the main features of high-dimensional data by compressing it into a low-dimensional form. A latent vector can be an intermediate representation generated after processing input data in a neural network model. For example, it can be expressed as a numeric array that encapsulates the geometric and topological features of a 3D shape. A latent vector can be extracted as data representing the characteristics of 3D shape information. Latent vectors can be used as inputs to both low-fidelity and high-fidelity neural networks.

[0074] A "low-fidelity neural network" can refer to a neural network trained with relatively low-resolution or simplified data. Low-fidelity neural networks can be utilized as models that require fast computation and few resources. Low-fidelity neural networks can be trained based on low-fidelity data sets. Low-fidelity neural networks can be used to predict field values ​​at specific coordinates in a 3D shape model. In this case, the predicted field values ​​may be referred to as "first field values," "low-fidelity predicted values," or "base predicted values."

[0075] A "high-fidelity neural network" can refer to a neural network trained on high-resolution or complex data. High-fidelity neural networks can be used to model complex relationships and make precise predictions. High-fidelity neural networks can be trained on high-fidelity data sets. High-fidelity neural networks can be used to predict field values ​​at specific coordinates in a 3D shape model. In this case, the high-fidelity neural network can produce a final prediction value by correcting the underlying prediction value predicted by the low-fidelity neural network. This final prediction value may be referred to as a "second field value," a "high-fidelity prediction value," or a "corrected prediction value."

[0076] A "correction term" can refer to a value that is added to the predicted value to correct the result. This correction term can be used to improve the model's prediction accuracy. This correction term can be computed using a high-fidelity neural network. The correction term can be output as a residual learned from the difference between the first and second field values.

[0077] A "coded distance function" can refer to a function that encodes and represents the distance between a point and a surface in three-dimensional space. For example, the inside of a surface can be expressed as a negative number, and the outside as a positive number, to distinguish the inside and outside of a shape. A coded distance function can be used to mathematically represent three-dimensional shapes. A coded distance function can be used to transform three-dimensional shape information into a three-dimensional grid. A coded distance function can be utilized as input data for extracting latent vectors.

[0078] A "3D convolutional neural network" can refer to a convolutional neural network designed to process 3D data. A 3D convolutional neural network can be used to analyze data such as images, volume data, or 3D models. A 3D convolutional neural network can be used to extract latent vectors from a 3D grid. In particular, it can effectively extract shape features by learning the spatial hierarchical structure of 3D data. A 3D convolutional neural network can be utilized to learn the features of 3D shape information.

[0079] A "low-fidelity dataset" can refer to a dataset with low resolution or a simplified form. Low-fidelity datasets can be used for fast processing and low resource consumption, enabling cost-effective model training. Low-fidelity datasets can be obtained from models discretized into grids with a primary resolution. Low-fidelity datasets can be used to train low-fidelity neural networks.

[0080] A "high-fidelity dataset" can refer to a dataset with high resolution or complex geometry. High-fidelity datasets can be used for precise analysis and prediction. While they provide more accurate data that closely resembles physical phenomena, they can be expensive to acquire. High-fidelity datasets can be obtained from models discretized into a grid with a second resolution higher than the first. High-fidelity datasets can be used to train high-fidelity neural networks.

[0081] A "unified loss function" can refer to a function that combines multiple loss factors into a single loss value. This unified loss function can be used to optimize a model's learning performance. It can include prediction errors for both low-fidelity data and high-fidelity data. This allows the model to learn the characteristics of both types of data in a balanced way. This unified loss function can be used to train both low-fidelity and high-fidelity neural networks.

[0082] A "mapping neural network" can refer to a neural network that learns to transform one data space into another. Mapping neural networks can be used to compensate for structural differences in data. For example, they can transform data sets with different numbers or types of input variables so that they can be processed by a single model. Mapping neural networks can transform high-fidelity data input variables into low-fidelity data input variables. The variables transformed by the mapping neural network can then be provided as input to the low-fidelity neural network.

[0083] "Residual" can refer to the difference between the predicted value and the actual value. Residuals can be used to evaluate or calibrate a model's predictive performance. By having the model learn residuals instead of predicting the values ​​themselves, faster and more stable learning can be achieved. The residual can be defined as the difference between the first and second field values. The residual can be learned by a high-fidelity neural network and output as a correction term.

[0084]

[0085] In this disclosure, "behavior" refers to a concept encompassing the movement, response, and state changes exhibited by an object or system in response to given physical laws or conditions (e.g., external load, self-rotation, fluid interaction, thermal conditions, etc.). This includes changes in physical quantities such as displacement, stress, and temperature observable at a specific point in time or over a continuous time interval, and may represent behavioral aspects of an object that can be predicted by an electronic device based on a condition vector. Various forms such as "dynamic behavior," "rotary behavior," and "fluid-structure interaction behavior" all fall under the category of behavior.

[0086] In this disclosure, a "scene" refers to a visual or physical configuration representing a specific environment or situation. It may be composed of various elements, such as objects (single or multiple), backgrounds, and lighting. A scene may include input (e.g., 3D data) related to the shape, motion, material properties, and physical quantities of objects, and may also express the relative arrangement or interaction of objects. An electronic device may receive input regarding the scene and extract a condition vector.

[0087] In this disclosure, a "condition vector" is a vector that numerically represents a specific condition or situation, and can be utilized in various forms in data analysis or machine learning. A condition vector is a collection of information extracted based on scene input (e.g., object shape, motion, physical quantities, etc.) and may also be referred to as a "scenario vector." Furthermore, a condition vector can be used as basic data for predicting object behavior or field values ​​in pre-trained artificial intelligence models.

[0088] In this disclosure, an "extractor" refers to a device that analyzes input data to extract specific characteristics. For example, the extractor may extract a shape vector by mapping 3D shape information of an object to a latent space, or may derive vector-based information by scaling kinematic parameters. The extractor may include a neural network model, a deep learning model, or the like, and may extract vectors using frequency response, but is not limited thereto.

[0089] In the present disclosure, "kinematic parameters" refer to parameters (e.g., velocity, acceleration, position, etc.) that indicate the motion state of an object. An electronic device may convert the parameters into vectors through a preprocessing process such as scaling, so that they can be input into a later model or utilized when generating a condition vector. For example, kinematic parameters may include the position, velocity, acceleration, orientation, attitude, angular velocity, or angular acceleration of an object.

[0090] In this disclosure, "Dynamic Field Prediction" refers to a process of estimating physical distributions (e.g., displacement, stress, temperature, flow velocity, etc.) that vary over time. The dynamic field prediction method can be implemented by estimating the behavior of an object or system at a specific point in time or over a continuous period using a pre-trained artificial intelligence model, etc., based on a condition vector, a latent vector, simulation data, or measurement data. Dynamic field prediction is a process for predicting nonlinear or time-varying characteristics of complex physical phenomena, and can increase accuracy by processing temporal data using time stamps.

[0091] Referring to FIG. 17, which illustrates various examples of 3D data, an electronic device can acquire 3D data for 3D shape information (S) using various 3D shape embedding methods. The 3D data acquired through the 3D shape embedding method is data used to express an object in a 3D space, and examples thereof include, but are not limited to, 3D SDF, 3D point cloud data, 3D mesh, 3D voxel, etc. For example, a 3D point cloud includes coordinate information of numerous points constituting the surface of an object measured by a 3D scanner. As an example, 3D mesh data includes information expressing the shape of a 3D object as points, lines, and surfaces.

[0092] FIG. 1 is a diagram illustrating a configuration of an electronic device according to various embodiments.

[0093] Referring to FIG. 1, an electronic device (100) may represent a hardware configuration for performing various embodiments of the present disclosure. The electronic device (100) may be a device for predicting a three-dimensional physical field. The electronic device (100) may include a processor (110), a memory (120), a storage device (130), a communication circuit (140), and an input / output interface (150).

[0094] An electronic device (100) can perform a function of predicting a three-dimensional physical field based on a multi-fidelity neural network. The electronic device (100) can effectively fuse a small amount of high-quality data with a large amount of low-quality data. Through this, the electronic device (100) can implement a system that predicts a three-dimensional physical field quickly and accurately. For example, the electronic device (100) may be a server computer, a desktop computer, a workstation, a laptop computer, a cloud computing node, a mobile device, a tablet computer, or an embedded system. However, the present disclosure is not limited thereto.

[0095] The processor (110) can control the overall operation of the electronic device (100). The processor (110) can execute an operating system or application program stored in the memory (120). The processor (110) can perform various logical operations to control other components of the electronic device (100).

[0096] In one embodiment of the present disclosure, the processor (110) may be configured to execute instructions stored in the memory (120). The processor (110) may perform an operation of receiving three-dimensional shape information and extracting a latent vector representing a characteristic of the shape information. In addition, the processor (110) may perform an operation of predicting a first field value through a pre-trained low-fidelity neural network using the latent vector and the first coordinate as inputs, and calculating a correction term through a high-fidelity neural network based on the predicted first field value to predict a final second field value.

[0097] In one embodiment of the present disclosure, the processor (110) may be configured with one or more processing cores. For example, the processor (110) may be a central processing unit (CPU), a graphics processing unit (GPU), a neural processing unit (NPU), a tensor processing unit (TPU), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or a digital signal processor (DSP). However, the present disclosure is not limited thereto.

[0098] The memory (120) may be electronically connected to the processor (110). The memory (120) may temporarily store instructions to be performed by the processor (110) or data during calculation. The memory (120) may be a volatile memory. For example, the memory (120) may be a random access memory (RAM), a dynamic RAM (DRAM), a static RAM (SRAM), a synchronous dynamic RAM (SDRAM), a double data rate SDRAM (DDR SDRAM), a graphics DDR SDRAM (GDDR SDRAM), or a low-power DDR SDRAM (LPDDR SDRAM). However, the present disclosure is not limited thereto.

[0099] The storage device (130) can permanently store a large amount of data. The storage device (130) may be a non-volatile storage medium. The storage device (130) may store an operating system, an application program for performing 3D field prediction, low-fidelity and high-fidelity neural network models, and low-fidelity and high-fidelity data sets used for training. For example, the storage device (130) may be a hard disk drive (HDD), a solid state drive (SSD), flash memory, a non-volatile memory express (NVMe)-based storage device, a secure digital (SD) card, a universal flash storage (UFS), or an optical disk drive. However, the present disclosure is not limited thereto.

[0100] The communication circuit (140) can support the electronic device (100) to transmit and receive data with other external devices or networks. The communication circuit (140) can support both wired and wireless communication methods. Through the communication circuit (140), 3D shape information can be received from an external server, or predicted 3D field results can be transmitted to other devices. For example, the communication circuit (140) can be an Ethernet module, a Wi-Fi module, a Bluetooth module, a Near Field Communication (NFC) module, a 5th generation (5G) communication module, a Long-Term Evolution (LTE) module, or a Universal Serial Bus (USB) port. However, the present disclosure is not limited thereto.

[0101] The input / output interface (150) can provide a path for interaction between a user and the electronic device (100). The input / output interface (150) can receive signals from an input device or transmit signals to an output device. The user can input parameters for field prediction or visually confirm the predicted 3D field results through the input / output interface (150). For example, the input / output interface (150) can be connected to a display panel, a touchscreen, a keyboard, a mouse, a speaker, a microphone, or a camera. However, the present disclosure is not limited thereto.

[0102]

[0103] FIG. 2 is a diagram for explaining a field prediction method using a field prediction system according to various embodiments.

[0104] Referring to FIG. 2, the electronic device (100) can predict a three-dimensional physical field using a multi-fidelity neural network. The processor (110) can implement a functional configuration of a field prediction system (1000). The processor (110) can receive three-dimensional shape information (1) and perform an operation of extracting a latent vector representing a characteristic of the shape information. The processor (110) can perform an operation of predicting a first field value in the three-dimensional coordinate (3) through a pre-trained low-fidelity neural network (1210) by using the latent vector and the three-dimensional coordinate (3) as inputs. The processor (110) can input the latent vector, the three-dimensional coordinate (3), and the predicted first field value. The processor (110) can perform an operation of calculating a correction term for the first field value through at least one high-fidelity neural network (1220). The processor (110) can perform an operation of predicting a second field value based on a first field value and a correction term. The field prediction system (1000) can receive three-dimensional shape information (1) and three-dimensional coordinates (3) as input. The field prediction system (1000) can output a field prediction result at the node level.

[0105] The field prediction system (1000) may include an extractor (1100) and a prediction network (1200). The extractor (1100) may process input 3D shape information (1) to generate a latent vector representing a feature. The prediction network (1200) may utilize the latent vector transmitted from the extractor (1100) and externally input 3D coordinates (3). The prediction network (1200) may predict a physical field value at the corresponding coordinates.

[0106] The extractor (1100) may perform a function of extracting a latent vector that compresses the geometric features of a shape from three-dimensional shape information (1). In one embodiment of the present disclosure, the extractor (1100) may include a preprocessing step of converting the three-dimensional shape information (1) into a three-dimensional grid (2) in the form of a signed distance function (SDF). Thereafter, the extractor (1100) may receive the three-dimensional grid (2) as input and output a latent vector.

[0107] In one embodiment of the present disclosure, the extractor (1100) may be implemented as a 3D convolutional neural network. The 3D convolutional neural network can effectively learn spatial features contained in a 3D grid (2). The 3D convolutional neural network can express spatial features as low-dimensional latent vectors. The generated latent vectors can encapsulate the unique characteristics of the 3D shape information (1). Therefore, the generated latent vectors can support the prediction network (1200) to perform consistent predictions for various shapes.

[0108] The prediction network (1200) may include a low-fidelity neural network (1210) and a high-fidelity neural network (1220). The prediction network (1200) may receive a latent vector generated by the extractor (1100) and specific three-dimensional coordinates (3) for which prediction is desired as input. The prediction network (1200) may perform fast and accurate field value prediction through a multi-fidelity approach.

[0109] A low-fidelity neural network (1210) can predict a first field value using a latent vector and a three-dimensional coordinate (3) as input. The first field value may be an initial prediction for a physical field. The first field value may be quickly generated based on low-fidelity simulation data with low computational cost. The low-fidelity neural network (1210) may be pre-trained based on a low-fidelity data set. For example, the low-fidelity data set may be obtained from a model discretized into a mesh having a first resolution.

[0110] In one embodiment of the present disclosure, the low-fidelity neural network (1210) may be implemented with various artificial neural network structures. For example, the low-fidelity neural network (1210) may be a multi-layer perceptron (MLP), a recurrent neural network (RNN), a transformer network, a convolutional neural network (CNN), an autoencoder, a generative adversarial network (GAN), or a radial basis function network (RBFN). However, the present disclosure is not limited thereto. A training data set used for training the low-fidelity neural network (1210) may include a plurality of coordinate values ​​and a label value corresponding to each coordinate value. Additionally, the low-fidelity neural network (1210) may include a coordinate-based implicit neural regressor, which may perform the function of predicting field values ​​based on input coordinates.

[0111] The high-fidelity neural network (1220) can perform the role of predicting the final field value by correcting the prediction result of the low-fidelity neural network (1210). The high-fidelity neural network (1220) can receive as input all of the latent vector, the three-dimensional coordinates (3), and the first field value predicted by the low-fidelity neural network (1210). Based on these inputs, the high-fidelity neural network (1220) can calculate a correction term for the first field value.

[0112] In one embodiment of the present disclosure, the high-fidelity neural network (1220) may be configured to learn the residual, which is the difference between the first field value and the second field value. The correction term output by the high-fidelity neural network (1220) may correspond to the learned residual. The final predicted second field value may be calculated by adding the correction term to the first field value. This may have the effect of combining the speed of low-fidelity prediction with the accuracy of high-fidelity prediction.

[0113] A high-fidelity neural network (1220) can be trained based on a high-fidelity data set. For example, the high-fidelity data set can be obtained from a model discretized into a second grid with a higher resolution than the grid of the low-fidelity data set. The high-fidelity neural network (1220) can be effectively trained even with a relatively small amount of high-quality data. This may be because the high-fidelity neural network (1220) does not directly learn the entire field values, but only learns the differences from the low-fidelity predictions.

[0114] In one embodiment of the present disclosure, the processor (110) may further perform an operation of generating a training data set for training a low-fidelity neural network (1210) or a high-fidelity neural network (1220). The training data set may include a plurality of coordinate values ​​and a plurality of label values ​​corresponding to each of the coordinate values. Each of the label values ​​may be determined based on the corresponding coordinate value. The processor (110) may identify a preset number (N) of nearest nodes among a plurality of nodes included in the low-fidelity data set or the high-fidelity data set. In addition, the processor (110) may determine a label value by calculating an average of the field values ​​of the identified nearest nodes.

[0115] In one embodiment of the present disclosure, a low-fidelity neural network (1210) and a high-fidelity neural network (1220) can be trained together based on a unified loss function. The unified loss function can be defined to include both prediction errors for low-fidelity data and prediction errors for high-fidelity data. This joint learning approach can optimize the interaction between the two neural networks. This can maximize the performance of the overall field prediction system (1000).

[0116]

[0117] The electronic device (100) can sample three-dimensional coordinates (3).

[0118] Referring to FIG. 18, the electronic device (100) can input 3D CAD data or 3D data expressing a 3D shape and output node coordinates of the 3D shape. Here, the node coordinates can all be expressed as x, y, and z by an orthogonal coordinate system expressing a 3D space.

[0119] The electronic device (100) can output node coordinates for nodes of a 3D shape using one or a combination of two of volume mesh point sampling, surface mesh point sampling, and 3D space point sampling techniques.

[0120] For example, the electronic device (100) can obtain the coordinates (Pv) for the volume node of the 3D shape obtained by applying a volume mesh point sampling technique to 3D data as node coordinates (P).

[0121] As another example, the electronic device (100) may obtain, as the node coordinates (P), the coordinates (Pv) of the volume nodes of the 3D shape obtained by the closest volume mesh point sampling from the coordinates (Ps) of the surface nodes of the 3D shape obtained by the surface mesh point sampling technique. Alternatively, the electronic device (100) may obtain, as the node coordinates (P), the values ​​determined through linear interpolation of the coordinates (Pv) of the volume nodes of the 3D shape obtained by the closest n (e.g., 5) volume mesh point sampling from the coordinates (Ps) of the surface nodes of the 3D shape obtained by the surface mesh point sampling.

[0122] As another example, the electronic device (100) can obtain, as the node coordinates (P), the coordinates (Pv) of the volume node of the 3D shape obtained by volume mesh point sampling that are closest to the coordinates (Pp) for the 3D space in which the 3D shape obtained by 3D space point sampling exists. Alternatively, the electronic device (100) can obtain, as the node coordinates (P), the values ​​of the coordinates (Pv) of the volume nodes of the 3D shape obtained by volume mesh point sampling that are closest to the coordinates (Pp) for the 3D space in which the 3D shape obtained by 3D space point sampling exists, through linear interpolation or nearest interpolation.

[0123] As described above, the electronic device (100) can output node coordinates (P) for a plurality of 3D shapes using one or a combination of two of the volume mesh point sampling, surface mesh point sampling, and 3D space point sampling techniques.

[0124] Volume Mesh Point Sampling is a method of extracting nodes of a 3D shape by uniformly sampling the volume inside the 3D shape.

[0125] As an example of volume mesh point sampling, a volume node of a 3D shape may be a grid node that divides the 3D space into grid cells of a certain size using a grid division technique and may be the center points of each cell. As another example, a volume node of a 3D shape may be a grid node that divides the 3D space into elements of a tetrahedron or a hexahedron and may be the center points of each element. As another example, a volume node of a 3D shape may be a point sampled by randomly selecting points within the 3D space using a random sampling technique. As another example, a volume node of a 3D shape may be a point sampled by adjusting the sampling density according to the complexity of the 3D shape using an adaptive sampling technique, where important parts are sampled densely and less important parts are sampled loosely.

[0126] For example, the electronic device (100) can receive 3D CAD data representing a 3D shape or a Visualization Toolkit (VTK) file as 3D data, and output the node coordinates (P) of the 3D shape. Here, the VTK file is a file format for scientific data visualization and analysis, and is widely used in 3D computer graphics, image processing, and visualization work, and supports various data structures for storing and visualizing meshes, point clouds, and other scientific data.

[0127] Since the VTK file of the 3D shape has a volume mesh for the 3D shape and coordinate information for the nodes of each volume mesh, the coordinate electronic device (100) can output the node coordinates (Pv) of the 3D shape through coordinate random sampling for the nodes of the VTK file. The coordinate sampling unit (230) can obtain the node coordinates (Pv) of the 3D shape as the node coordinates (121) for the nodes of the 3D shape.

[0128] Surf Mesh Point Sampling is a method of extracting nodes of a 3D shape by uniformly sampling the surface of the 3D shape.

[0129] As an example of surface mesh point sampling, the surface nodes of a 3D shape may be the center points of each triangular or rectangular patch divided by a grid division technique on the 3D shape surface. As another example, the surface nodes of a 3D shape may be points formed to evenly cover the 3D shape surface by a Poisson disk sampling technique. As another example, the surface nodes of a 3D shape may be points randomly generated and selected on the surface.

[0130] For example, the electronic device (100) can receive 3D CAD data representing a 3D shape or an STL file as 3D data, and output coordinates (Ps) for surface nodes of the 3D shape. Here, the STL file represents a 3D shape as a set of triangular faces, and is a file format mainly used in 3D printing and computer graphics.

[0131] Since the STL file of the 3D shape has a surface mesh for the 3D shape and coordinate information for each node of the surface mesh, the electronic device (100) can output the coordinates (Ps) of the surface nodes of the 3D shape through coordinate random sampling for the nodes of the STL file. The coordinate sampling unit (230) can obtain the coordinates (Pv) of the volume nodes of the 3D shape close to the coordinates (Ps) of the nodes of the surface mesh as the node coordinates (P) of the 3D shape.

[0132] 3D Space Point Sampling is a method for extracting the coordinates of nodes in a 3D shape by sampling points (spatial nodes) throughout 3D space and generating point cloud data. For example, the spatial nodes of a 3D shape may be a collection of points within 3D space obtained from a 3D scanner. Alternatively, the spatial nodes of a 3D shape may be randomly generated and selected points within 3D space.

[0133] For example, the electronic device (100) can input 3D CAD data or 3D data expressing a 3D shape, randomly sample coordinates of points generated in the 3D space where the 3D shape is located, and output the 3D space coordinates (Pp) of the 3D shape. The coordinate sampling unit (230) can obtain the coordinates (Pv) of a volume node of the 3D shape that is close to the 3D space coordinates (Pp) of the 3D shape randomly selected in the 3D space as the node coordinates (P) for the 3D shape.

[0134] FIG. 3 is a flowchart illustrating a field prediction method using a field prediction system according to various embodiments.

[0135] Referring to FIG. 3 together with FIG. 2, a field prediction method using a field prediction system (1000) can be performed by a processor (110) of an electronic device (100).

[0136] In step S310, the processor (110) may receive three-dimensional shape information (1) and extract a latent vector representing the characteristics of the shape information (1). The processor (110) may generate a latent vector from the three-dimensional shape information (1) through an extractor (1100). The latent vector may imply geometric or topological characteristics of the shape. The latent vector may be a result of mapping high-dimensional shape information into a low-dimensional vector space.

[0137] In step S320, the processor (110) can predict a first field value at a first coordinate using a pre-trained low-fidelity neural network (1210) with a latent vector and a first coordinate as input. The low-fidelity neural network (1210) can be included in the prediction network (1200) and can be pre-trained based on a low-fidelity data set. The low-fidelity neural network (1210) can quickly predict a physical field value at a specific three-dimensional coordinate (3).

[0138] In step S330, the processor (110) may calculate a correction term for the first field value through at least one high-fidelity neural network (1220) by inputting a latent vector, a first coordinate, and a predicted first field value. The high-fidelity neural network (1220) may correct an error in the first field value, which is a prediction result of the low-fidelity neural network (1210). The high-fidelity neural network (1220) may learn the difference between the low-fidelity prediction result and actual high-fidelity data to produce the correction term.

[0139] In step S340, the processor (110) can predict a second field value based on the first field value and the correction term. The processor (110) can apply the correction term calculated by the high-fidelity neural network (1220) to the first field value predicted by the low-fidelity neural network (1210) to derive a final second field value. By deriving the second field value, the processor (110) can improve the accuracy of the prediction.

[0140] In one embodiment of the present disclosure, the operation of predicting the second field value may include obtaining the second field value by adding a correction term to the first field value. For example, the correction term may be a residual between the first field value and the actual field value. The processor (110) can obtain a more accurate second field value by adding the correction term to the first field value.

[0141] In one embodiment of the present disclosure, the operation of extracting a latent vector may include an operation of converting three-dimensional shape information (1) into a three-dimensional grid (2) in the form of a signed distance function (SDF). The operation of extracting a latent vector may include an operation of extracting a latent vector from the three-dimensional grid (2) using a three-dimensional convolutional neural network.

[0142] In one embodiment of the present disclosure, a low-fidelity neural network (1210) may be trained based on a low-fidelity data set. At least one high-fidelity neural network (1220) may be trained based on a high-fidelity data set.

[0143] In one embodiment of the present disclosure, the processor (110) may be further configured to perform an operation of generating a training data set for training a low-fidelity neural network (1210) or at least one high-fidelity neural network (1220). The training data set may include a plurality of coordinate values ​​and a plurality of label values ​​corresponding to each of the coordinate values. Each of the label values ​​may identify a preset number of nearest nodes among a plurality of nodes included in the low-fidelity data set or the high-fidelity data set based on the corresponding coordinate value. Each of the label values ​​may be determined by calculating an average of field values ​​of the identified nearest nodes.

[0144] In one embodiment of the present disclosure, at least one high-fidelity neural network (1220) may include a first high-fidelity neural network that models a linear relationship for an input and a second high-fidelity neural network that models a nonlinear relationship. The operation of calculating a correction term may include the operation of calculating a correction term based on a first output value output from the first high-fidelity neural network and a second output value output from the second high-fidelity neural network.

[0145] In one embodiment of the present disclosure, a low-fidelity neural network (1210) and at least one high-fidelity neural network (1220) may be trained based on an integrated loss function that includes both prediction errors for low-fidelity data and prediction errors for high-fidelity data. The integrated loss function may be defined as a weighted sum of loss terms for each fidelity data.

[0146] In one embodiment of the present disclosure, the electronic device (100) may further include a mapping neural network. The mapping neural network may convert input variables of high-fidelity data used for training the high-fidelity neural network (1220) into input variables of low-fidelity data used for training the low-fidelity neural network (1210). The variables converted by the mapping neural network may be provided as inputs to the low-fidelity neural network (1210).

[0147] In one embodiment of the present disclosure, at least one high-fidelity neural network (1220) may be configured to learn a residual, which is a difference between a first field value and a second field value. The learned residual may be output as a correction term.

[0148] FIG. 4 is a flowchart illustrating a method for extracting a latent vector according to various embodiments.

[0149] Referring to FIG. 4 together with FIG. 3, step S310 of FIG. 3 may include the steps illustrated in FIG. 4. The steps of FIG. 4 may be performed by the processor (110).

[0150] In step S311, the processor (110) may convert the 3D shape information into a 3D grid in the form of a Signed Distance Function (SDF). For example, the processor (110) may receive 3D computer-aided design (CAD) shape data. For example, the 3D computer-aided design (CAD) shape data may be a STEP (Standard for the Exchange of Product model data) file, an IGES (Initial Graphics Exchange Specification) file, an STL (Stereolithography) file, an OBJ (Object) file, a 3DM file, a DWG (Drawing) file, or a CATPart file. However, the present disclosure is not limited thereto. A neural network may have limitations in directly processing 3D computer-aided design (CAD) data. The processor (110) may convert the 3D shape information into a 3D grid in the form of a Signed Distance Function (SDF). A signed distance function (SDF) can be a value representing the shortest distance from each point in the grid to the surface of the shape.

[0151] In one embodiment of the present disclosure, step S311 may include a step of preprocessing three-dimensional shape information. The processor (110) may perform a preprocessing operation to process data into a form suitable for input to a neural network. The preprocessing operation may include normalization, scaling, or noise removal of the three-dimensional shape data.

[0152] In step S312, the processor (110) can extract a latent vector (z) from a 3D grid using a 3D convolutional neural network. An encoder configured with a 3D convolutional neural network can process a coded distance function (SDF) in the form of a 3D grid as input. The encoder can compress information including geometric features of a shape from the input 3D grid. The encoder can generate a latent vector (z) using the compressed information. The generated latent vector (z) can be used as unique identification information or a shape code of the corresponding 3D shape.

[0153] In one embodiment of the present disclosure, the operation of extracting the latent vector (z) may include the operation of converting the 3D shape information into a 3D grid in the form of a signed distance function (SDF). The operation of extracting the latent vector (z) may further include the operation of extracting the latent vector (z) from the 3D grid using a 3D convolutional neural network. Complex 3D shape information can be converted into a low-dimensional vector representation that is easy for the neural network to process through the latent vector extraction operation.

[0154] Figure 5 is a diagram illustrating an example of low-fidelity data and high-fidelity data.

[0155] Referring to FIG. 5, an example of a low-fidelity data set (510) and a high-fidelity data set (520) may be illustrated. The low-fidelity data set (510) and the high-fidelity data set (520) may constitute a data set used to train a multi-fidelity neural network. The electronic device (100) may train a low-fidelity neural network (1210) based on the low-fidelity data set (510). Additionally, the electronic device (100) may train a high-fidelity neural network (1220) based on the high-fidelity data set (520).

[0156] A low-fidelity data set (510) may be a simulation result that can be obtained with relatively low computational cost. The low-fidelity data set (510) may represent the overall trend of a physical phenomenon. However, the low-fidelity data set (510) may have relatively low precision in localized areas. As illustrated in Fig. 5 (a), the low-fidelity data set (510) may be obtained from a model discretized into a grid with a first resolution, i.e., a coarse mesh. Simulations using a coarse mesh may have the advantage of fast computation speed. Therefore, it may be easy to secure a large amount of low-fidelity data sets (510).

[0157] The low-fidelity data set (510) can be used for pre-training of the low-fidelity neural network (1210). The low-fidelity neural network (1210) can learn the basic relationship between 3D shape information and physical field values ​​through a large amount of low-fidelity data sets (510). This pre-training process can increase the learning efficiency of the overall prediction model. In addition, the low-fidelity data set (510) can be used for rapid performance prediction in the early design stage where it is difficult to obtain a high-fidelity data set (520). For example, the low-fidelity data set (510) can be a coarse mesh-based analysis result, a simplified model analysis result using beam elements, a simplified model analysis result using shell elements, an inviscid flow analysis result, 2D cross-sectional analysis data, simulation data based on specific assumptions, a result applying simplified physical laws, or a result using a low-order numerical analysis technique. However, the present disclosure is not limited thereto.

[0158] The high-fidelity data set (520) may be a simulation result or experimental data with high accuracy that is very close to actual physical phenomena. The high-fidelity data set (520) may require a lot of computational cost and time. Therefore, only a small amount of the high-fidelity data set (520) can be secured. As illustrated in Fig. 5 (b), the high-fidelity data set (520) may be obtained from a model discretized into a fine mesh, i.e., a second resolution higher than the first resolution. Simulations using a fine mesh can precisely simulate stress concentration areas or complex flow phenomena.

[0159] The high-fidelity data set (520) can be used to train the high-fidelity neural network (1220). The high-fidelity neural network (1220) can learn the difference, i.e., the residual, between the prediction result of the low-fidelity neural network (1210) and the actual high-fidelity data set (520). By correcting the low-fidelity prediction result using a small amount of the high-fidelity data set (520), the overall system can achieve both computational efficiency and prediction accuracy. For example, the high-fidelity data set (520) can be a fine mesh-based analysis result, a solid model analysis result, a viscous flow analysis result, 3D full shape analysis data, experimental data measured from an actual sensor, a result obtained by solving a complete partial differential equation (PDE), or a result using a high-dimensional numerical analysis technique. However, the present disclosure is not limited thereto.

[0160] In one embodiment of the present disclosure, the electronic device (100) may perform an operation of generating a learning data set for training a low-fidelity neural network (1210) or a high-fidelity neural network (1220). The learning data set may include a plurality of coordinate values ​​and a plurality of label values ​​corresponding to each of the coordinate values. Each of the label values ​​may be determined by identifying a preset number (N) of nearest nodes among a plurality of nodes included in the low-fidelity data set (510) or the high-fidelity data set (520), based on the corresponding coordinate value. The electronic device (100) may determine the label value by calculating an average of the field values ​​of the identified nearest nodes. Through this process, the electronic device (100) may convert simulation data in an irregular grid shape into a form suitable for neural network training.

[0161] Figure 6 is a diagram illustrating an example of a low-fidelity prediction result and a high-fidelity prediction result.

[0162] Referring to FIG. 6, following FIG. 5, examples of a low-fidelity prediction result (610) and a high-fidelity prediction result (620) may be illustrated. The electronic device (100) may generate the low-fidelity prediction result (610) through a low-fidelity neural network (1210). The electronic device (100) may correct the low-fidelity prediction result (610) using a high-fidelity neural network (1220). The electronic device (100) may obtain the high-fidelity prediction result (620) through the corrected low-fidelity prediction result (610).

[0163] The low-fidelity prediction result (610) may be an initial 3D physical field predicted through a low-fidelity neural network (1210). The processor (110) of the electronic device (100) may provide the latent vector obtained from the extractor (1100) and the 3D coordinates of the query target as inputs to the low-fidelity neural network (1210). The low-fidelity neural network (1210) may be pre-trained based on the low-fidelity data set (510). The low-fidelity neural network (1210) may output the first field value with low computational cost. Referring to (a) of FIG. 6, the low-fidelity prediction result (610) may be expressed in the form of a point cloud including a plurality of nodes. The form of the low-fidelity prediction result (610) may reflect the characteristics of the low-fidelity data set (510).

[0164] The low-fidelity prediction result (610) may not be the final prediction result. The low-fidelity prediction result (610) may function as an intermediate result for high-fidelity prediction. The processor (110) may utilize the low-fidelity prediction result (610) as input to the high-fidelity neural network (1220). The processor (110) may perform a correction process to improve the prediction accuracy. The multi-fidelity approach may contribute to increasing the efficiency of the overall prediction process.

[0165] The high-fidelity prediction result (620) may be a 3D physical field obtained by correcting the low-fidelity prediction result (610). The high-fidelity prediction result (620) may reflect the accuracy and detailed physical characteristics of the high-fidelity data set (520). Referring to Fig. 6 (b), the high-fidelity prediction result (620) may be expressed in the form of a smooth and continuous surface. The form of the high-fidelity prediction result (620) may more precisely describe physical phenomena.

[0166] The processor (110) may perform the steps of the field prediction method of FIG. 3 to obtain a high-fidelity prediction result (620). The processor (110) may input a latent vector, a three-dimensional coordinate, and a first field value predicted from a low-fidelity neural network (1210) into the high-fidelity neural network (1220) (S330). The first field value may be the low-fidelity prediction result (610). The high-fidelity neural network (1220) may be configured to learn a residual, which is a difference between the first field value and an actual high-fidelity field value. The high-fidelity neural network (1220) may output the residual as a correction term. The processor (110) may add the correction term to the first field value. The processor (110) may predict a value obtained by adding the correction term to the first field value as a second field value (S340). The second field value may be a high-fidelity prediction result (620).

[0167] In one embodiment of the present disclosure, the electronic device (100) may perform an operation of generating a learning data set for training a low-fidelity neural network (1210) or a high-fidelity neural network (1220). The learning data set may include a plurality of coordinate values ​​and a plurality of label values ​​corresponding to each of the coordinate values. Each of the label values ​​may be determined through a process of identifying a preset number (N) of nearest nodes among a plurality of nodes included in the low-fidelity data set (510) or the high-fidelity data set (520), based on the corresponding coordinate value. For example, the result of identifying nearest nodes for a specific coordinate value in the low-fidelity data set (510) and the result of identifying nearest nodes for the same coordinate value in the high-fidelity data set (520) may be determined as nodes at different locations due to a difference in resolution between the two data sets. These differences imply that prediction results from low-fidelity datasets may not exactly match those from high-fidelity datasets, requiring more sophisticated interpolation methods when generating training datasets.

[0168] Each label value can be determined by calculating the average of the field values ​​of the identified closest nodes. For example, if the field values ​​of the closest nodes identified for a specific coordinate value in the low-fidelity data set (510) are {A1, A2, A3}, the label value can be determined by calculating the average of the corresponding field values. On the other hand, in the high-fidelity data set (520), the field values ​​of the closest nodes identified for the same coordinate value can appear as {B1, B2, B3, B4}, and in this case, the label value can also be determined by calculating the average of the corresponding field values. This method can contribute to correcting positional mismatch that may occur due to the difference in resolution between the low-fidelity data set and the high-fidelity data set.

[0169] Interpolation can generate structured data from simulation data with an irregular grid structure. For example, if the spacing between nodes in a low-fidelity data set is irregular, interpolation can transform it into structured data with a consistent spacing. Structured data can be used for neural network training, and in particular, it can improve model robustness by reducing the influence of noise. This technical method plays a crucial role in minimizing training data mismatch that can arise from positional differences between low-fidelity and high-fidelity prediction results, thereby enhancing the accuracy of neural network training.

[0170] FIG. 7 is a diagram illustrating an example of a multi-fidelity neural network-based field prediction system according to various embodiments.

[0171] Referring to FIG. 7, together with FIGS. 2 to 6, an example of an electronic device (1000) performing field prediction based on a multi-fidelity neural network may be illustrated. The electronic device (1000) may be implemented as a field prediction system. The electronic device (1000) may have a multi-fidelity neural network structure. The electronic device (1000) may include a low-fidelity neural network (710), a high-fidelity neural network (720), and a calculator (730). The electronic device (1000) may receive input data (701) and output a final prediction value (703).

[0172] The multi-fidelity neural network according to the present disclosure can learn residuals based on low-fidelity prediction values ​​and correction terms based on [Formula 1] below.

[0173] [Formula 1]

[0174]

[0175] The input data (701) may be information provided to a neural network for field prediction. The input data (701) may include a latent vector extracted from three-dimensional shape information by at least one processor. In addition, the input data (701) may include first coordinates, which are coordinate information of a specific point for which a field value is to be predicted. For example, the input data (701) may be composed of a first input (x1) representing a latent vector and a second input (x2) representing the first coordinate. However, the present disclosure is not limited thereto.

[0176] A low-fidelity neural network (710) can predict initial field values ​​based on input data (701). The low-fidelity neural network (710) can be pre-trained based on a low-fidelity data set. The low-fidelity neural network (710) can perform fast predictions with relatively low computational cost to generate low-fidelity prediction values ​​(702). The low-fidelity neural network (710) can receive input data (701). The low-fidelity neural network (710) can perform non-linear transformation while passing the data through multiple hidden layers. The nodes of each hidden layer can process the output of the previous layer through an activation function. Finally, the low-fidelity neural network (710) can output a low-fidelity prediction value (702), which is a predicted first field value.

[0177] The low-fidelity prediction value (702) may be a first field value output from the low-fidelity neural network (710). The low-fidelity prediction value (702) may be used as a base value for calculating the final prediction value (703). In addition, the low-fidelity prediction value (702) may be input into the high-fidelity neural network (720) and utilized for calculating a correction term.

[0178] The high-fidelity neural network (720) may perform a role of calculating a correction term for correcting the low-fidelity prediction value (702), which is the first field value. In one embodiment of the present disclosure, the high-fidelity neural network (720) may receive a latent vector, a first coordinate, and a predicted first field value as inputs. Here, the latent vector and the first coordinate may be included in the input data (701). The first field value may correspond to the low-fidelity prediction value (702). The high-fidelity neural network (720) may be trained based on a high-fidelity data set. The high-fidelity neural network (720) may model a complex nonlinear relationship between the first field value and the actual high-fidelity value.

[0179] In one embodiment of the present disclosure, the high-fidelity neural network (720) may be configured to learn the residual, which is the difference between the first field value and the second field value. The high-fidelity neural network (720) may output the learned residual as a correction term (ε). This residual learning method may enable more efficient learning than directly predicting the entire high-fidelity value.

[0180] In one embodiment of the present disclosure, a low-fidelity neural network (710) and a high-fidelity neural network (720) can be trained together based on a unified loss function. The unified loss function can include both prediction errors for low-fidelity data and prediction errors for high-fidelity data. This allows the two neural networks to operate complementarily, optimizing overall prediction performance.

[0181] The multi-fidelity neural network according to the present disclosure can be trained based on the integrated loss function of [Equation 2].

[0182] [Formula 2]

[0183]

[0184] The operator (730) may be configured to combine the low-fidelity prediction value (702) and the correction term output by the high-fidelity neural network (720). In one embodiment of the present disclosure, the operator (730) may include an operation of obtaining a second field value by adding the correction term to the first field value. The operator (730) may perform an addition operation to sum the two input values. The output of the operator (730) may be the final prediction value (703), which is the corrected field value.

[0185] The final predicted value (703) is the final output of the electronic device (1000) and may be a high-fidelity predicted field value. The final predicted value (703) may correspond to a second field value calculated by adding a correction term to the low-fidelity predicted value (702). The final predicted value (703) may be provided as a node-level output at a specific coordinate. The final predicted value (703) may be used to represent the distribution of a three-dimensional physical field.

[0186] FIG. 8 is a diagram illustrating an example of a field prediction system including a mapping neural network according to various embodiments.

[0187] Referring to FIG. 8, a field prediction system (1000) may illustrate an embodiment including a mapping neural network (810). The field prediction system (1000) according to the present disclosure may be implemented by being included in an electronic device. Details of the field prediction system (1000) that overlap with those described in FIG. 7 may be omitted. The field prediction system (1000) may receive input data (801) and output a final prediction value (803). The field prediction system (1000) may include a mapping neural network (810), a low-fidelity neural network (820), and a high-fidelity neural network (830).

[0188] The input data (801) may represent initial information provided to the field prediction system (1000) for field prediction. The input data (801) may include input variables of high-fidelity data used for training the high-fidelity neural network (830). For example, the input data (801) may include common variables (xc) commonly used with low-fidelity data and non-common variables (x_h^nc) that exist only in high-fidelity data.

[0189] The mapping neural network (810) can perform the function of transforming variables between different data domains. Specifically, the mapping neural network (810) can transform the input variables of high-fidelity data used for training the high-fidelity neural network (830) into the input variables of low-fidelity data used for training the low-fidelity neural network (820). This transformation may be necessary when the low-fidelity neural network (820) and the high-fidelity neural network (830) use different types of input variables.

[0190] The mapping neural network (810) can receive as input a non-common variable (x_h^nc) included in the input data (801). The mapping neural network (810) can process the input non-common variable (x_h^nc). The mapping neural network (810) can map the processed non-common variable (x_h^nc) into a variable (x_l^nc) in a form that the low-fidelity neural network (820) can understand. The mapping neural network (810) can be set to learn the relationship between the input space of high-fidelity data and the input space of low-fidelity data.

[0191] In one embodiment of the present disclosure, the mapping neural network (810) may be implemented with a multi-layer perceptron (MLP) structure. For example, the mapping neural network (810) may be a recurrent neural network (RNN), a long short-term memory (LSTM), a gated recurrent unit (GRU), a convolutional neural network (CNN), a transformer, an autoencoder, or a generative adversarial network (GAN). However, the present disclosure is not limited thereto. In addition, the field prediction system (1000) may be configured not to include the mapping neural network (810) when the types of input variables of low-fidelity data and high-fidelity data are the same. Therefore, whether or not to include a mapping neural network (810) can be selectively determined depending on the specific problem to be solved and the characteristics of the data set used.

[0192] Preprocessed data (802) may refer to data processed to be input to a low-fidelity neural network (820). The preprocessed data (802) may be generated by combining a common variable (xc) of input data (801) with a transformed non-common variable (x_l^nc) output from a mapping neural network (810). Through this combination, the low-fidelity neural network (820) can indirectly utilize information of high-fidelity data. Through this, the low-fidelity neural network (820) can improve the accuracy of prediction.

[0193] Regarding the low-fidelity neural network (820), any details that overlap with those described in FIG. 7 may be omitted. The low-fidelity neural network (820) may receive preprocessed data (802) as input. The low-fidelity neural network (820) may predict initial field values ​​(yl) based on the preprocessed data (802). The low-fidelity neural network (820) may be pre-trained based on a low-cost, low-fidelity data set. Through this, the low-fidelity neural network (820) may provide a fast prediction speed.

[0194] For the high-fidelity neural network (830), any overlapping information with that described in FIG. 7 may be omitted. The high-fidelity neural network (830) may receive initial input data (801) and the predicted value (yl) of the low-fidelity neural network (820) as input and calculate a correction term (ε). The high-fidelity neural network (830) may be designed to learn the residual between the low-fidelity predicted value and the actual high-fidelity value. Through this, the high-fidelity neural network (830) may increase the accuracy of the prediction.

[0195] The final prediction value (803) may represent the final output result of the field prediction system (1000). The final prediction value (803) may be calculated by adding the initial prediction value (yl) output by the low-fidelity neural network (820) and the correction term (ε) output by the high-fidelity neural network (830). Through this structure, the field prediction system (1000) can utilize the fast prediction speed of the low-fidelity neural network (820). In addition, the field prediction system (1000) can combine the high accuracy of the high-fidelity neural network (830). As a result, the field prediction system (1000) can perform efficient field prediction.

[0196] FIG. 9 is a diagram for explaining a specific implementation method of a multi-fidelity neural network according to various embodiments.

[0197] Referring to FIG. 9, the field prediction system (1000) may illustrate a specific implementation method of a multi-fidelity neural network. Details regarding the field prediction system (1000) that overlap with those described in FIGS. 2, 7, and 8 may be omitted. The field prediction system (1000) illustrated in FIG. 9 may separate a high-fidelity neural network into two individual neural networks that model linear and nonlinear relationships. The field prediction system (1000) may thus have a structure that enhances prediction accuracy.

[0198] The field prediction system (1000) may include a low-fidelity neural network (910), a first high-fidelity neural network (921), and a second high-fidelity neural network (922). The low-fidelity neural network (910) may perform a role of predicting an initial field value based on input data. The first high-fidelity neural network (921) and the second high-fidelity neural network (922) may correct the prediction result of the low-fidelity neural network (910). The first high-fidelity neural network (921) and the second high-fidelity neural network (922) may function to increase the accuracy of the final predicted value.

[0199] The low-fidelity neural network (910) can receive as input a latent vector and specific coordinates extracted from three-dimensional shape information. The low-fidelity neural network (910) can be pre-trained based on a low-fidelity data set. The low-fidelity neural network (910) can process the input information. The low-fidelity neural network (910) can output a low-fidelity predicted value (901), which is an initial field value at the corresponding coordinate.

[0200] The low-fidelity prediction value (901) may be a first field value generated by a low-fidelity neural network (910). The low-fidelity prediction value (901) may be a prediction based on simulation results with low computational cost. The low-fidelity prediction value (901) may be used as a base value for the final prediction. The low-fidelity prediction value (901) may be provided as input to the subsequent first high-fidelity neural network (921) and second high-fidelity neural network (922).

[0201] In one embodiment of the present disclosure, the high-fidelity neural network may include a first high-fidelity neural network (921) that models a linear relationship for an input. The high-fidelity neural network may also include a second high-fidelity neural network (922) that models a nonlinear relationship. The electronic device may learn linear characteristics inherent in complex physical phenomena based on a decoupled structure. Furthermore, the decoupled structure may learn nonlinear characteristics to optimize the performance of the overall prediction model. The first high-fidelity neural network (921) and the second high-fidelity neural network (922) may be configured to learn a residual, which is the difference between a first field value and a second field value. The first high-fidelity neural network (921) and the second high-fidelity neural network (922) may output the learned residual as a correction term.

[0202] The first high-fidelity neural network (921) can be designed to model a linear relationship between input and output. The first high-fidelity neural network (921) can receive latent vectors, coordinates, and low-fidelity predicted values ​​(901) as inputs. The first high-fidelity neural network (921) can calculate a first output value (902) corresponding to a linear correction term. The first high-fidelity neural network (921) can have a relatively simple network structure. The first high-fidelity neural network (921) can effectively capture and correct the overall trend or linear error of the data.

[0203] The second high-fidelity neural network (922) may be designed to model a nonlinear relationship between input and output. The second high-fidelity neural network (922) may receive the same input as the first high-fidelity neural network (921). The second high-fidelity neural network (922) may calculate a second output value (903) corresponding to a nonlinear correction term. The second high-fidelity neural network (922) may have a complex structure, such as a deep neural network. For example, the deep neural network may be a convolutional neural network (CNN), a recurrent neural network (RNN), a transformer, a graph neural network (GNN), an autoencoder, a generative adversarial network (GAN), or a multilayer perceptron (MLP). However, the present disclosure is not limited thereto. The second high-fidelity neural network (922) can precisely correct errors caused by local and complex physical phenomena that a linear model cannot capture.

[0204] The first output value (902) may be a value output from the first high-fidelity neural network (921). The second output value (903) may be a value output from the second high-fidelity neural network (922). The correction term may be calculated based on the first output value (902) and the second output value (903). For example, the electronic device (100) may form the overall correction term by adding the first output value (902) and the second output value (903).

[0205] In one embodiment of the present disclosure, the electronic device (100) can obtain a second field value by adding a correction term to the first field value. Specifically, the electronic device (100) can obtain a final node-level output, i.e., the second field value, by adding the first output value (902) and the second output value (903) to the low-fidelity prediction value (901). This method can combine the speed of low-fidelity prediction with the accuracy of high-fidelity correction. As a result, it is possible to achieve high-precision field prediction while maintaining computational efficiency.

[0206] FIG. 10 is a diagram illustrating a method for an electronic device to perform dynamic field prediction using an artificial intelligence model according to various embodiments.

[0207] Referring to FIG. 10, in step S1010, the electronic device (100) may obtain input regarding a scene including at least one object. The electronic device (100) may obtain input including data (or variables) related to the shape, motion, material properties, or physical quantities of the objects constituting the scene. For example, the input regarding the scene may include, but is not limited to, shape data obtained through 3D scanning technology or 3D modeling technology, motion data (e.g., speed, acceleration, etc.) collected through a sensor, material property data such as the density, strength, or thermal conductivity of the object, or physical characteristic data such as the elastic coefficient, damping, and coefficient of friction.

[0208] Additionally, the input to the scene may include multiple inputs reflecting different characteristics of the object. For example, the input to the scene may include an input related to the shape of the at least one object, an input related to the motion of the at least one object, an input related to a physical property of the at least one object, or an input related to a physical quantity of the at least one object.

[0209] Additionally, the input to the scene may include at least one variable associated with a situation in which at least one object is placed. The input to the scene may include at least one variable for physically representing the situation of at least one object. The input to the scene may include at least one variable for describing the situation of at least one object.

[0210] In step S1020, the electronic device (100) may extract a condition vector corresponding to the scene based on an input regarding the scene. The condition vector is data that numerically expresses the characteristics of an object constituting the scene, and may include information regarding the shape, motion, material properties, or physical quantities of the object. For example, the condition vector may include a shape vector corresponding to the shape of the object, kinematic parameters such as the position, velocity, acceleration, or rotational velocity of the object, etc.

[0211] Additionally, the electronic device (100) can analyze scene input data using a predefined algorithm and generate a condition vector based on this. Furthermore, the condition vector can be automatically generated using an artificial intelligence algorithm. Condition vectors generated in this manner can provide key data necessary for predicting the dynamic behavior of an object.

[0212] An electronic device (100) can extract a condition vector by acquiring an implicit neural representation of at least one object using an artificial intelligence model including an implicit neural network. In this case, the implicit neural network is an artificial intelligence model in the form of an implicit function, and is a model that predicts a performance value at a point (node) in a three-dimensional space containing specific shape information.

[0213] In step S1030, the electronic device (100) can predict the behavior of at least one object at a first point in time based on a condition vector using a pre-trained artificial intelligence model. The pre-trained artificial intelligence model can be trained to predict the dynamic behavior of the object and can predict the behavior of the object by receiving the condition vector as input. At this time, the behavior of the object is a concept including the performance, state, or action of the object, and may include, for example, deformation, stress distribution, reaction force, interaction with a flexible body, etc.

[0214] Additionally, the electronic device (100) can receive a condition vector as input and simulate the dynamic state of an object. Furthermore, the predicted results can be provided to the user via 3D visualization technology. In this way, the electronic device (100) can efficiently predict dynamic behavior in complex scenes containing multiple objects.

[0215] The electronic device (100) can learn an artificial intelligence model using simulation data (or interpretation data) for at least one object. Specifically, the electronic device (100) can learn an artificial intelligence model by configuring learning data based on variables for a scene and simulation data corresponding to the scene.

[0216] The electronic device (100) can perform the following process to learn a performance prediction artificial intelligence model using simulation data (or analysis data). First, the electronic device (100) can give various simulation conditions (e.g., shape, material properties, boundary conditions, load conditions, temperature conditions, etc.) to an object and collect simulation data including corresponding behavior (e.g., displacement, stress, temperature distribution, etc.) or other performance-related indicators. At this time, the simulation data can include result data corresponding to multiple time intervals (time stamps) or different conditions, and can include dynamic characteristics, motion (kinematic parameters), physical quantities (velocity, acceleration, reaction force, stress, etc.) of the object. In addition, the electronic device (100) can map data representing the 3D shape or structure of the object to a latent space, extract a latent vector, and use it as learning data. For example, an artificial intelligence model can apply various neural network structures such as a multilayer perceptron (MLP), a convolutional neural network (CNN), a recurrent neural network (RNN), and a transformer architecture. The electronic device (100) can perform learning by updating weights and biases using an optimization algorithm (e.g., stochastic gradient descent, Adam, RMSprop, etc.) to minimize errors (e.g., L1 or L2 loss, cross-entropy loss, etc.) between the predicted results and actual simulation data while performing iterative learning on the model.

[0217] The electronic device (100) can predict object performance in real-world environments with high reliability by pre-training an AI model for various situations, load conditions, boundary conditions, shape changes, etc. using simulation data. This enables effective predictions through the AI ​​model even when complex analysis processes are difficult to perform in real time, thereby streamlining product design and verification processes or being utilized in monitoring and control systems in real-world operating environments.

[0218] An electronic device according to an embodiment of the present disclosure utilizes a dynamic field prediction model to accelerate real-time flexible body dynamics simulations, thereby predicting interactions between flexible bodies at a significantly faster rate than conventional finite element method (FEM)-based simulations. This enhances the potential for real-time simulations and virtual environments.

[0219] Additionally, the electronic device according to the embodiment of the present disclosure can perform flexible body dynamics predictions for various situations and scenarios using a dynamic field prediction model. The electronic device can effectively express various initial conditions and environmental changes through a condition vector or scenario vector, thereby improving the performance of flexible body dynamics predictions for complex scenarios.

[0220] Furthermore, electronic devices according to embodiments of the present disclosure can internalize physical laws by incorporating physics-based constraints, utilizing Physics-Informed Neural Networks (PINNs). This allows the electronic device to generate prediction results that satisfy physical laws such as conservation of energy and conservation of momentum, thereby enabling physically accurate predictions even with a small amount of training data, thereby improving data efficiency.

[0221] In addition, the electronic device according to the embodiment of the present disclosure can predict interactions between multiple flexible bodies by predicting interactions between multiple objects, and can perform effective simulation and analysis for complex multi-body dynamic systems.

[0222] FIG. 11 is a diagram illustrating a method for an electronic device to extract a condition vector according to various embodiments.

[0223] Referring to FIG. 11, the electronic device can extract a condition vector (condition vector(c)) based on an input for a scene using a plurality of extractors corresponding to a plurality of characteristics.

[0224] Here, the input for a scene may be data expressing elements included in the scene in a specific data format. Specifically, the input for the scene may include data related to the shape of at least one object included in the scene and data related to the movement of at least one object. For example, the input may be image data, video data, sensor data, 3D scan data, text data, voice data, or other data types. However, the present disclosure is not limited thereto.

[0225] The electronic device may include a preprocessing unit (210). The preprocessing unit (210) may perform a preprocessing function to extract a condition vector (Condition Vector(c)) from scene data. The preprocessing unit (210) may be implemented in various forms, for example, as a hardware-based processing device, a software-based processing module, or a combination thereof. However, the present disclosure is not limited thereto.

[0226] The electronic device includes a first extractor (220) and a second extractor (230), and the preprocessing unit (210) can generate a condition vector by extracting a plurality of vectors by separating shape and motion characteristics based on input data. The preprocessing unit (210) can classify the input for the scene into the first input and the second input. For example, the preprocessing unit (210) can classify shape data of at least one object as the first input based on the input for the scene and input it to the first extractor (220), and classify motion data of at least one object as the second input and input it to the second extractor (230), but is not limited thereto.

[0227] In one embodiment of the present disclosure, the preprocessing unit (210) may include a function for preprocessing data acquired from a scene and classifying it into first and second inputs. For example, the preprocessing unit (210) may extract shape data through an image processing algorithm or analyze sensor data to calculate kinematic parameters. Such preprocessing can increase the accuracy of the condition vector and enhance the reliability of dynamic behavior prediction.

[0228] Additionally, the electronic device can input a first input to a first extractor (220) to extract a first vector corresponding to a first characteristic associated with a shape, and input a second input to a second extractor (230) to extract a second vector corresponding to a second characteristic associated with a motion.

[0229] At this time, the electronic device can generate a condition vector by merging the first vector and the second vector. The electronic device can obtain the condition vector by merging the first vector and the second vector based on a merging algorithm (e.g., a concatenate method).

[0230] The first extractor (220) can extract a first vector by mapping a first input corresponding to the shape of at least one object to a latent space. The first extractor (220) can process shape data to generate a representation in the latent space. The first extractor (220) can include, but is not limited to, a convolutional neural network (CNN), a recurrent neural network (RNN), a transformer, or an autoencoder, for example.

[0231] The first extractor (220) can process the input first input and map the shape data to a latent space to generate a unique representation of the shape. The first extractor (220) can analyze the shape data to effectively extract the shape features of the object. In one embodiment of the present disclosure, the first extractor (220) can be designed to process various shape data. For example, the first extractor (220) can process 3D shape data, 2D image data, or point cloud data. Through this, the first extractor (220) can generate a representation in the latent space based on various shape data. In one embodiment of the present disclosure, the first extractor (220) can have a structure that can operate adaptively depending on the resolution or size of the input data. Through this, the first extractor (220) can generate an optimal shape representation depending on the characteristics of the input data.

[0232] In one embodiment of the present disclosure, the first extractor (220) can process shape data using a pre-trained model. The pre-trained model includes weights learned from a large-scale dataset, enabling the first extractor (220) to more accurately extract features of the shape data. Furthermore, the first extractor (220) can provide high generalization performance when processing new shape data based on the learned weights.

[0233] The second extractor (230) can process data corresponding to the kinematic parameters of at least one object to extract a second vector. The second extractor (230) can adjust the data based on the kinematic parameters, thereby contributing to generating a condition vector (Condition Vector (c)). The second extractor (230) can be implemented in various types. For example, the second extractor (230) can include, but is not limited to, a neural network-based model, a statistical model, or a rule-based algorithm.

[0234] The second extractor (230) can extract a second vector by adjusting a second input corresponding to a kinematic parameter of at least one object to a predetermined scale, and the electronic device can generate a condition vector by merging the first vector and the second vector. The second extractor (230) can maintain data consistency and improve the accuracy of the condition vector (Condition Vector (c)) by adjusting the scale of the kinematic parameter.

[0235] In one embodiment of the present disclosure, the second extractor (230) can be extended to include a different scaling method for kinematic parameters or additional data preprocessing steps. For example, the second extractor (230) can perform nonlinear transformation of kinematic parameters or dynamically change the scaling method according to specific environmental conditions. This allows the second extractor (230) to optimize the performance of generating condition vectors (Condition Vector(c)) in various environments.

[0236] In one embodiment of the present disclosure, the second extractor (230) can supplement the kinematic parameters by integrating additional data sources. For example, the second extractor (230) can utilize sensor data to improve the accuracy of the kinematic parameters. This allows the second extractor (230) to generate a more precise condition vector (Condition Vector (c)).

[0237] In one embodiment of the present disclosure, the electronic device may be extensible to accommodate various data types. For example, the electronic device may be designed to process additional data types (e.g., thermal image data, acoustic data, etc.). Such extensions may support more precise dynamic prediction in complex scenes containing multiple objects.

[0238] A condition vector (Condition Vector(c)) is a vector generated based on the shape and motion characteristics of at least one object included in a scene, and can be used as input data for predicting the dynamic behavior of the object. The condition vector (Condition Vector(c)) can be generated by processing data about the scene to extract a first characteristic related to the shape of the object and a second characteristic related to the motion, and then combining them. For example, the condition vector (Condition Vector(c)) can include shape information, motion information, physical characteristics, environmental factors, temporal changes, interaction information, and other related data. However, the present disclosure is not limited thereto.

[0239] In one embodiment of the present disclosure, a condition vector (Condition Vector(c)) can be input into a pre-trained artificial intelligence model and used to predict the dynamic state of an object. The condition vector (Condition Vector(c)) enables field prediction based on 3D data, thereby addressing the lack of scalability of existing prediction methods based on 2D data. Furthermore, the condition vector (Condition Vector(c)) can support efficient prediction of dynamics over time even in complex scenes containing multiple objects.

[0240] In one embodiment of the present disclosure, a condition vector (Condition Vector(c)) can be generated based on data acquired from various sensors. For example, data on the shape, motion, and physical characteristics of an object can be acquired through a camera, LiDAR, ultrasonic sensor, etc., and converted into a condition vector (Condition Vector(c)) to be used for predicting dynamic behavior. This method allows for simulating the dynamic state of an object and providing prediction results through 3D visualization technology.

[0241] An electronic device according to an embodiment of the present disclosure can recognize the situation of a scene based on input about the scene, and can adopt a method of extracting a condition vector corresponding to the situation. The electronic device can determine a dominant characteristic for generating a condition vector describing the situation of the scene based on the situation of the scene, and can extract the determined characteristic to configure the condition vector. Specifically, the electronic device can determine a characteristic to be extracted based on situation information recognized based on input about the scene, and can generate a condition vector using an extractor corresponding to the determined characteristic.

[0242] An electronic device may recognize various scene characteristics (e.g., shape of an object, material of an object, kinematic parameters of an object, etc.) based on input about a scene, and determine at least one key characteristic based on the recognized characteristics. The electronic device may extract at least one vector corresponding to the scene using an extractor corresponding to the determined key characteristic, and obtain a condition vector based on the at least one vector. For example, the electronic device may determine the key characteristic by evaluating the dominance of a plurality of characteristics with respect to a situation. Furthermore, the electronic device may determine the key characteristic by calculating weights of the plurality of characteristics with respect to a situation condition.

[0243] For example, if the initial shape of at least one object included in the scene is diverse, a condition vector can be generated based on the characteristics of the shape. Furthermore, if the initial collision conditions of at least one object included in the scene are diverse, a condition vector can be generated based on the kinematic parameters. Furthermore, if both the initial shape and the initial collision conditions of at least one object included in the scene are diverse, a condition vector can be generated based on the characteristics of the shape and the kinematic parameters.

[0244] In addition, although FIG. 11 illustrates generating a condition vector through two extractors, the present invention is not limited thereto, and the electronic device may further generate a condition vector by using a plurality of extractors each corresponding to various characteristics of the scene (e.g., material properties, boundary conditions, physical parameters (inertia coefficient, etc.)).

[0245] FIG. 12 is a diagram illustrating an example of an electronic device extracting a condition vector according to various embodiments.

[0246] Referring to FIG. 12, in step S1021, at least one processor may acquire a latent vector based on the 3D shape of at least one object. Specifically, at least one processor may convert shape information of the object into a latent vector by utilizing a pre-trained artificial intelligence model. At this time, at least one processor may preprocess the 3D shape data of at least one object. For example, at least one processor may perform a preprocessing process including noise removal, data normalization, or resolution adjustment of shape data.

[0247] In step S1022, at least one processor may extract a condition vector corresponding to the scene based on the latent vector. The at least one processor may analyze the latent vector to generate a condition vector including the shape, motion characteristics, physical characteristics, etc. of the object. For example, the shape characteristics may indicate the size, shape, and surface texture of the object, and the motion characteristics may include the speed, acceleration, and rotation of the object. The physical characteristics may include information such as the density, mass, or material of the object. Such a condition vector may provide key data necessary for predicting the dynamic behavior of the object.

[0248] In one embodiment of the present disclosure, step S1022 may include mapping the latent vector in a multidimensional space to generate a condition vector. For example, at least one processor may map the latent vector from a high-dimensional space to a low-dimensional space to generate the condition vector. This mapping process can reduce the dimensionality of the data, thereby increasing computational efficiency and enhancing the interpretability of the condition vector. Furthermore, the mapping process can be performed using a pre-trained artificial intelligence model, thereby generating consistent condition vectors for various scene data.

[0249] In one embodiment of the present disclosure, step S1021 may include a step in which at least one processor acquires position coordinates of nodes corresponding to a 3D shape. For example, at least one processor may convert 3D shape data into a graph structure and express it as nodes and edges. Nodes may represent specific points on an object, and edges may represent relationships between nodes. Such a graph structure efficiently expresses shape information of an object and can be utilized in the process of generating a condition vector.

[0250] In one embodiment of the present disclosure, step S1022 may include a step in which at least one processor provides input data to a pre-trained artificial intelligence model based on a condition vector. For example, the condition vector may be used as input to the pre-trained model to predict the dynamic behavior of an object. These prediction results may be provided to a user via 3D visualization technology or transmitted to an external device. This enables efficient and accurate dynamic prediction even in complex scenes containing multiple objects.

[0251] FIG. 13 is a diagram illustrating an example of an electronic device predicting the behavior of an object according to various embodiments.

[0252] Referring to FIG. 13, in step S1031, at least one processor can predict a first physical quantity of a first point of at least one object at a first time based on a condition vector. At this time, the first physical quantity can include all physical quantities that can change depending on the interaction of at least one object, such as impact amount, stress, speed, acceleration, temperature of the object, pressure, density, electrical characteristics, magnetic characteristics, or strain.

[0253] In step S1032, at least one processor can predict the behavior of at least one object at a first point in time based on a first physical quantity.

[0254] In this disclosure, "behavior" refers to a concept encompassing the movement, response, and state changes exhibited by an object or system in response to given physical laws or conditions (e.g., external load, self-rotation, fluid interaction, thermal conditions, etc.). This includes changes in physical quantities such as displacement, stress, and temperature observable at a specific point in time or over a continuous time interval, and may represent behavioral aspects of an object that can be predicted by an electronic device based on a condition vector. Various forms such as "dynamic behavior," "rotary behavior," and "fluid-structure interaction behavior" all fall under the category of behavior.

[0255] For example, at least one processor can utilize a pre-trained artificial intelligence model to analyze physical quantity data and predict the behavior of an object. At least one processor can use various algorithms and models to predict the behavior of an object, enabling accurate predictions even in complex multi-body systems.

[0256] At least one processor can input a condition vector into a pre-trained artificial intelligence model to predict a first physical quantity of at least one object at a first point in time. Furthermore, at least one processor can input a condition vector, information about the first point in time, and information about the first point in time into the pre-trained artificial intelligence model to predict a first physical quantity of at least one object at a first point in time.

[0257] Additionally, at least one processor may input all position coordinates corresponding to at least one object along with a condition vector into a pre-trained artificial intelligence model to predict a first physical quantity for the entire region of at least one object. At least one processor may determine the behavior of at least one object based on the first physical quantity for the entire region of at least one object.

[0258] Electronic devices can be designed to operate under a variety of data types and conditions, which can improve the accuracy and efficiency of 3D data-based dynamic behavior prediction.

[0259] FIG. 14 is a diagram illustrating another example of an electronic device predicting the behavior of an object according to various embodiments.

[0260] Referring to FIG. 14, in step S1035, at least one processor can predict field values ​​in the global coordinate system of at least one object at a first point in time based on a condition vector. Here, the field values ​​may include not only the position, velocity, and acceleration of the object, but also additional physical characteristics such as the possibility of collision, rotational velocity, or degree of deformation of the object. Through this, the dynamic behavior of the object can be predicted more accurately even in a complex scene including multiple objects. For example, a plurality of field values ​​in the global coordinate system of at least one object can be calculated based on a condition vector reflecting the shape or motion characteristics of the object included in the scene. Here, the field values ​​are data representing the physical state of the object and can provide basic data necessary for predicting the behavior of the object.

[0261] In step S1036, at least one processor can predict the behavior of at least one object at a first point in time based on field values. For example, if the field values ​​include the object's velocity and acceleration, the object's movement path can be calculated or the possibility of collision can be analyzed. Additionally, if the field values ​​include the object's rotational velocity, the object's rotational behavior can be predicted.

[0262] The above-described prediction steps can each be performed using pre-trained artificial intelligence models and can process various data types (e.g., 3D shape data, 2D image data, etc.). Through this, the present disclosure can overcome the limitations of existing 2D data-based prediction methods and improve the accuracy and efficiency of 3D data-based dynamic behavior prediction.

[0263] FIG. 15 is a diagram illustrating an exemplary method for an electronic device to perform prediction according to various embodiments.

[0264] Referring to FIG. 15, an electronic device can predict the behavior of at least one object included in a scene using a predictor (MLP). At this time, the predictor (MLP) may include an artificial intelligence model having a multi-layer perceptron structure that generates a specific output value based on input data. The predictor (MLP) is generally a neural network-based model and can learn a nonlinear relationship between input data and output data. For example, the predictor (MLP) can be used in various application fields such as regression analysis, classification, time series data prediction, image processing, and natural language processing. However, the present disclosure is not limited thereto.

[0265] An electronic device can input a condition vector (c), a specific coordinate (x) in the global coordinate system, and a previous state (Previous State) into a predictor (MLP), thereby generating a prediction result for the dynamic behavior of an object at a specific point in time. Here, the previous state (Previous State) means data reflecting the behavior (or behavior prediction value) of the object before a specific point in time, and can be expressed as historical data or past observation results.

[0266] An electronic device can perform auto-regressive time-series prediction by setting the predicted value at a previous time point as the input of an artificial intelligence model. Specifically, the predictor (MLP) can predict the state at a second time point (T=t) based on the previous state (Previous State Vector) at a first time point (T=t-1). Specifically, the predictor (MLP) can receive a condition vector (Condition Vector) and a previous state (Previous State) as inputs and predict the dynamic behavior of an object at the second time point (t). Such prediction can be made through a pre-trained artificial intelligence model, and the model can be trained based on simulation data. In this case, the interval between the first time point (t-1) and the second time point (t) can correspond to the timestamp of the trained simulation data.

[0267] In one embodiment of the present disclosure, the predictor (MLP) can additionally predict the state at a third time point (t+1) based on the prediction result at a second time point (t). This structure enables continuous dynamic behavior prediction across multiple time points and enables more precise analysis of object state changes over time. Through this continuous prediction process, the predictor (MLP) can perform efficient and accurate dynamic prediction even in complex scenes containing multiple objects.

[0268] In one embodiment of the present disclosure, the predictor (MLP) may include a data preprocessing process to optimize the characteristics of the input data. For example, processes such as noise removal, normalization, and dimensionality reduction of the input data may improve the learning and prediction performance of the predictor. This preprocessing process can improve computational efficiency by mapping high-dimensional data to low-dimensional data, and enhance the generalization ability of the predictor.

[0269] An electronic device can predict the behavior of an object by reflecting information about a scene that changes over time by setting a condition vector reflecting time-series conditions as input to a predictor (MLP). Specifically, the electronic device can identify kinematic parameters from the scene input and generate a condition vector reflecting time-series conditions by reflecting changes in the object's position over time based on the kinematic parameters. The electronic device can predict the behavior of the object at the current point in time based on the condition vector reflecting time-series conditions and the previous state.

[0270] Global coordinates are a coordinate system used to define a specific point in space, and can express the position and orientation of an object globally. Global coordinates are generally defined based on the X, Y, and Z axes in three-dimensional space, and can express the position and orientation of an object in absolute terms. For example, global coordinates can be used to fix the position of an object to a specific point or to calculate the relative position between objects. However, the present disclosure is not limited thereto.

[0271] The electronic device can predict the dynamic behavior of an object across the global field by inputting a specific point (x) in the global coordinate system to the predictor (MLP), and then calculating a predicted value at a specific point (x) at a specific time. In other words, the electronic device can use the predictor to calculate field values ​​based on 3D data, and can predict the dynamic behavior of the object based on the field values ​​based on 3D data.

[0272] Electronic devices can predict the dynamic behavior of objects by considering interactions between them, even in complex scenes involving multiple objects. Based on the state at time 1 (t-1), the electronic device provides dynamic characteristics such as collisions, rotation, and movement between objects at time 1 to an AI model. By outputting predicted values ​​that reflect the state at the previous time, the electronic device can more accurately predict the state of the object at time 2 (t).

[0273] An electronic device may select at least one primary coordinate in a global coordinate system included in a scene based on predetermined criteria. Specifically, the electronic device may select at least one primary coordinate by selecting a point with a high weight in predicting the behavior of at least one object in a scene including at least one object. The electronic device may select the primary coordinate based on a change in a physical quantity at a point on at least one object. For example, the electronic device may select coordinates corresponding to an expected collision point between multiple objects, a point where stress is concentrated in at least one object, a point where interaction occurs in at least one object, etc., but is not limited thereto.

[0274] An electronic device can dynamically adjust parameters associated with a prediction algorithm based on the coordinates for which behavior is to be predicted. The parameters associated with the prediction algorithm may include time intervals (or intervals, cycles) for time-series prediction, parameters for dynamic prediction, condition vectors, and the like. For example, the electronic device can adjust the prediction cycle based on the coordinates for which behavior is to be predicted, thereby performing time-series predictions at shorter or longer time intervals. Furthermore, for example, the electronic device can determine whether to calculate a condition vector based on the coordinates for which behavior is to be predicted, and can be configured to recalculate the condition vector for key coordinates.

[0275] As a specific example, the electronic device may be configured to change the time interval for time-series prediction when key coordinates are input into the predictor. Specifically, when key coordinates are input into the predictor, the electronic device may adjust the time interval between the first time point (t-1) and the second time point (t) to be shorter. By setting the electronic device to predict at short intervals for points requiring high-precision prediction, the electronic device can produce high-resolution prediction results for key locations. Furthermore, by setting the electronic device to focus computations on key locations, the electronic device can efficiently control computational costs.

[0276] FIG. 16 is a diagram illustrating another exemplary method for an electronic device to perform prediction according to various embodiments.

[0277] Referring to Fig. 16, the predictor (MLP) can predict the behavior (u) of an object by receiving a time value (t), a global coordinate (x), and a condition vector (c) as input. The predictor (MLP) can predict the state of the object at a specific point in time based on the input time value (t), and can calculate the dynamic behavior related to the position of the object by reflecting spatial information through the global coordinate (x). In addition, the predictor (MLP) can perform a prediction that reflects the conditions of the scene by utilizing the condition vector (c). The behavior (u), which is the output of the predictor (MLP), can represent the dynamic field value of the object based on the time value (t), the global coordinate (x), and the condition vector (c).

[0278] In one embodiment of the present disclosure, the predictor (MLP) can be designed based on a multi-layer perceptron structure, and can predict the behavior (u) of an object at successive points in time by processing an input time value (t), a global coordinate (x), and a condition vector (c). The predictor (MLP) can improve computational efficiency by mapping high-dimensional data to low-dimensional data, and can improve the accuracy of prediction by removing noise or normalizing input data through a data preprocessing process.

[0279] The time value (t) can be combined with the global coordinate (x) and the condition vector (c) and used as input to the predictor (MLP). For example, the time value (t) can be an integer, a real number, a timestamp, a time interval, a relative time, an absolute time, or a periodic time value. However, the present disclosure is not limited thereto.

[0280] An electronic device can predict the dynamic behavior (u) of an object by inputting a time value (t) based on a preset time interval. The time value (t) can include absolute and relative time values, providing flexibility to analyze and predict the dynamic state of an object from various temporal perspectives.

[0281] Global coordinate(x) may be a coordinate representing the position of an object in three-dimensional space, a coordinate representing a position on a two-dimensional plane, or a coordinate representing a relative position to a specific reference point. However, the present disclosure is not limited thereto.

[0282] The behavior (u) is the output of the predictor (MLP), and can represent the dynamic field values ​​of the object based on the time value (t), global coordinate (x), and condition vector (c). The electronic device can predict the behavior (u) by predicting the state of the object over time and space.

[0283] In one embodiment of the present disclosure, behavior (u) can be used to simulate the dynamic state of an object. Behavior (u) predicts the dynamic state of an object based on a time value (t), a global coordinate (x), and a condition vector (c), thereby simulating collisions, rotations, and movement paths of the object. Such simulations can be useful for analyzing or predicting the dynamic behavior of an object. Behavior (u) can improve efficiency by mapping high-dimensional data to low-dimensional data, thereby overcoming the limitations of existing 2D data-based methods. Behavior (u) can handle various data types and improve accuracy through data normalization and preprocessing.

[0284] In one embodiment of the present disclosure, the behavior (u) can be the output of a predictor (MLP) and generate field values ​​representing the dynamic state of an object. The behavior (u) interacts with a time value (t), a global coordinate (x), and a condition vector (c) to predict the state of the object, thereby enabling simulation or analysis of the object's dynamic behavior. This structure can contribute to more accurately and efficiently predicting the object's dynamic state.

[0285] The electronic device can perform car crash simulations by utilizing the dynamic behavior prediction technology of the above-described object.

[0286] Electronic devices can accumulate simulation data for numerous vehicle models and various crash scenarios (such as side impacts, frontal impacts, and rear-end collisions). Specifically, electronic devices can store a vast amount of scene information, including different vehicle shapes, materials (material properties), and configurations, along with analysis results (simulation data) that calculate stress and strain distributions for each vehicle component.

[0287] Additionally, electronic devices can generate Latent Vectors by mapping collected vehicle component initial geometry information (e.g., frame, body, door lining, etc.) and physical properties such as material, stiffness, and density to a latent space using an extractor. The extracted Latent Vectors can compressively represent the structural and material properties of the corresponding component or the entire vehicle.

[0288] Additionally, electronic devices can obtain a condition vector (or scenario vector) by quantifying initial conditions such as the impact angle, impact speed, vehicle mass distribution, and passenger occupancy status. The condition vector is a key external condition for crash simulation, and when combined with the Latent Vector obtained above, it can fully represent a car crash scene.

[0289] Furthermore, the electronic device can dynamically predict the deformation and stress distribution of vehicle components according to a crash scenario by providing a condition vector (scenario vector + latent vector) to a pre-trained predictor. Specifically, the electronic device can predict the dynamic behavior of a vehicle in a crash scenario by providing information about the time point and information about the query point as input to the predictor along with the condition vector. This allows the electronic device to perform simulation predictions at a faster speed than conventional physics-based simulations (e.g., FEA) and quickly derive dynamic fields (e.g., stress field, strain field) over continuous time intervals.

[0290] Additionally, electronic devices can visualize predicted crash behavior (behavior fields) to assess safety during the design phase or to support decision-making for the design of enhanced crash structures.

[0291] The electronic device can perform medical simulations by utilizing the dynamic behavior prediction technology of the above-described object.

[0292] Electronic devices can generate 3D models of human tissues and organs by analyzing medical imaging data such as CT and MRI. Then, the shape and material properties (e.g., elastic modulus, viscoelasticity, etc.) of the 3D model can be mapped to a latent space using an extractor to generate a latent vector.

[0293] Additionally, the electronic device can quantify initial conditions regarding the interaction between the surgical tool and the tissue (e.g., insertion location, insertion speed, applied external force, etc.) to obtain a condition vector (or scenario vector). In this case, the condition vector represents external factors that influence tissue deformation in a surgical simulation environment.

[0294] By providing the Latent Vector and Condition Vector to a pre-trained predictor, the electronic device can predict tissue deformation and reaction forces in real time during virtual surgery. This can support the rapid calculation of tissue behavior (strain field, stress field, etc.) over continuous time intervals, enabling its use in surgical technique training and planning.

[0295] Furthermore, electronic devices can provide realistic tactile feedback in surgical simulators or remote surgical environments by linking predicted force data to haptic devices. This allows medical staff to safely practice in an environment similar to the sensation of actual surgery, thereby improving the accuracy and proficiency of surgical outcomes.

[0296] Electronic devices can support simulation and control algorithm development in the field of robotics by utilizing the dynamic behavior prediction technology of the above-described objects.

[0297] Electronic devices can build a database of various robot manipulator and gripper models. By extracting information about the shape, joint structure, and material properties of robot components and expressing them as latent vectors, complex kinematic characteristics can be effectively compressed.

[0298] The electronic device can configure the initial conditions defining the interaction between the robot and the target object (e.g., the position of the target object, the shape of the contact surface, the relative velocity, etc.) as a condition vector (or scenario vector).

[0299] Furthermore, the electronic device can combine the Latent Vector and Condition Vector to predict dynamic field values, such as deformation, reaction force, and vibration, that occur during robot tasks in real time. This enables faster and more accurate dynamic response calculations than conventional physics-based models when handling flexible or deformable materials.

[0300] Electronic devices can efficiently prevent collisions during contact, control gripping strength, and ensure safety by reflecting predicted results into robot control algorithms.

[0301] Electronic devices can perform home appliance drop test simulations by utilizing the dynamic behavior prediction technology of the above-described object.

[0302] Electronic devices can accumulate scene information including various initial conditions such as the drop point, drop angle, and product center of gravity along with simulation data to simulate the impact and deformation that occurs when home appliances such as TVs, refrigerators, and smartphones are dropped from a certain height.

[0303] Additionally, the electronic device can generate a Latent Vector by mapping the product shape and properties to the latent space, and configure external conditions such as drop height and impact surface material as a condition vector.

[0304] By providing the aforementioned latent vector and condition vector to a pre-trained predictor, electronic devices can quickly predict product deformation, stress concentration areas, and potential damage upon impact. This can be used to verify drop impact safety in the early stages of product design and inform decision-making for reinforcement design.

[0305] Electronic devices can utilize the dynamic behavior prediction technology of the above-described objects to perform warhead destructive force and penetration deformation analysis in the defense field.

[0306] By accumulating analysis results for warhead shape, explosive yield, blast pressure, and characteristics of impact targets (e.g., metal plates, composite materials), electronic devices can secure a vast amount of simulation data for various combat scenarios.

[0307] Additionally, the electronic device can generate a Latent Vector by mapping the properties and shape of the warhead and target to the latent space, and set initial conditions such as the collision angle, collision velocity, and explosion temperature as a condition vector.

[0308] Electronic devices can dynamically predict fragment trajectories, penetration depth, and shockwave propagation resulting from explosions or collisions using the aforementioned Latent Vector and Condition Vector. This allows for rapid assessment of the effectiveness and protective performance of defense weapon systems, which can then be reflected in strategic planning or defense system design.

[0309] The electronic device can perform analysis of rotating bodies (motors, propellers, turbines, etc.) by utilizing the dynamic behavior prediction technology of the above-described objects.

[0310] The electronic device can compose a scene by combining the geometric information of the rotating body (e.g. number of blades, profile, etc.), material properties, and working fluid conditions (pressure, temperature, flow rate).

[0311] Additionally, the electronic device can generate a latent vector by mapping the shape and properties of the rotating body to a latent space, and obtain initial conditions such as rotational speed and flow conditions as a condition vector.

[0312] By inputting the aforementioned Latent Vector and Condition Vector into a pre-trained predictor, electronic devices can rapidly predict dynamic behaviors of rotating bodies, such as stress, vibration modes, and thermal distribution, that change over time. This contributes to performance monitoring and defect prediction of rotating bodies, and can be utilized during the design phase to improve durability and optimize efficiency.

[0313] Although the embodiments described above have been described by way of limited examples and drawings, those skilled in the art will appreciate that various modifications and variations can be made based on the above teachings. For example, appropriate results can still be achieved even if the described techniques are performed in a different order than described, and / or components of the described systems, structures, devices, circuits, etc. are combined or combined in a different manner than described, or are replaced or substituted with other components or equivalents.

[0314] Therefore, other implementations, other embodiments, and equivalents to the claims also fall within the scope of the claims described below.

Claims

In electronic devices, A memory implemented to store multiple instructions; and At least one processor electronically connected to said memory and configured to execute at least some of the plurality of instructions stored in said memory; At least one processor, An action of obtaining input for a scene containing at least one object; An operation of extracting a condition vector corresponding to the scene based on input for the scene; An electronic device configured to perform an operation of predicting the behavior of at least one object at a first point in time based on the condition vector using a pre-learned artificial intelligence model. In the first paragraph, An electronic device, wherein the input for the scene comprises at least one of an input associated with a shape of the at least one object, an input associated with a motion of the at least one object, an input associated with a physical property of the at least one object, or an input associated with a physical quantity of the at least one object. In the first paragraph, The operation of extracting the above condition vector is as follows: An electronic device comprising: an operation of classifying input for the scene into a first input and a second input, inputting the first input into a first extractor, extracting a first vector corresponding to a first characteristic associated with a shape, extracting a second vector corresponding to a second characteristic associated with a movement, and extracting the condition vector based on the first vector and the second vector. In the third paragraph, An electronic device characterized in that the first extractor extracts a first vector by mapping a first input corresponding to a shape of the at least one object to a latent space, the second extractor extracts the second vector by adjusting a second input corresponding to a kinematic parameter of the at least one object to a predetermined scale, and extracts the condition vector by merging the second vector with the first vector. In the first paragraph, The operation of extracting the above condition vector is as follows: An operation of obtaining a potential vector based on a 3D shape of at least one object; and An operation of extracting a condition vector corresponding to the scene based on the latent vector; An electronic device characterized in that the at least one processor is set to input the condition vector and the position coordinates of nodes corresponding to the 3D shape at the first point in time into the pre-learned artificial intelligence model. In the first paragraph, The above predicted behavior is, An operation of predicting a first physical quantity of a first point of the at least one object at the first time based on the condition vector; and An electronic device comprising: an operation of predicting the behavior of at least one object at the first point in time based on the first physical quantity; In the first paragraph, The above predicted behavior is, An operation of predicting field values ​​in the entire region of at least one object at the first time point based on the condition vector; and An electronic device comprising: an operation of predicting the behavior of at least one object at the first point in time based on the field values; In the first paragraph, At least one processor, An electronic device further configured to perform an operation of predicting a behavior of at least one object at a second time point based on a behavior of at least one object at the first time point. In paragraph 8, The above pre-trained artificial intelligence model is trained based on simulation data, An electronic device, characterized in that the interval between the first time point and the second time point corresponds to a time stamp of the simulation data. In the first paragraph, At least one processor, An operation of selecting at least one major coordinate in the global coordinates included in the scene based on a predetermined criterion; is further set to be performed; An electronic device characterized in that, when the above key coordinates are input into the above pre-learned artificial intelligence model, the time interval for time series prediction is set to be changed. In the dynamic behavior prediction method, By at least one processor electronically connected to the memory and configured to execute at least some of the plurality of instructions stored in the memory, A step of obtaining input for a scene containing at least one object; A step of extracting a condition vector corresponding to the scene based on the input for the scene; and A method for predicting dynamic behavior, comprising: a step of predicting the behavior of at least one object at a first point in time based on the condition vector using a pre-learned artificial intelligence model. In Article 11, A method for predicting dynamic behavior, wherein the input for the scene includes at least one of an input associated with the shape of the at least one object, an input associated with the motion of the at least one object, an input associated with a physical property of the at least one object, or an input associated with a physical quantity of the at least one object. In Article 11, The step of extracting the above condition vector is: A method for predicting dynamic behavior, comprising: a step of classifying input for the scene into a first input and a second input, inputting the first input into a first extractor, extracting a first vector corresponding to a first characteristic associated with a shape, extracting a second vector corresponding to a second characteristic associated with a movement, and extracting the condition vector based on the first vector and the second vector. In Article 13, A dynamic behavior prediction method, characterized in that the first extractor extracts a first vector by mapping a first input corresponding to a shape of the at least one object to a latent space, the second extractor extracts the second vector by adjusting a second input corresponding to a kinematic parameter of the at least one object to a predetermined scale, and extracts the condition vector by merging the second vector with the first vector. In Article 11, The step of extracting the above condition vector is: A step of obtaining a potential vector based on the 3D shape of at least one object; and A step of extracting a condition vector corresponding to the scene based on the latent vector; A dynamic behavior prediction method, characterized in that the at least one processor is set to input the condition vector and the position coordinates of nodes corresponding to the 3D shape at the first point in time into the pre-learned artificial intelligence model.

Images