Machine-learning for CAD model retrieval
A machine learning-based method using neural networks to assess CAD model compatibility addresses the inefficiencies in existing CAD model retrieval systems, enhancing design and assembly outcomes by ensuring parts with high fit scores are identified and assembled.
Patent Information
- Application Number
- JP2024040661
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2023-05-02
- Filing Date
- 2024-03-15
- Publication Date
- 2024-11-15
AI Technical Summary
Existing CAD model retrieval systems lack efficient methods for determining the compatibility of mechanical parts during the design process, leading to suboptimal assembly and manufacturing outcomes.
A computer-implemented method using machine learning, specifically a neural network trained on boundary representations (B-Rep) of mechanical parts, to determine a fit score for compatibility, enabling efficient retrieval of compatible CAD models.
The method allows for resource-efficient and time-effective retrieval of CAD models with high mating compatibility, improving design efficiency and the physical performance of mechanical assemblies by ensuring parts that fit well together are identified.
Smart Images

Figure 2024160936000001_ABST
Abstract
Description
[Technical field]
[0001] The present disclosure relates to the field of computer programs and systems, and more particularly to methods, systems and programs for machine learning for CAD model retrieval. [Background technology]
[0002] There are many systems and programs available for designing, engineering and manufacturing objects. CAD is an acronym for Computer-Aided Design and refers to software solutions for example to design objects. CAE is an acronym for Computer-Aided Engineering and refers to software solutions for example to simulate the physical behavior of future products. CAM is an acronym for Computer-Aided Manufacturing and refers to software solutions for example to define manufacturing processes and operations. In such computer-aided design systems, the graphical user interface plays a key role regarding the efficiency of the technology. These technologies may also be integrated into product lifecycle management (PLM) systems. PLM is a business strategy that helps companies share product data, apply common processes and leverage corporate knowledge for product development from conception to the end of the product lifecycle, going beyond the concept of the extended enterprise. Dassault Systèmes' PLM solutions (under its trademarks CATIA, ENOVIA and DELMIA) provide an Engineering Hub to organize product engineering knowledge, a Manufacturing Hub to manage manufacturing engineering knowledge, and an Enterprise Hub to enable enterprise integration and connectivity to both the Engineering and Manufacturing Hubs. The system provides an open object model that links products, processes and resources, enabling dynamic, knowledge-based product creation and decision support that drives optimized product definition, manufacturing preparation, production and service.
[0003] In this context, there remains a need for improved solutions for retrieving CAD models from databases. Summary of the Invention
[0004] Accordingly, there is provided a machine learning computer-implemented method for CAD model retrieval based on mating scores, the method comprising: providing a dataset of pairs of boundary representations (B-Rep) representing mechanical parts, each pair being labeled with mating compatibility data that represents a range of mating compatibility between the mechanical parts represented by the pair, and training a neural network based on the dataset, the neural network configured to take as input pairs of B-Rep representing the mechanical parts and output mating scores for the pairs in a single embedding, each single embedding corresponding to a B-Rep of the pair, and the mating score representing a score of mating compatibility between the mechanical parts represented by the pair.
[0005] The method may include one or more of the following. The neural network, a neural network encoder configured to take a pair of B-Reps as input and output a single embedding for the pair of B-Reps; a score neural network configured to output a mating score; has. The neural network encoder comprises a Siamese graph neural network encoder and / or the score neural network is a multi-layer perceptron neural network. The neural network includes a pooling module applied to the output of a Siamese graph neural network encoder and passed as an input to the multi-layer perceptron neural network. The neural network encoder and the score neural network are trained simultaneously using a loss; The loss penalizes the disparity between the fit compatibility data and the fit score.
[0006] There is also provided a neural network trainable according to any of the methods described above.
[0007] Also provided is a method of using the neural network, the method comprising the steps of providing a database of B-Reps representing machine parts, each B-Rep being associated with a single embedding obtained by applying the neural network to the B-Rep.
[0008] This method of use may include one or more of the following: Applying a neural network to B-Rep to obtain a single embedding of B-Rep; Querying the database with single embeddings, including applying a neural network to each pair of the single embedding and one of the single embeddings of each of one or more B-Reps contained in the database; retrieving one or more B-Reps from the database based on a match score obtained from application of the neural network; Includes. Further comprising the step of ranking one or more B-Reps based on their fit score.
[0009] The training method or the method of use may comprise applying the neural network to the B-Reps to output a respective single embedding and performing database indexing, which associates each single embedding with each B-Rep.
[0010] Also provided is a database obtainable by the learning method.
[0011] Also provided is a computer program comprising instructions which, when executed by a computer system, cause the computer system to carry out the method according to the learning method and / or the use method.
[0012] Also provided is a computer-readable recording medium having the program, the neural network and / or the database recorded thereon.
[0013] Also provided is a computer system including a processor coupled to a memory, the memory having the program, the neural network, and / or the database recorded therein. [Brief description of the drawings]
[0014] [Figure 1] An example of a system is shown. [Diagram 2] The method is illustrated below. [Figure 3A] The method is illustrated below. [Figure 3B] The method is illustrated below. [Figure 3C] The method is illustrated below. [Figure 4] The method is illustrated below. [Diagram 5] The method is illustrated below. [Figure 6] The method is illustrated below. [Figure 7] The method is illustrated below. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0015] A computer-implemented machine learning method for CAD model retrieval based on mating scores is proposed. The method (also referred to as the "learning method" or equivalently the "training method") includes providing a dataset of boundary representation (B-Rep) pairs, where the B-Rep pairs represent mechanical parts. Each pair is labeled with mating compatibility data. The mating compatibility data represents a range of mating compatibility between the mechanical parts represented in the pair. The method also includes training a neural network based on the dataset. The neural network is configured to take as input the B-Reps of the pairs representing the mechanical parts. The neural network is also configured to output mating scores for embeddings of the pairs. Each single embedding corresponds to the B-Rep of the pair. The mating scores represent a score of mating compatibility between the mechanical parts represented in the pair.
[0016] Such a method forms an improved solution for CAD model retrieval. In effect, the neural network is configured (i.e., trained) to output a single embedding mating score corresponding to an input pair of B-Reps. Since the mating score represents a score of mating compatibility between the mechanical parts represented by the pair, the neural network is trained to determine how well the mechanical parts represented by the pair of B-Reps assemble together.
[0017] Thereby, the neural network trained by the method may be used to search one or more B-Reps from a database based on the fit score obtained by application of the neural network. In other words, a CAD model search for one or more B-Reps is performed by using the fit score as a search criterion in the database.
[0018] Therefore, the B-Reps obtained from the database may be searched based on respecting the mating compatibility between the mechanical parts, thereby ensuring that the retrieved B-Rep or B-Reps are representative of mechanical parts that can effectively mate in the real world with the input query part. In this way, the method and the neural network form a solution to assist the designer in selecting B-Reps of mating mechanical parts during the design phase of the corresponding assembly of the mechanical parts.
[0019] Moreover, the method not only assists the designer to search for CAD models of mechanical parts with mating compatibility, but also does so efficiently in terms of computing resources (e.g., in terms of CPU and memory usage) and is therefore resource-efficient (the method can be executed on, for example, a GPU). Indeed, the search is performed based on single embeddings and mating scores calculated by a neural network (e.g., single embeddings are output by a Siamese graph neural network encoder, and mating scores are output by a score neural network, which will be described later), and therefore involves processing vectors that belong to a relatively low-dimensional space. This allows for an efficient search in terms of CPU and memory usage, for example, compared to mating compatibility searches based on more geometric methods (e.g., mating compatibility identification based on ray casting, etc.). As mentioned above, the search has the flexibility to be performed on either a CPU or a GPU.
[0020] The usage, or the process of using it in the online phase (described below), brings about time-efficient improvements: in fact, thanks to the single embedding, this usage allows fast queries.
[0021] Moreover, our method is more space-efficient: it relies on computing the "signature" of the B-Rep model, which is relatively lighter than the original model.
[0022] Additionally, new CAD models can be added to the indexed database, for example by applying a neural network to the new CAD models and adding them to the database in association with their respective single embeddings, allowing unseen models to be used to query the database, for example.
[0023] Moreover, the method is data-driven: in fact, the method can help construct pairs and add them to the dataset on which the method was trained.
[0024] The method and / or the method of use can be included in a design process, forming a solution to assist a designer in designing an assembly of mechanical parts corresponding to the mating of two or more mechanical parts. The design process includes the following steps: Optionally, obtaining a neural network (eg by performing the method). Providing a database of B-Reps, each associated with a single embedding. · Querying the database by applying a neural network to each pair of a single embedding of one or more B-Reps contained in the database and one of the single embeddings resulting from applying a neural network to the B-Reps (e.g., by performing the method of use) and searching for one or more B-Reps based on a fit score resulting from the application of the neural network (e.g., the score is the output of a neural network encoder, as described below). Assembling the B-Rep corresponding to the single embedding with one or more (e.g. all) of the retrieved B-Reps, thereby obtaining a B-Rep representing the assembly of the mechanical parts.
[0025] The design process may further include using the resulting B-Reps, which represent an assembly of mechanical parts, to manufacture the assembly of parts.
[0026] The step of using the obtained B-Rep to manufacture an assembly of parts includes any real-world action or series of actions that are involved / participate in the manufacture of the assembly of parts represented by the obtained B-Rep. The step of using the obtained B-Rep to manufacture the assembly of parts includes, for example, the following steps: Editing the obtained B-Reps taking into account their production. Performing simulations based on the obtained B-Reps and / or the corresponding feature trees, such as simulations (e.g. structural simulations, thermodynamic simulations, aerodynamic simulations) for validation of mechanical, usage and / or manufacturing properties and / or constraints. Editing the obtained B-Rep (or its feature tree) based on the results of the simulation. Optionally (i.e. depending on the manufacturing process used, the manufacture of the mechanical product may or may not include this step), (e.g. automatically) determining a manufacturing file / CAM file based on the (e.g. edited) B-Rep (or its feature tree) for the manufacture / production of the manufactured product. Sending the B-Rep and / or manufacturing / CAM files to the factory (e.g. in the form of CAD files), and / or (Based on the determined manufacturing / CAM files or B-Reps, manufacturing / producing (e.g. automatically) an assembly of parts originally represented in the obtained B-Reps. This may include (e.g. automatically) supplying the manufacturing / CAM files and / or CAD files to a machine that performs the manufacturing process.
[0027] This final step of production / manufacturing may be referred to as the manufacturing or production step, in which an assembly of parts is manufactured / fabricated based on the CAD model (e.g., B-Rep) and / or CAM file, e.g., the CAD model and / or CAD file are fed to one or more manufacturing machines or computer systems controlling the machines. The manufacturing step may include performing any known manufacturing process or sequence of manufacturing processes, e.g., one or more additive manufacturing steps, one or more cutting steps (e.g., laser cutting or plasma cutting steps), one or more stamping steps, one or more forging steps, one or more bending steps, one or more deep drawing steps, one or more forming steps, one or more machining steps (e.g., milling steps), and / or one or more punching steps.
[0028] Because it allows a resource-efficient search for mechanical parts that match the entered query part, it not only improves the efficiency and ergonomics of the design (assisting the designer during the design), but also improves the mechanical assembly of the parts manufactured downstream of the design. Indeed, since the method allows the search for the most suitable mechanical parts (or more) to be combined with a given (query) mechanical part from a database, the assembly of parts obtained by assembling these mechanical parts (or rather their B-Reps) during the design will show improved physical performance, in particular improved mating strength. In fact, the assembly results from assembling the parts that, thanks to the method, have been identified as being most suitable for mating.
[0029] The method is computer-implemented. This means that the steps of the method (or substantially all steps) are performed by at least one computer, or any system. The steps of the method are therefore performed by a computer, possibly fully automatically, or semi-automatically. In an example, the triggering of at least some steps of the method is performed through user-computer interaction. The level of user-computer interaction required may depend on the balance between the level of automation foreseen and the need to realise the wishes of the user. In an example, this level may be user-defined and / or predefined.
[0030] A typical example of a computer implementation of the method is executing such a method on a system adapted for this purpose. The system may comprise a processor coupled to a memory and a graphical user interface (GUI), the memory having a computer program recorded thereon, the computer program including instructions for carrying out the method. The memory may store a database. The memory is any hardware adapted for such storage, possibly comprising several physically distinct parts (e.g. one for the program and one for the database).
[0031] FIG. 1 shows an example of a system, which is a client computer system, such as a user's workstation.
[0032] The client computer of this embodiment has a central processing unit (CPU) 1010 connected to an internal communication bus 1000, and a random access memory (RAM) 1070 also connected to the bus. The client computer further comprises a graphical processing unit (GPU) 1110 associated with a VRAM 1100 connected to the bus. The VRAM 1100 is also known in the art as a frame buffer. A mass storage controller 1020 manages access to mass memory devices such as a hard drive 1030. Mass memory devices suitable for embodying computer program instructions and data include all forms of non-volatile memory including, by way of example, semiconductor memory devices such as EPROMs, EEPROMs, flash memory devices, magnetic disks such as internal hard disks and removable disks, and magneto-optical disks. Any of the above may be supplemented by or incorporated in specially designed ASICs (Application Specific Integrated Circuits). A network adapter 1050 manages access to a network 1060. The client computer may also include a haptic device 1090 such as a cursor control device, a keyboard, etc. A cursor control device is used in the client computer to selectively position a cursor at any position on the display 1080. The cursor control device also allows various commands to be selected and control signals to be input. The cursor control device includes a number of signal generators for inputting control signals to the system. Typically, the cursor control device is a mouse, and the buttons on the mouse are used to generate the signals. Alternatively or additionally, the client computer system may include a sensitive pad and / or a sensitive screen.
[0033] A computer program may include instructions executable by a computer, which instructions include means for causing the system to carry out the method. The program may be recorded on any data storage medium, including the memory of the system. The program may be implemented, for example, in digital electronic circuitry, computer hardware, firmware, software, or a combination thereof. The program may be implemented as an apparatus, for example an article embodied in a computer-readable storage medium for execution by a programmable processor. The method steps may be performed by a programmable processor executing a program of instructions that performs the functions of the method by manipulating input data to generate output. Thus, the processor may be programmably coupled to receive data and instructions from, and to transmit data and instructions to, the data storage system, at least one input device, and at least one output device. The application program may be implemented in a high level procedural or object oriented programming language, or in assembly or machine language as appropriate. In any case, the language may be a compiled or interpreted language. The program may be a full installation program or an update program. Application of the program on the system will in any case display instructions for carrying out the method. The computer program may alternatively be stored and executed on a server in a cloud computing environment, which communicates with one or more clients over a network, such that a processor executes instructions constituted by the program to perform the method on the cloud computing environment.
[0034] The learning method is a machine learning for CAD model retrieval based on a fit score. In other words, the method leverages machine learning to provide a solution for retrieving (e.g., retrieving from a database) a CAD model based on a fit score. In yet another way, the fit score is a criterion for CAD model retrieval, for example, through a lookup in a database or other data structure in a non-volatile memory that takes the fit score into account. Such a criterion can be used, for example, by retrieving other fit scores associated with the CAD model in the database or other data structure (e.g., already stored in the database or obtained by performing an appropriate transformation) and determining whether there is a match (exactly or approximately within a tolerance of, for example, 5% or more, e.g., 10%) between the fit score output by the neural network and the other fit scores to return the CAD model.
[0035] As known per se from the field of machine learning, the processing of an input by a neural network involves applying an operation to the input, the operation being defined by data including the weight values. Training a neural network therefore involves determining the values of the weights based on a data set (configured for such training), such a data set sometimes being called a training data set or a training data set. The data set thus includes pieces of data each forming a respective training sample. The training samples are representative of the diversity of situations in which the neural network will be used after training. The data set may consist of a number of training samples of 100 or more, 1000 or more, 10000 or more (e.g. 25000 or more), or 100000 or more.
[0036] The method includes providing a dataset, which may include obtaining such dataset from a data storage medium and / or downloading from a remote location, or alternatively, creating a dataset, which may include annotating B-Reps, as further described below.
[0037] Such a data set is a pair of boundary representations (B-Reps). A B-Rep as described herein is a 3D representation of a mechanical part. In particular, a B-Rep is a persistent data representation of a mechanical part. A B-Rep may be the result of a calculation and / or a series of operations performed during the design phase of the represented mechanical part. The shape of the mechanical part as it appears on the computer screen when the modeled object is rendered may be a B-Rep (e.g., a tessellation).
[0038] B-Rep contains topological entities and geometric entities. Topological entities are faces, edges and vertices. Geometric entities are 3D objects such as faces, planes, curves, lines and points. By definition, a face is a bounded part of a surface called a support surface. An edge is a boundary part of a curve, called a support curve. A vertex is a point in 3D space. They are related in the following way: The boundary part of a curve is defined by two points (vertices) lying on the curve. The boundary part of a surface is defined by its boundary, which is the set of edges lying on the surface. The boundaries of edges of a face are connected by sharing vertices. Faces are connected by sharing edges. Two faces are adjacent if they share an edge. Similarly, two edges are adjacent if they share a vertex. In CAD systems, B-Rep collects the relationship between topological entities and supporting geometry, and the mathematical description of the supporting geometry, in an appropriate data structure, through the "connected by" relationship. An internal edge of a B-Rep is an edge shared by exactly two faces. By definition, boundary edges are not shared and only touch one face. By definition, a boundary face is bounded by at least one boundary edge. A B-Rep is said to be closed if all its edges are interior edges. A B-Rep is said to be open if it contains at least one boundary edge. Closed B-Reps are used to model large 3D volumes, since they define the interior portion of the space that (virtually) surrounds the material. Open B-Reps are used to model 3D skins, which represent 3D objects whose thickness is small enough that it can be neglected.
[0039] The advantage of B-Rep over any other representation type used in CAD modeling is that it can accurately represent any shape. All other representations used, such as point clouds, distance fields, meshes, etc., make an approximation of the shape they represent by discretization. B-Rep, on the other hand, contains surface equations that represent the exact design, and therefore constitutes a true "master model" for further manufacturing, whether it is the generation of toolpaths for CNC or discretization to the appropriate sample density for a given 3D printer technology. In other words, with B-Rep, the 3D model can accurately represent the manufactured object. B-Rep is also advantageous for the simulation of the behavior of the 3D model. For stress, thermal, electromagnetic and other analyses, it supports local refinement of the simulation mesh to capture the physical phenomena, and for kinematics, it supports true contact modeling between curved surfaces. Finally, B-Rep can have a small memory and file footprint. The STEP data format may be used to represent B-Rep in computer systems. It is easy to read, since it is an ASCII structure. Firstly, because this representation only contains a parametric surface. In other representations, such as meshes, the equivalent surface contains up to several thousand triangles. Second, B-Reps do not contain any historical information.
[0040] Besides vertices, faces and edges, each B-Rep may also contain co-edges. Co-edges are oriented edges, i.e. edges with additional orientation information to the edges of the B-Rep. Co-edges thereby define the order around the faces of the B-Rep and their adjacency relations, i.e. the topology of the B-Rep. Co-edges fully link the entities of the B-Rep. Each co-edge has one parent face (i.e. the face for which it defines the boundary), one parent edge (i.e. by definition a co-edge is an oriented edge and therefore has a parent edge), one fellow co-edge that is a co-edge with the same parent edge, one next co-edge (i.e. a co-edge is an oriented edge and therefore has a parent edge), one mating co-edge that is a co-edge with the same parent edge, one next co-edge that corresponds to the next co-edge in the loop to which the co-edge belongs, and one previous co-edge that corresponds to the previous co-edge in the loop to which the co-edge belongs. Each coedge maintains adjacency with its parent face, the mating coedge on the adjacent face, the parent face, and the next and previous coedges around the parent edge.
[0041] A B-Rep may be represented by a B-Rep graph. Such a graph is composed of graph nodes representing the elements (edges, faces, and / or co-edges) of the B-Rep, and graph edges representing the topological relationships between the elements represented by the nodes connected by the edges, and thus represents both the geometry and the topology of the B-Rep. In an example, each graph node may represent a respective edge, a respective face, or a respective co-edge of a respective B-Rep. This means that for each edge of a B-Rep there is a respective graph node, for each face of a B-Rep there is a respective graph node, and / or for each co-edge of a B-Rep there is a respective graph node. In another example, the B-Rep graph may be composed of graph nodes representing only each face and edge of a B-Rep. Alternatively, the B-Rep graph may be composed of graph nodes representing only each edge of a B-Rep. Each graph node may further include geometric and / or topological features associated with (i.e., attached to) the graph node. A feature is data, for example a vector (also called a feature vector), associated with a node that describes the geometric and / or topological data that characterizes the B-Rep element represented by the graph node.
[0042] The B-Rep (e.g., represented in a STEP file format) is converted to a B-Rep graph (e.g., using a STEP file parser if the B-Rep is represented in a STEP file format) or a CAD backend is used to load the B-Rep parts and extract relevant features. For example, the conversion to a B-Rep graph may include mapping B-Rep edges and faces (i.e., excluding topological entities and, e.g., co-edges) to graph nodes (i.e., graph vertices) and mapping the adjacencies of the topological entities to graph edges. The vertex features encode the geometric properties of the corresponding entities. In another example, the conversion may also include passing an additional adjacency matrix indicating the adjacent entities of each type of topological entity.
[0043] A B-Rep can represent the geometry of a mechanical part or assembly of parts (or equivalently, an assembly of parts, which can be considered as the part itself from the point of view of the method, or the method can be applied independently to each part of the assembly), or more generally, any rigid body assembly (e.g., a moving mechanism). Thus, a B-Rep can represent an industrial product, which may be any mechanical part. Mechanical parts are, for example: Ground vehicle parts (including, for example, automobiles and light truck equipment, racing cars, motorcycles, trucks and motor equipment, trucks and buses, trains), aviation vehicle parts (including, for example, airframe equipment, aerospace equipment, propulsion equipment, defense products, aviation equipment, space equipment), naval vehicle parts (including, for example, naval equipment, merchant marine, marine equipment, yachts and workboats, marine equipment), general machinery parts (including, for example, industrial manufacturing machinery, heavy machinery / equipment, installation equipment, industrial machinery products, metal fabrication products, tire manufacturing products), electromechanical or electronic components (including, for example, consumer electronics products, security and / or control and / or instrumentation products, computer and communication equipment, semiconductors, medical equipment and devices), consumer goods (including, for example, furniture, home and garden products, leisure products, fashion products, rigid goods retailer products, flexible goods retailer products), packaging (including, for example, food and beverage, tobacco, beauty and personal care, household product packaging).
[0044] Each pair of B-Reps (in the provided dataset) is labeled with mating compatibility data. By "mating compatibility data" is meant any value (e.g., the value may be a binary or real value, or an alphanumeric value) associated with a pair of B-Reps. The mating compatibility data represents a range of mating compatibility between the mechanical parts represented by the pair, e.g., a binary value or a real (non-negative) value indicating such a range. By "mating compatibility" is meant an indication of the extent to which the mechanical parts represented by both B-Reps can be mated (i.e., assembled). Thus, the mating compatibility data may have a value indicating whether the mechanical parts represented by both B-Reps can be mated together, e.g., 1 if the parts can be mated and 0 if they cannot be mated. The mating compatibility data may also include intermediate values between 0 and 1. For example, mating compatibility data consisting of a value 0 may represent no mating compatibility between both mechanical parts, and a value 1 may represent that the mechanical parts represented by the pair of B-Reps can be mated together. Since B-Reps may mate together (due to compatibility data with value 1), a mate can be defined between two mechanical parts that constrain them. This is the case when the mate compatibility data is a binary value, where a value of zero indicates no mate compatibility and a value of one indicates perfect mate compatibility. Here, the lowest value of the real value range indicates no mate compatibility, the highest value indicates a higher mate compatibility, and values in between indicate a less compatible mate than a higher mate compatibility. It should be understood that the above values are examples and other values and ranges may be used (e.g., 0 and 10 instead of 0 and 1, or the range [0,10] instead of [0,1]). In an example, the mate compatibility data may have a higher mate compatibility value (closer to 1) for a pair of B-Reps that can mate with constraints that greatly reduce the degrees of freedom (e.g., cylindrical constraints, or slider constraints), and a lower compatibility value for a pair of parts that can mate with constraints that still allow a large degree of freedom of movement (e.g., planar constraints).
[0045] In an example, providing the dataset includes annotating pairs of B-Reps and mating compatibility data. For example, providing the dataset may include obtaining pairs of B-Reps (e.g., representing machine parts) from an existing dataset (that does not include prior mating compatibility data).
[0046] In some examples, for a given pair of B-Reps in the existing dataset, annotation may include positively labeling the pair (i.e., creating a mating compatibility data with a value of 1) if the pair was mated together at least once in the existing dataset. Alternatively, annotation may include negatively labeling the pair (i.e., creating a mating compatibility data with a value of zero) by randomly selecting two parts in the existing dataset that are not joined together in an assembly. Thus, the dataset may include not only positively labeled pairs, but also negatively labeled pairs. This increases the robustness of the learning. In fact, negatively labeled pairs correspond to pairs that cannot be mated. The neural network is thus trained to detect that pairs of B-Reps cannot be mated.
[0047] In another example, for a given pair of B-Reps in an existing dataset, the annotation may count the number of unique assemblies that the pair mates with. The annotation can assign higher mating compatibility scores to pairs that are represented in multiple assemblies. In other words, pairs that mate frequently (e.g., a bolt and its nut) are given a higher value. For a given pair, the annotation could look like this:
number
[0048] In these examples, the annotation may negatively annotate a pair if it is not matched together across the dataset.
[0049] The method includes training a neural network based on a dataset. A neural network is a function of a collection of connected nodes, also called "neurons." Each neuron receives inputs and outputs results to other neurons connected to it. The connections between neurons have weights, which are adjusted by training. In the context of this disclosure, "training a neural network based on a dataset" means that the dataset is a learning / training dataset for a neural network, based on which the values of the weights neural network are set. That is, the training determines the values of the weights of the neural network based on pairs of B-Reps and their mating compatibility data.
[0050] The learning may be supervised learning. The method may set at least a part of the data in the data structure of the dataset to be used as ground truth data used for learning in supervised mode. As known from the field of machine learning, the neural network may thus compare its output with the ground truth data, and the weights of the neural network may be adjusted by the learning so that the output of the neural network matches the ground truth data.
[0051] The neural network is configured to take as input pairs of B-Reps representing machine parts. The B-Reps can be taken as input in any manner. In an example, the neural network can take as input a B-Rep graph representation of each B-Rep of the pair. In other words, the neural network can take as input, for each B-Rep of the pair, a graph representation of the B-Rep representing the topological relationships between the elements represented by nodes connected by edges.
[0052] The neural network includes a neural network encoder. A neural network encoder is a neural network configured to determine (i.e., to be trained) a mapping (encoding) between the original input space of the dataset (i.e., the space of B-Reps) and a lower-dimensional space known as the embedding space or latent space. The neural network encoder is configured to take as input pairs of B-Reps (e.g., as B-Rep graph representations) and output a single embedding (e.g., the single embedding is a vector that represents a topological and / or geometric entity) for each B-Rep of the pair. In other words, the neural network encoder outputs a single embedding for each input B-Rep.
[0053] The neural network is configured to output a mating score for a pair of single embeddings, each single embedding corresponding to a pair of B-Reps. By single embedding we mean any data, such as a (multi-dimensional) vector with values representative of the pair of B-Reps, for example their topology and / or geometry. In other words, a single embedding may be any data that allows identification of the topological and / or geometric entities of the pair of B-Reps.
[0054] The neural network may comprise a score neural network. The score neural network is a neural network (e.g., a deep neural network model) configured to output a match score based on its input. The score neural network is configured to take as input a concatenation of a pair of single embeddings and to output a match score. That is, the score neural network is configured to take as input a concatenated single embedding and to output a match score.
[0055] The mating score represents a score of the mating compatibility between the mechanical parts represented by the pair. The score may be a (e.g., non-negative) value representative of the mating compatibility between the mechanical parts represented by the pair of single embeddings. In other words, the mating score output by the method is a value representative of the degree of possibility that the mechanical parts represented by both single embeddings can be combined. In training, the values of the neural network weights are determined based on the dataset, so that the neural network utilizes the mating compatibility data learned by the neural network to output the mating score. In other words, the output mating score depends on the values of the neural network weights set by training. In fact, since the neural network is trained on pairs of B-Reps labeled with mating compatibility data, the mating score matches the range of mating compatibility values captured in the training dataset (i.e., mating compatibility data). This allows the pairs of B-Reps labeled with mating compatibility data to become ground truth data that the neural network is trained to respect when predicting the mating score, and supervised learning can be performed. This allows the neural network weights to be adjusted through training so that the output of the neural network (i.e., the mating score) matches the ground truth data (i.e., the values constructed by the mating compatibility data).
[0056] The method improves CAD model retrieval. In fact, the mating score output by the neural network provides a criterion for retrieving CAD models (in a data structure such as a database) corresponding to the input query B-Rep. For example, CAD model retrieval includes a step of retrieving CAD models (represented by B-Rep) in a data structure by inputting the input query B-Rep. Thus, the CAD model retrieval is highly efficient and returns appropriate results. This is all thanks to the neural network learning method of the method. Meanwhile, the method simultaneously exploits the topology and / or geometry of the B-Rep (represented by the mating compatibility data) and each mating compatibility pair, thus improving the accuracy of CAD model retrieval based on mating compatibility. Meanwhile, the mating score provides an indicator of mating compatibility with the input query part, thus making the CAD model retrieval more efficient.
[0057] Moreover, the accuracy and efficiency of the neural network for CAD model retrieval is further improved when the neural network is composed of a neural network encoder and a score neural network. Indeed, on the one hand, the neural network encoder focuses on extracting topological and / or geometric entities of B-Rep in the low-dimensional representation. On the other hand, the score neural network focuses on concatenating the single embeddings and determining the scores, thus determining the mating scores, which represent the mating suitability of B-Rep. The single embedding is indeed a low-dimensional (in other words, compact) representation of B-Rep, which is less costly in terms of memory requirements (storage, memory footprint) than the corresponding B-Rep model.
[0058] The neural network encoder may be composed of a Siamese graph neural network (GNN) encoder. The Siamese graph neural network encoder may be a neural network encoder that is applied twice with the same weights to each B-Rep of an input pair. That is, in each training iteration, the first B-Rep of the pair is given as input to the Siamese graph neural network encoder, then the other B-Rep of the pair is given as input, the resulting single embedding is concatenated and the score module is applied, and finally the weights of both the encoder and the score module are updated by backpropagation of the score loss function. In other words, the GNN encoder is applied twice during one training iteration, but its weights are updated once at the end. Therefore, by applying the same network twice, the same weights are applied to both parts. In other words, the word "Siamese" simply indicates that there are two different inputs, but one network is applied. In other words, the weights of the Siamese graph neural network encoder (set in training) can be applied to the nodes and edges of the graph of each input B-Rep (listing their topological entities, their geometric features and / or adjacency relationships) to compute each single embedding. In the example, the training updates the weights of the Siamese graph neural network encoder by backpropagating the loss results after both input B-Reps are processed. The Siamese graph neural network encoder can perform any kind of data processing on the input pair of B-Reps to obtain a graph of B-Reps, for example, by applying a STEP parser to the B-Reps if they are in STEP format.
[0059] Additionally or alternatively, the score neural network may be a multi-layer perceptron neural network. A multi-layer perceptron neural network includes fully connected neurons. A multi-layer perceptron neural network is composed of an input layer, one or more (e.g., multiple) hidden layers, and an output layer. The multi-layer perceptron neural network may be trained to perform a mapping between a single embedding and a concatenation of the fit scores, i.e., real-valued values. A multi-layer perceptron neural network may be
number
[0060] This further improves the efficiency of CAD model search. On the other hand, the Siamese graph neural network encoder uses identical weights to output a single embedding for each graph representation (i.e., the B-Rep graph) of the input B-Reps. Thus, the goal of the encoder is to compute a signature of the single embedding, from which the second network (the score module) becomes good at computing scores of mating compatibility, minimizing the loss. As a result, the neural network can more efficiently determine mating compatibility between input B-Reps.
[0061] The neural network may include a pooling module applied to the output of the Siamese graph neural network encoder. The pooling module may have a pooling layer that downsamples the output of the Siamese graph neural network encoder, i.e., the single embedding. The result of the application of the pooling module may be passed as an input to the multi-layer perceptron neural network, i.e., the results of the application of the pooling module for each single embedding may be concatenated and then passed to the multi-layer perceptron neural network. In an example, the neural network encoder may further include a pooling module at the output of the Siamese network, in other words, at the output of the Siamese network for each B-Rep of the pair. In further words, the single embedding for each B-Rep of the pair is the output after the application of the pooling module.
[0062] The use of the pooling module improves the accuracy and efficiency of determining the fit compatibility by ensuring that the single embedding is a low-dimensional representation, which significantly reduces the memory required.
[0063] In an example, a Siamese graph neural network encoder can output multiple vector embeddings for each input B-Rep graph. A pooling module may be applied to each of the multiple vector embeddings to obtain a single embedding. The pooling module can apply multiple pooling operations, e.g., max pooling (calculating the maximum of the coordinates of the single embeddings), sum pooling, or average pooling operations. In other words, the pooling module is either max pooling, sum pooling, or average pooling. The encoder can, for example, first output one embedding per graph node of the B-Rep graph representation (i.e., outputting one vector embedding per B-Rep, e.g., face or edge), so that pooling is applied to have one fixed-size embedding per part.
[0064] A graph representation of a B-Rep containing n topological entities (e.g. faces, edges, etc.) may be constructed and passed to the GNN encoder. Thus, the graph has n nodes. The GNN encoder outputs a vector embedding (1-dimensional tensor) of fixed size (e.g. 256) for each of these nodes, resulting in an embedding of [n×256] for the whole part. The embedding output by the encoder is [m×256]. The pooling module is applied at the end of the Siamese graph neural network encoder (before concatenation). The application of the Siamese graph neural network encoder allows to obtain a [1×256] tensor for each part, regardless of the amount of topological entities n and m (the score neural network receives as input the concatenation of pairs of single embeddings with a constant size). Concatenating these results in a vector of size 512, resulting in a very compact memory usage.
[0065] The neural network encoder and the score neural network can be trained simultaneously with the loss. By "trained simultaneously," we mean that the weights of the neural network encoder and the weights of the score neural network are adjusted simultaneously (i.e., in the same training process) with the loss. In an example, the neural network encoder is a Siamese graph neural network encoder, and the loss can adjust the weights of the Siamese graph neural network encoder simultaneously along with the weights of the score neural network. For example, the loss can output a single loss value and be backpropagated to adjust the weights of the Siamese graph neural network encoder and the score neural network.
[0066] The loss may penalize disparity between the mating fit data and the mating score. The loss may penalize disparity by measuring a distance function between the mating fit data and the mating score. For example, if the mating fit data consists of binary values, the distance function may be binary cross-entropy (the loss is the binary cross-entropy loss). Thus, the mating score is a binary value that matches the binary value of the mating fit data. The binary cross-entropy loss has the form:
number
[0067] Alternatively and / or additionally, the distance function can constitute a Euclidean distance (where the loss is, for example, a mean squared loss) if the fit compatibility data is real-valued, thereby forcing the fit scores to be real-valued. The mean squared error loss takes the form:
number
[0068] This improves the efficiency of CAD model search. The loss indeed improves efficiency by simultaneously updating weights to obtain low-dimensional representations of topological and / or geometric features of pairs of B-Reps, while at the same time forcing the mating scores to be consistent with the mating compatibility data.
[0069] In an example, the method can simultaneously train a Siamese graph neural network encoder that takes pairs of B-Reps and mating compatibility data as input, a pooling module that is applied to the output of the Siamese graph neural network encoder to output a single embedding for each B-Rep of the pair, and a compatibility score module that consists of a score neural network that takes the concatenation of the single embeddings as input and outputs a mating score. In this example, training is done end-to-end. A single loss value is calculated and back-propagated to adjust the weights of the Siamese graph neural network encoder and the compatibility score module. The loss measures the distance between the predicted scores output by the compatibility score module and the mating compatibility data.
[0070] Further provided is a computer-implemented neural network data structure having neural networks trainable (e.g., trained according to the training method) according to the present method, i.e., weights of the neural network trained according to the present method. In other words, the weights of the neural network can be reliably set after training according to the present method has been performed (i.e., without further post-processing of such weights).
[0071] Further provided is a method of using a neural network trained according to the training method, the method including providing a database of B-Reps representing machine parts, each B-Rep in the database being associated with a single embedding obtained by applying the neural network to the B-Rep.
[0072] A process is provided that further includes the step of performing the method and / or the method of use. Performing the method may also be referred to as an "offline phase" and performing the method of use may also be referred to as an "online phase." The offline phase and the online phase may be performed in different systems, different entities, or different locations.
[0073] "Database" means any collection of data (i.e., information) organized for search and retrieval (e.g., a relational database based on a predefined structured language, e.g., SQL). A database stored in memory allows for rapid search and retrieval by a computer. A database is structured to facilitate the storage, retrieval, modification, and deletion of data in conjunction with various data processing operations. A database contains one or more fields. It contains a file or set of files that can be broken down into records. A field is the basic unit of data storage. A user can retrieve data primarily through queries. Using keywords and sorting commands, a user can quickly find, sort, group, and select fields of many records to search or create reports on specific data aggregations according to the rules of the database management system being used.
[0074] Further provided is a database obtainable in accordance with this method of use, i.e. a collection of data organized for searching and retrieval, where the data pieces consist of (e.g. consist of) respective B-Reps and are associated with respective single embeddings.
[0075] By "related" we mean that the B-Rep and the single embedding are related and have a relationship in the database, i.e., by applying a neural network to the B-Rep, there is a mutual link between the two objects in the database. However, the association of a B-Rep with its single embedding in the database is a matter of database implementation and need not be discussed further.
[0076] The method of use may include obtaining a single embedding of B-Rep (also referred to as a "new B-Rep") by applying a neural network to B-Rep. In an example, the neural network may include a Siamese graph neural network encoder and a score neural network trained according to the training method. The method of use may apply the Siamese graph neural network encoder to B-Rep and obtain a single embedding from the Siamese graph neural network encoder.
[0077] The method of use may also include a step of querying the database using the single embedding. The query includes applying a neural network to each pair consisting of a single embedding and one of the single embeddings of each of one or more B-Reps included in the database. In an example (e.g., where the neural network may be composed of a Siamese graph neural network encoder and a score neural network), the method may set the single embedding as part of the query (e.g., via an SQL instruction) and search the database using the single embedding. The method of use may, for example, apply the neural network to one or more B-Reps included in the database to obtain each single embedding (e.g., at the output of the Siamese graph neural network encoder, or at the output of the pooling module when such B-Rep is not associated with each single embedding) when executing the query, i.e., while traversing the database of B-Reps with the query. The score neural network may take as input the concatenation of a set of single embeddings (i.e., the single embedding obtained from the new B-Rep and each of one or more B-Reps included in the database) and output a corresponding fit score.
[0078] The method of use may also include retrieving one or more B-Reps from the database based on the mating score obtained from application of the neural network. For example, the method of use may output the B-Rep from the database that has the highest mating score for the new B-Rep. Alternatively, the method may output multiple B-Reps that have a score above a predetermined threshold, for example. In an example, if multiple B-Reps from the database have the same highest mating score (up to a numerical tolerance), the method may output all of the B-Reps.
[0079] As a result, B-Reps retrieved from the database are searched for their mating compatibility using the new B-Rep as a reference, making the search for B-Reps very efficient.
[0080] The method of use may include a step of ranking one or more B-Reps based on the fit score, so that the method may provide suitable results. In practice, the method of use may rank one or more B-Reps in descending order of fit score to prioritize more relevant retrieved B-Reps. The B-Reps are displayed to the end user, for example, according to the order established by the ranking.
[0081] The method or method of use may further comprise a step of performing database indexing. Database indexing means a set of operations that allows storing B-Reps in association with their respective data in the database. Database indexing may be performed by applying a neural network to the B-Reps to output each single embedding (e.g. from the output of a neural network encoder or from the output of a pooling module). Database indexing may associate each B-Rep with each single embedding. In other words, database indexing may store the B-Rep in the database in association with its respective single embedding. This process may also comprise a step of performing database indexing. Database indexing may be part of an offline stage. The method of use may comprise obtaining one or more (e.g. a plurality) B-Reps from an existing database. In this way, the method of use may create a database (e.g. a completely new database, if none existed before) consisting of one or more B-Reps, each associated with a respective single embedding.
[0082] The method can, for example, query a database with a single embedding against a database with a single embedding obtained, which improves search efficiency compared to existing databases and allows searches based on a low-dimensional representation with a single embedding.
[0083] Next, an example of the method will be described with reference to FIGS.
[0084] Figure 2 illustrates a pipeline 2000 of the method. The method includes an offline training stage 2100 of a neural network according to the method, and an offline database indexing 2200, i.e., an offline phase of creating a database of B-Reps, each B-Rep being associated with a signature. The pipeline 2000 also includes an online stage 2300, i.e., a method of use.
[0085] Next, the offline learning stage 2100 will be described.
[0086] The offline training 2100 aims to provide a database 2110 of B-Rep pairs with their mating compatibility data ("compatibility labels"). The offline training simultaneously trains a Siamese B-Rep Encoder 2120 that takes as input pairs of B-Rep parts and outputs their respective (multi-dimensional) vector embeddings (also called "signatures") 2121 and 2122, and a compatibility score module (which is a score deep neural network (DNN) module) 2140 that takes as input a concatenation 2130 of vector embeddings 2130 with annotations associated with the input pairs of B-Rep parts. The compatibility score module 2140 outputs a compatibility score 2150. The offline training 2100 is end-to-end. A single loss value is computed and back-propagated to adjust the weights of the two modules 2120 and 2140.
[0087] Since the offline learning stage 2100 is supervised learning, we describe the annotation of data samples, i.e., labeling them with mating compatibility data. The data samples are pairs of CAD models in B-Rep format (e.g., STEP), annotated with numerical values that measure the compatibility / relatedness of two parts that go together in an assembly.
[0088] Data Preprocessing. Data samples may be preprocessed so that they can be passed to the DNN architecture.
[0089] Architectural design. The Siamese B-Rep encoder returns one fixed-size vector embedding for each part. The score module takes as input the concatenation of the two embeddings of a pair output by the Siamese B-Rep encoder and outputs a compatibility score.
[0090] Training Loss. The training loss function measures the distance between the predicted scores output by the score module and the paired target mating compatibility data.
[0091] Training 2100 is performed over several iterations on the annotated dataset.
[0092] Next, offline database indexing 2200 is described. The trained deep learning module can be used to generalize to unseen data (data samples other than the training dataset). The trained deep learning module can be used to create a database that associates B-Reps with single embeddings (e.g., from an existing database). Thus, offline indexing 2200 applies the trained deep learning module and stores the output vector single embeddings (e.g., resulting from outputs 2121 and 2122).
[0093] Next, the online stage 2300 will be described.
[0094] Given an input query B-Rep part, the online stage 2300 comprises extracting from the database indexed in 2200 the most relevant parts that match the query.
[0095] The online stage 2300 includes the following subtasks:
[0096] Obtain a single embedding / signature of the query vector. Since the query B-Rep may not necessarily exist in the database indexed in 2200 (e.g., it may be a part just designed by a CAD user), the online stage 2300 can apply the learned deep learning module 2100 to obtain the corresponding single embedding 2310.
[0097] Concatenate the single embeddings of all (query part, index part) pairs. A concatenation between the query part signature obtained in 2310 and each embedding signature indexed in the database 2200 is performed, ready to be passed to the learned score module 2140. Thus, if the indexed database contains n parts, the concatenation 2320 is obtained by n concatenations.
[0098] At 2320, the learned fit score module 2140 is applied to the resulting connections.
[0099] The indexed parts 2340 are sorted in descending order of predicted compatibility score. In this way, for the top k parts, the method returns the top k indexed parts that match the query.
[0100] Data annotation and labeling using fit-match data for supervised learning will be described in more detail below.
[0101] Here we present some examples of pairwise annotations of fit compatibility data that are suitable for performing the desired task: Given a dataset containing CAD assemblies, CAD parts in B-Rep format, and fits / joint data for assembling them into a multi-part assembly (defining the fits between the parts of the assembly), the following approach is possible: Approach A: We label as positive (set value to 1) part pairs that have mating data in the dataset, i.e., pairs of parts that mated together at least once in one of the assemblies of the dataset. To create negative samples (set value to 0), we randomly select two parts from the entire dataset such that they are never mated together in one assembly. Another implementation parameter is the balance between positive and negative samples in the training set. Approach B: For each pair of parts that are mated together in the dataset, count the number of unique assemblies this pair is mated to, and assign higher mating compatibility scores to pairs that are represented in multiple assemblies. That way, pairs that mate well (e.g., screws and nuts) will have higher values. For a given pair, the mating compatibility data looks like this:
number
[0102] The data pre-processing of the B-Rep parts will be further described with reference to Figures 3A to 3C. Figure 3A is a diagram showing the data pre-processing in a schematic manner, and Figures 3B to 3C are further perspective views of the exploded parts shown in Figure 3A.
[0103] The data pre-processing applied to the B-Rep part may rely on a DNN encoder, which is covered in the following section. For example, the neural network implementation may be based on a message passing based encoder, such as the Siamese Graph Neural Network B-Rep Encoder 3100, which takes as input a graph B-Rep Representation 3200 (B-Rep Graph). Such a B-Rep Graph can be constructed using a STEP file parser or by using a CAD backend to load the part and extract the relevant features. For example, the method can construct a graph where the vertices are the topological entities of the edges and faces of the object, and the graph edges are defined by the adjacencies of these topological entities. The vertex features encode the geometric properties of the corresponding entities (such as the type of edge or face, the area of the face, the length of the edge, etc.). It may also be necessary to additionally pass an adjacency matrix indicating the neighboring entities of each type of topological entity.
[0104] We further explain the design of neural networks.
[0105] The neural network converts vector embeddings into a single embedding per object.
number
[0106] The Mating Score module
number
[0107] Learning loss is explained in more detail below.
[0108] The learning loss is linked to the mating compatibility data, since it is necessary to measure the distance between the predicted score values and the labeled compatibility scores.
[0109] When using binary labels (approach A), the learning task becomes a binary classification problem, which can be trained using a binary cross-entropy loss.
number
[0110] For other labelling approaches, such as approach B, the mean squared error loss is a good fit.
number
[0111] Indexing of offline datasets is further explained with reference to FIG.
[0112] To perform database indexing, a trained encoder 4200 with graph pooling can be applied to the B-Reps in the existing database 4100 and stored in association with an output single embedding 4300. Thus, each B-Rep can be mapped to its fixed size vector single embedding. The method may also perform a data pre-processing routine (to obtain a graph B-Rep for each B-Rep) so that a graph neural network can be applied to the samples. When the trained encoder was used for experiments, indexing the parts database took about 15 seconds for about 3600 parts while running inference mode on the CPU.
[0113] The online query / inference stage is further described with reference to FIG.
[0114] The input query portion 5100 is concatenated with one embedding stored in the database. This concatenation is input to a score module 5200. The method can rank the retrieved match scores 5300.
[0115] Executing the query and generating a fit score between the query part and ~3600 index parts using 2140 took ~0.4 seconds on the CPU. Thus, the method can be considered production-ready. In fact, it allows online queries in less than 1 second, can be integrated into commercial CAD products, and can be seamlessly integrated into the CAD designer's workflow.
[0116] 6 and 7 are search results for CAD models. 6100 is the input of the query. The method obtains 6200, which are the top four parts with the best mating scores for 6100. 7100 is also the input of the query. The method obtains 7200, which are the top four parts with the best mating scores for 7100.
Claims
1. 1. A machine learning computer-implemented method for CAD model retrieval based on fit scores, comprising: providing a dataset of boundary representation (B-Rep) pairs representing mechanical parts, each pair being labeled with mating compatibility data, the mating compatibility data representing a range of mating compatibility between the mechanical parts represented by the pair; training a neural network based on the dataset, the neural network configured to take as input pairs of B-Reps representing mechanical parts and output a mating score for the pair of single embeddings, each single embedding corresponding to a B-Rep of the pair, the mating score representing a score of mating compatibility between the mechanical parts represented by the pair; 23. A computer-implemented method comprising:
2. The neural network, a neural network encoder configured to take a pair of B-Reps as input and output a single embedding for the pair of B-Reps; a score neural network configured to output a mating score; The computer-implemented method of claim 1 , comprising:
3. The neural network encoder includes a Siamese graph neural network encoder, and / or the score neural network is a multi-layer perceptron neural network. The computer-implemented method of claim 2 .
4. The neural network includes a pooling module applied to an output of a Siamese graph neural network encoder and passed as an input to the multi-layer perceptron neural network.
4. The computer-implemented method of claim 3.
5. the neural network encoder and the score neural network are trained simultaneously using a loss; The loss penalizes the disparity between the mating compatibility data and the mating score. A computer-implemented method according to any one of claims 2 to 4.
6. A neural network trainable according to the method of any one of claims 1 to 5.
7. A method for using a neural network according to claim 6, comprising the steps of: providing a database of B-Reps representing machine parts; Each B-Rep is associated with a single embedding, obtained by applying a neural network to the B-Rep. How to use neural networks.
8. applying a neural network to B-Rep to obtain a single embedding of B-Rep; Querying the database with single embeddings, including applying a neural network to each pair of the single embedding and one of the single embeddings of each of one or more B-Reps contained in the database; retrieving one or more B-Reps from the database based on a match score obtained from application of the neural network; The method of claim 7, comprising:
9. Further comprising the step of ranking one or more B-Reps based on the fit score.
9. The method of claim 8.
10. applying the neural network to the B-Reps to output a respective single embedding, performing database indexing, the database indexing associating each single embedding with each B-Rep.
10. The method of any one of claims 1 to 5 or 7 to 9.
11. A database obtainable by the method of claim 10.
12. A computer program comprising instructions which, when executed by a computer system, cause the computer system to perform the method according to any one of claims 1 to 5, the method according to any one of claims 7 to 9 and / or the method according to claim 10.
13. A computer-readable recording medium having recorded thereon the program according to claim 12, the neural network according to claim 6, and / or the database according to claim 11.
14. A computer system including a processor coupled to a memory, the memory storing the program according to claim 12, the neural network according to claim 6, and / or the database according to claim 11. Computer system.
