Computer-implemented method for training an artificial intelligence model intended to predict physical features of a brake caliper, prediction method and related systems

Abstract

A computer-implemented method trains an artificial intelligence model based on neural networks to predict physical features of a brake caliper from a geometric representation, such as a mesh. The method provides for receiving geometric representations of brake calipers, each comprising local regions such as vertices and faces, and local information relating to the geometry. The artificial intelligence model is trained using input data comprising the received local information and output data relating to physical features of the brake caliper, including fields of stresses due to the pressure of brake fluid and the friction force, and the integral of the elastic deformation energy. The system comprises an input interface to access training data and a processor module configured to train the artificial intelligence model. The trained artificial intelligence model predicts physical features of a brake caliper from the geometric representation, providing output data such as stress fields and deformation energy.

Description

" COMPUTER- IMPLEMENTED METHOD FOR TRAINING AN ARTIFICIAL INTELLIGENCE MODEL INTENDED TO PREDICT PHYSICAL FEATURES OF A BRAKE CALIPER, PREDICTION METHOD AND RELATED SYSTEMS"DESCRIPTIONField of Application

[0001] The present invention generally falls within the field of simulation of the mechanical characteristics of braking systems for vehicles. Moreover, the present invention generally relates to the field of artificial intelligence or machine learning, and more specifically to the application of artificial intelligence or machine learning techniques for predicting physical and mechanical properties of complex mechanical components.

[0002] In particular, the subject matter of the present invention is a computer-implemented method for training an artificial intelligence or machine learning model based on neural networks intended to predict physical features of a brake caliper from its geometric mesh, and a related system. Moreover, the subject matter of the present invention also includes a computer-implemented method for predicting physical features of a brake caliper and a related system.

[0003] Furthermore, the subject matter of the present invention includes a computer program for executing sucha method and a computer product in which said program is stored.State of the Art

[0004] In the automotive industry, the design and optimization of brake components are crucial to ensure performance and safety. Accurate prediction of the physical and mechanical characteristics of brake calipers is essential to improve design and reduce development time.

[0005] Traditional numerical simulations, such as those based on the finite element method (FEM), are commonly used to analyse the performance of brake calipers. However, these simulations require a considerable expenditure of time and computational resources.

[0006] Traditional numerical simulations, such as those based on the finite element method (FEM), are computationally very intensive; therefore, only a limited number of input configurations can be tested within a reasonable time frame. This limits the ability to efficiently and rapidly optimise the design of brake calipers.

[0007] The use of surrogate models based on artificial intelligence offers an opportunity to improve design processes, reduce development time and optimise the useof computational resources.

[0008] Surrogate models can approximate the output of FEM simulations without numerically solving the equations, allowing a greater number of input configurations to be tested in reasonable time.

[0009] In this context, document EP3757903A1 describes a methodology that provides for training a neural network based on previous simulations in order to optimise the geometry of a physical object based on a particular metric. The aim is to identify the most promising geometries to then perform FEM simulations, thereby reducing the number of FEM simulations, which entail significantly higher computational costs and times. This solution uses convolutional and linear layers and models such as PointNet and PointNet++ applied to point clouds. This solution, by reducing the number of geometries to be subjected to the FEM process, does not, however, fully eliminate the execution of FEM simulation, which still remains the bottleneck in terms of timing for the final determination of the physical parameters to be simulated.

[0010] Document W02020055659A1 describes a methodology that uses two neural networks. The first network classifies the object to decide the type of simulation to be performed among high-fidelity, low-fidelity or surrogate model. The surrogate model, a second neuralnetwork, replaces the FEM simulation based on the output of the first network. However, this methodology does not entirely replace the FEM simulation, but only under certain conditions.Summary of the Invention

[0011] There is therefore a need for a method and a system capable of efficiently and accurately predicting the physical properties of brake calipers using a surrogate model based on neural networks, reducing the time and computational resources required for FEM simulations and enabling faster design iterations.

[0012] This and other objects are achieved by means of a computer-implemented method for training an artificial intelligence model based on neural networks intended to predict physical features of a brake caliper from a geometric representation thereof, for example from a geometric mesh, and a related system, and by means of a computer-implemented method for predicting physical features of a brake caliper and a related system, in accordance with the appended independent claims. The dependent claims describe preferred or advantageous embodiments of the invention, involving further advantageous aspects.

[0013] According to one aspect of the present invention, a computer-implemented method for training anartificial intelligence or machine learning model based on neural networks intended to predict physical features of a brake caliper from a geometric representation thereof, for example from a geometric mesh, comprises performing the following steps on an electronic processor:

[0014] a) receiving geometric representations of brake calipers, each geometric representation comprising one or more local regions, for example vertices and faces of a geometric mesh;

[0015] b) receiving local information relating to the geometry of each brake caliper;

[0016] c) training an artificial intelligence model based on neural networks intended to predict the physical properties of an input geometric representation of brake caliper provided as input data to the artificial intelligence model, said step of training the artificial intelligence model being performed with input data comprising the local information received and output data relating to physical features of the brake caliper comprising at least one of the following:

[0017] - a field of stresses due to the pressure of a brake fluid on the brake caliper and to the friction force generated between pad and brake disc, said field of stresses being obtained for each brake caliper fromfinite element simulations and comprising a value of the stresses for one or more local regions of the input geometric representation, for example for each vertex and / or face of the input geometric mesh;

[0018] - integral of the elastic deformation energy of the brake caliper.

[0019] According to another aspect, the method provides that the local information includes thicknesses and / or angles and / or curvatures of local regions of the brake caliper.

[0020] According to a further aspect, the method provides that the local information includes at least the normals at the vertices and the local curvature calculated locally at the vertices.

[0021] According to another aspect, the method comprises, before the step of receiving local information, performing the following steps to determine the local information relating to the geometry of each brake caliper: applying geometric transformations to the coordinates of the vertices to make uniform the orientation of the brake calipers in space and align the centre of gravity of the brake calipers, for example aligning it with the origin of the axes of the three-dimensional virtual space (xyz); calculating local information relating to the geometry of each brakecaliper said local information comprising at least normals at the vertices and locally calculated local curvature at the vertices.

[0022] According to a further aspect, the method comprises the step of receiving one or more of the following pieces of global information, relating to global geometric properties: effective radius of the caliper; number of pistons, total area of pistons, caliper fixing type (axial or radial), outer diameter of the disc, overall thickness of the disc, whether or not the tie rod is present, pad support type (flat or pins), centre-to-centre distance: distance of the centre between the holes of the two screws to constrain the caliper to the hub carrier, height: z-coordinate of the axes of the screw holes, maximum radius: maximum radius considering the coordinates of the points of the caliper in a cylindrical system centred at the origin, theta: angular span of the points considering the coordinates of the points of the caliper in a cylindrical system centred at the origin; and wherein in the training step, the input data comprise at least one or more of said pieces of global information.

[0023] According to another aspect, the method provides that the artificial intelligence model includes a DiffusionNet neural network architecture.

[0024] According to a further aspect, the method provides that the field of stresses comprising a value of the stresses for each vertex of the geometric mesh is obtained by means of interpolation techniques to align the output data of finite element simulations with the geometric mesh of each brake caliper.

[0025] According to another aspect, the method provides that the geometric transformations applied to the coordinates of the vertices include rotations and translations to make uniform the orientation of the brake calipers in space.

[0026] According to a further aspect, the method provides that the artificial intelligence model is trained using an adapted loss function for regression problems.

[0027] According to another aspect, the method provides that the artificial intelligence model is trained using a neural network architecture comprising a diffusion layer which utilises the heat equation to propagate information through the surface of the brake caliper.

[0028] According to a further aspect, a computer-implemented method for predicting physical features of a brake caliper from its input geometric mesh comprises performing the following steps on one or more electronicprocessors:

[0029] - receiving a trained artificial intelligence model as obtained by the method according to any one of the preceding aspects;

[0030] - predicting the physical features of the brake caliper using the trained artificial intelligence model, said predicting comprising: using the input geometric representation as input data for the trained artificial intelligence model;

[0031] - providing as output data from the trained artificial intelligence model a prediction of at least one physical feature of the brake caliper.

[0032] According to another aspect, a system for training an artificial intelligence or machine learning model based on neural networks intended to predict physical features of a brake caliper from its geometric mesh comprises:

[0033] - an input interface to access: training data comprising a plurality of training instances representing different geometric meshes of brake calipers, each training instance comprising: a geometric mesh of a respective brake caliper, comprising vertices and faces; local information relating to the geometry of each brake caliper; - data defining an artificial intelligence model configured to receive a respective geometric mesh asinput and to generate an output of the artificial intelligence model;

[0034] - a processor module configured to train the artificial intelligence model using the training data to obtain a trained artificial intelligence model intended to predict the physical properties of an input geometric mesh of brake caliper provided as input data to the artificial intelligence model, said training being performed with input data comprising the local information received and output data relating to physical features of the brake caliper comprising at least one of the following:- a field of stresses due to the pressure of a brake fluid on the brake caliper and to the friction force generated between pad and brake disc, said field of stresses being obtained for each brake caliper from finite element simulations and comprising a value of the stresses for each vertex of the geometric mesh;- integral of the elastic deformation energy of the brake caliper.

[0035] According to a further aspect, a system for predicting physical features of a brake caliper from its input geometric mesh comprises:

[0036] - an input interface for accessing data defining a trained artificial intelligence model asobtained by the computer-implemented method according to any one of the embodiments described in the present document;

[0037] - a processor module configured to predict the physical features of the brake caliper using the trained artificial intelligence model, said prediction comprising:

[0038] - using the input geometric mesh as input data for the trained artificial intelligence model;

[0039] - providing as output data from the trained artificial intelligence model a prediction of at least one physical feature of the brake caliper.Description of the Drawings

[0040] The features and advantages of the computer-implemented method for training an artificial intelligence model and related system, of the computer-implemented method for predicting physical features of a brake caliper and related system, according to the present invention, will however be apparent from the following description of some preferred embodiments, given by way of example and not limitation, with reference to the accompanying figures, in which:- Figure 1 shows a diagram of the computer-implemented method for training an artificial intelligence model, in accordance with one embodiment of the present invention;Figure 2 shows a diagram of a computer-implemented method for predicting physical features of a brake caliper from its input geometric representation, for example a geometric mesh (100), in accordance with one embodiment of the present invention;- Figure 3 shows an example of a point-based architecture of an artificial intelligence model in accordance with one embodiment of the present invention;- Figure 4 shows another example of an architecture of an artificial intelligence model in accordance with one embodiment of the present invention;- Figure 5 shows an example of a geometric representation of a brake caliper and of some global parameters, in accordance with one embodiment of the present invention; - Figure 5a shows a diagram of a system 400 for training an artificial intelligence or machine learning model based on neural networks intended to predict physical features of a brake caliper, in accordance with one embodiment of the present invention;- Figure 6 shows a diagram of a system for predicting physical features of a brake caliper, in accordance with one embodiment of the present invention;- Figure 7 shows, in order from left to right, a chart of the field of stresses predicted by the artificial intelligence model for a brake caliper, a chart of thereference field of stresses simulated for the same brake caliper, and a chart of the stress error calculated as the difference between the predicted field of stresses and the reference field of stresses;- Figure 8 shows a graph representing the percentage errors of the deformation energy, wherein each dot in the graph is calculated as the percentage difference between a value of deformation energy predicted by the artificial intelligence model for a given brake caliper and a reference value of deformation energy for the same brake caliper simulated with a finite element method; in particular, the data series indicated by the number 0 relates to brake caliper data present only in the test dataset, while the data series indicated by the number 1 relates to brake caliper data for which there are examples also used in the training dataset of the artificial intelligence model.Detailed Description

[0041] The computer-implemented method for training an artificial intelligence model in accordance with the present invention is directed to the training of an artificial intelligence model 1 based on neural networks intended to predict physical features of a brake caliper from a geometric representation thereof, for example from its geometric mesh 100.

[0042] In the following discussion, without loss of generality, reference will be made to a geometric mesh 100 of a brake caliper, but it shall be understood to refer to any geometric representation of a brake caliper. For example, an alternative to the geometric mesh may be a set of specific geometric parameters (such as thicknesses, angles, curvatures) configured to create a set of simplified input variables for the artificial intelligence model. Therefore, where reference is made to a geometric mesh 100 or a 3D mesh, it shall also be understood, in a substitutive manner, as a generic geometric representation of a brake caliper.

[0043] The computer-implemented method for training an artificial intelligence model comprises a step a) of receiving geometric meshes of brake calipers 100, each geometric mesh of said geometric meshes comprising vertices and faces, and a step b) of receiving local information relating to the geometry of each brake caliper.

[0044] Preferably, each brake caliper is described by its geometry, represented by a 3D mesh comprising vertices (xyz coordinates in space) and faces (vertex connectivity).

[0045] In accordance with one embodiment, each geometric mesh 100 of said geometric meshes is a mesh file, forexample an .stl file (Standard Triangle Language). For example, a file with the following structure:

[0046] solid name

[0047] facet normal ni nj nk

[0048] outer loop

[0049] vertex v1x v1y v1z

[0050] vertex v2x v2y v2z

[0051] vertex v3x v3y v3z

[0052] endloop

[0053] endfacet

[0054] endsolid name

[0055] Such file is processed in such a way as to extract the information necessary to construct a 3D mesh, namely:

[0056] vertices: xyz coordinates of each vertex

[0057] matrix of size n × 3 (where n is the number of vertices), for example

[0058] [

[0059] [-49.73719, -51.801174, 164.13678 ],

[0060] [-56.78471, -51.800766, 166.51537 ],

[0061] [-49.52898, -48.392784, 164.51765 ],

[0062] ...,

[0063] ]

[0064] faces: each face is identified by the ids of the vertices that compose it, matrix of size (m × 3 wherem is the number of faces)

[0065] [

[0066] [ 0, 1, 2 ],

[0067] [ 1, 3, 4],

[0068] [ 3, 5, 6],

[0069] ...,

[0070] ]

[0071]

[0072] In accordance with one embodiment, geometric transformations (rotations and translations) are also applied to the coordinates of the vertices so that the geometric representations of the calipers, in this case the geometric meshes 100, all have the same orientation in space and, preferably, that the centre of gravity coincides with the origin of the xyz axes.

[0073] The two data structures (matrices) mentioned above are thus used to generate new local information (step b) ) related to the geometry of the calipers, for example:

[0074] normals at the vertices: xyz components of the normal vector calculated at each vertex, matrix of size n × 3 (where n is the number of vertices)

[0075] [

[0076] [-0.44225708, 0.4844753, -0.7547797 ],

[0077] [-0.17541838, 0.37867317, -0.9087546 ],

[0078] [-0.4241023, -0.35505542, -0.8331104 ],

[0079] ...,

[0080] ];

[0081] curvature: curvature calculated locally at the vertex, array of size n (where n is the number of vertices )

[0082] [0.01295676, 0.01418231, 0.01847907,..., 0.0189136, 0.01575623, 0.02365503].

[0083] For the subsequent step c) of training, in addition to the input data obtained in steps a) and b), for example the vertex matrix, face matrix, normals and curvature, the output data 200 is also required.Said output data 200, relating to physical features of the brake caliper, comprises at least one or all of the following data:- a field of stresses due to the pressure of a brake fluid on the brake caliper and to the friction force generated between pad and brake disc, said field of stresses being obtained for each brake caliper from finite element simulations and comprising a value of the stresses for each vertex of the geometric mesh;- integral of the elastic deformation energy of the brake caliper.

[0084] Preferably, said output data 200 relating to physical features of the brake caliper is, for example, previously calculated by means of a finite element method(FEM) simulation, for example using a software such as Abaqus.

[0085] Preferably, said output data 200 comprises at least one output datum for each of the vertices and / or faces of the geometric mesh 100.

[0086] In accordance with one embodiment, since it may occur that the geometric mesh of a brake caliper does not coincide with the mesh of the results of the FEM stress field simulation, the computer-implemented method for training an artificial intelligence model also provides for a step of interpolating the stress field from the FEM simulation so as to have a stress value and / or a value of the friction force generated between pad and brake disc assigned for each vertex of the geometric mesh 100.

[0087] In accordance with one embodiment, the step of training the artificial intelligence model 1 provides for:

[0088] - dividing the available geometric meshes 100 of brake calipers into a training set and a testing set 300. Subsequently, the training of the artificial intelligence model 1 is carried out using the training set and the artificial intelligence model 1 is validated (step d) on the testing set.

[0089] In accordance with one embodiment, the artificial intelligence model 1 has a point-based architecture 2. Anexample of a point-based architecture is shown in figure 3, for example a pointNET architecture.

[0090] In accordance with one embodiment, the artificial intelligence model 1 has a mesh and graph-based architecture. Such architecture is particularly suitable for classification and segmentation problems of 3D objects.

[0091] In accordance with one embodiment, the artificial intelligence model 1 has a point-based or mesh and graphbased architecture adapted for regression problems, by modifying the final layers of the neural network and adapting the loss function.

[0092] In particular, the artificial intelligence model is, for example, a DiffusionNet model 3, which is a model specifically suitable for learning on surfaces represented as triangular meshes.

[0093] In accordance with one embodiment, for example as shown in figure 4, said DiffusionNet model 3 also comprises a diffusion layer 31 which models, through the heat equation, the propagation of information across the surface of the object, allowing both local analysis and long-range communication via said diffusion layer. For this reason, this model is particularly suitable for predicting both local and global properties of the surface.

[0094] In accordance with one embodiment, the DiffusionNet model provides for the following parameters:

[0095] Epoch: 120

[0096] Learning rate: 1e-5

[0097] Loss function: MSE for the energy, Loss custom to penalise the error on the extreme values of the stress in the case of stresses;

[0098] Total samples: 206

[0099] Test samples: 164 for training and 42 for test

[0100] Output parameters:

[0101] - energy, a scalar value;

[0102] - stress, as a tensor of size n × 1 where n is the number of vertices of the mesh.

[0103] In accordance with one embodiment, the computer-implemented method for training an artificial intelligence model further comprises step bl) of receiving one or more global information items to be added to the input data for training, relating to global geometric properties of the brake caliper.

[0104] An example of geometric calculation of some global parameters is shown in figure 5.

[0105] In accordance with one embodiment, the global information may be selected from one or more, or all, of the following information:

[0106] - effective radius of the caliper;

[0107] - number of pistons,

[0108] - total area of pistons,

[0109] - caliper fixing type (axial or radial),

[0110] - outer diameter of the disc,

[0111] - overall thickness of the disc,

[0112] - presence or absence of the tie rod,

[0113] - pad support type (flat or pins),

[0114] - centre-to-centre distance 102: distance of the centre between the holes of the two screws to constress the caliper to the hub carrier,

[0115] - height 103: z-coordinate of the axes of the screw holes,

[0116] - maximum radius 101: maximum radius considering the coordinates of the points of the caliper in a cylindrical system centred at the origin,

[0117] - theta 104: angular span of the points considering the coordinates of the points of the caliper in a cylindrical system centred at the origin, calculated on a planar view of the geometric mesh of the caliper.

[0118] For example, for each caliper it is possible to encode a dictionary object of the global information, for example as follows, in Python code:

[0119] {

[0120] " Caliper code": 4.0,

[0121] " Project code": " XC2A2 ",

[0122] " Effective radius [mm] ": 158. 5,

[0123] " Pistons number": 6.0,

[0124] " Pistons area [mm^2] ": 328. 1,

[0125] " Fixing type": " Radial",

[0126] " Disk diameter [mm] ": 390.0,

[0127] " Disk thickness [mm] ": 36.0,

[0128] " Tie rod flag": 1,

[0129] " Abutment type": "flat",

[0130] " Interaxis distance [mm] ": 256.0,

[0131] " Height [mm] ": 102.0,

[0132] " Maximum radius [mm] ": 227. 217,

[0133] " Theta [deg] ": 119. 665

[0134] }.

[0135] Once the model 1 has been trained, it is then possible to use it to predict physical features of a brake caliper from an input geometric mesh thereof, as shown for example in the diagram of figure 2.

[0136] With reference for example to figure 2, starting from an input geometric mesh, the computer-implemented method for predicting physical features of a brake caliper comprises performing the following steps on one or more electronic processors:

[0137] - receiving or querying a trained artificial intelligence model 1 as obtained by the computer-implemented method for training an artificialintelligence model according to any of the method variants described in the present document;

[0138] - predicting the physical features of the brake caliper using the trained artificial intelligence model, said predicting comprising:

[0139] i) using the input geometric representation, for example a geometric mesh 100 of a brake caliper, as input data for the trained artificial intelligence model 1;

[0140] ii) providing as output data from the trained artificial intelligence model a prediction of at least one physical feature of the brake caliper, for example providing at least one value related to the field of stresses due to the pressure of a brake fluid on the brake caliper and / or to the friction force generated between pad and brake disc for the input geometric representation, for example for each vertex and / or face of the input geometric mesh 100 of a brake caliper.

[0141] In accordance with one embodiment, step i) also comprises a data processing step iO) in which the input geometric mesh 100 is processed to extract vertices and faces and preferably to calculate local curvature and normals at each vertex to be provided as input to the trained artificial intelligence model 1.

[0142] It is clear that, preferably, the computer-implemented method for predicting physical features of a brake caliper does not include any step of performing a finite element analysis (FEM) method to generate a field of stresses or the energy integral. Therefore, preferably, this method uses only output data relating to the field of stresses or to the energy integral generated solely by the trained artificial intelligence model 1.

[0143] It is clear that the present invention also concerns a system for training an artificial intelligence or machine learning model based on neural networks intended to predict physical features of a brake caliper from its geometric mesh. Likewise, the present invention also concerns a system for predicting physical features of a brake caliper by means of a trained artificial intelligence model 1.

[0144] The system 400 for training an artificial intelligence or machine learning model based on neural networks intended to predict physical features of a brake caliper from its geometric mesh comprises an input interface 20 for accessing:

[0145] - training data comprising a plurality of training instances representing different geometric meshes of brake calipers, each training instance comprising:

[0146] - a geometric mesh of a respective brakecaliper, comprising vertices and faces;

[0147] - local information relating to the geometry of each brake caliper;

[0148] - data defining an artificial intelligence model configured to receive a respective geometric mesh as input and to generate an output of the artificial intelligence model.

[0149] The system 400 for training an artificial intelligence model further comprises a processor module 30 configured to train the artificial intelligence model 1 using the training data to obtain a trained artificial intelligence model 1 intended to predict the physical properties of an input geometric mesh of a brake caliper provided as input data to the artificial intelligence model 1, said training being performed with input data comprising the local information received and output data relating to physical features of the brake caliper comprising at least one of the following:- a field of stresses due to the pressure of a brake fluid on the brake caliper and to the friction force generated between pad and brake disc, said field of stresses being obtained for each brake caliper from finite element simulations and comprising a value of the stresses for each vertex of the geometric mesh;- integral of the elastic deformation energy of the brakecaliper.

[0150] The system for predicting physical features of a brake caliper 500 from its input geometric mesh comprises:

[0151] - a system input interface 50 for accessing data defining a trained artificial intelligence model as obtained by the computer-implemented method for training an artificial intelligence model according to any one of the embodiments described in the present document;

[0152] - a processor module 51 configured to predict the physical features of the brake caliper using the trained artificial intelligence model 1, said prediction comprising:

[0153] - using an input geometric representation, for example an input geometric mesh 100 of a brake caliper, as input data for the trained artificial intelligence model 1;

[0154] - providing as output data from the trained artificial intelligence model 1 a prediction of at least one physical feature of the brake caliper.

[0155] The physical feature provided as output data may be one or more of the following features:

[0156] - a field of stresses due to the pressure of a brake fluid on the brake caliper and to the friction force generated between pad and brake disc, said field ofstresses comprising a stress value assigned to each local region of the input geometric representation, for example a stress value assigned to each vertex of the geometric mesh;

[0157] - integral of the elastic deformation energy of the brake caliper.

[0158] It is clear that the operator is able, via the system input interface, to upload the input geometric representation, for example an input geometric mesh 100 of the brake caliper, to be sent as input to the trained artificial intelligence model 1 and obtain, for example on the same system input interface, the prediction of the physical features of the brake caliper.

[0159] Innovatively, the method according to the present invention effectively overcomes the drawbacks cited with reference to the prior art.

[0160] Advantageously, the application of machine learning techniques, in particular neural networks, for predicting the physical and mechanical properties of brake calipers, makes it possible to learn and generalise from the data of previous FEM simulations.

[0161] Once trained, the artificial intelligence model can accurately predict the desired physical properties based solely on the geometric mesh of the object, without requiring finite element simulations, thereby drasticallyreducing the time required for analysis and enabling faster design iterations.

[0162] As can be seen from the attached figure 7, the model allows accurate prediction of stresses, with average percentage errors below 10%.

[0163] Furthermore, as can be seen from the attached figure 8, the model allows accurate prediction of deformation energy, with average percentage errors below 7%.

[0164] By receiving the geometric meshes of the brake calipers, each comprising vertices and faces, the method enables a detailed and accurate representation of the geometry of the brake caliper, which is essential for precise predictions of the physical properties.

[0165] By receiving local information relating to the geometry of each brake caliper, such as thicknesses, angles and curvatures, the model ' s ability to capture fine details and variations in the structure of the caliper is enhanced, leading to more accurate predictions of the physical properties.

[0166] Furthermore, by training an artificial intelligence model based on neural networks to predict the physical properties of the geometric mesh of a brake caliper using input data that includes local geometric information and output data from finite elementsimulations, it is ensured that the model can learn from high-fidelity simulation data. This approach significantly reduces the computational resources and time required compared to running full finite element simulations for every new design iteration.

[0167] By predicting the fields of stresses due to the pressure of the brake fluid and the integral of the elastic deformation energy of the brake caliper, a comprehensive analysis of the performance of the caliper under operating conditions is enabled. This capability allows rapid design iterations and optimisations, improving the overall efficiency of the brake caliper design process.

[0168] Furthermore, advantageously, by receiving global information relating to global geometric properties of the brake caliper, the method allows the inclusion of additional data that may influence the physical properties of the brake caliper.

[0169] In particular, by including this global information in the input data during the training phase, the artificial intelligence model can learn more complex relationships between the global geometric properties and the physical features of the brake caliper. This leads to more accurate and reliable predictions of the physical properties of the brake caliper, further improving theefficiency of the design process.

[0170] The ability to receive and use global information allows the method to adapt to a variety of brake caliper configurations, making it versatile and applicable to different design scenarios. This approach enhances the flexibility of the artificial intelligence model and its ability to generalise to new brake caliper configurations not seen during training.

[0171] The computer-implemented method that uses the artificial intelligence model trained according to the present invention enables accurate and reliable predictions of the physical properties of the brake caliper based solely on the input geometric mesh. In contrast to prior art solutions, which still require a subsequent FEM analysis to generate the physical properties, this approach enables fast and precise performance analyses of the brake caliper under operating conditions, improving the overall efficiency of the design process. Moreover, it allows faster design iterations, optimising the time and resources needed for the development of new brake caliper configurations.

[0172] It is clear that a person skilled in the art, in order to meet contingent needs, may make modifications to the invention described above, all of which fall within the scope of protection as defined by thefollowing claims.

Claims

CLAIMS1. A computer-implemented method for training an artificial intelligence or machine learning model (1) based on neural networks intended to predict physical features of a brake caliper from a geometric representation thereof, for example a geometric mesh (100), said method comprising performing the following steps on an electronic processor:a) receiving geometric representations, for example geometric meshes of brake calipers, each geometric representation (100) of said geometric representations comprising one or more local regions, for example vertices and faces;b) receiving local information relating to the geometry of each brake caliper;c) training an artificial intelligence model based on neural networks intended to predict the physical properties of an input geometric representation of brake caliper provided as input data to the artificial intelligence model, said step of training the artificial intelligence model being performed with input data comprising the local information received in step b) and output data (200) relating to physical features of the brake caliper comprising at least one of the following: - a field of stresses due to the pressure of a brakefluid on the brake caliper and to the friction force generated between pad and brake disc, said field of stresses being obtained for each brake caliper from finite element simulations and comprising a value of the stresses for one or more local regions of the input geometric representation, for example for each vertex of the input geometric mesh;- integral of the elastic deformation energy of the brake caliper.

2. Method according to claim 1, wherein the local information comprises thicknesses and / or angles and / or curvatures of local regions of the brake caliper.

3. Method according to claim 1, wherein the local information comprises at least the normals at the vertices and the local curvature calculated locally at the vertices.

4. Method according to claim 3, wherein the following steps are performed before step b) to determine the local information relating to the geometry of each brake caliper:al) applying geometric transformations to the coordinates of the vertices to make uniform the orientation of the brake calipers in space and align the center of gravity of the brake calipers, for example aligning it with the origin of the axes of the three-dimensional virtual space(xyz);a2) calculating local information relating to the geometry of each brake caliper, said local information comprising at least normals at the vertices and locally calculated local curvature at the vertices.

5. Method according to any one of the preceding claims, further comprising the step bl) of receiving one or more of the following pieces of global information, relating to global geometric properties:- effective radius of the caliper; number of pistons, total area of pistons, caliper fixing type (axial or radial), outer diameter of the disc, overall thickness of the disc, whether or not the tie rod is present, pad support type (flat or pins), center-to-center distance: distance of the center between the holes of the two screws to constrain the caliper to the hub carrier, height: z-coordinate of the axes of the screw holes, maximum radius: maximum radius considering the coordinates of the points of the caliper in a cylindrical system centered at the origin, theta: angular span of the points considering the coordinates of the points of the caliper in a cylindrical system centered at the origin;and wherein in step c) the input data comprise at least one or more of said pieces of global information.

6. Method according to any one of the preceding claims wherein the artificial intelligence model comprises a DiffusionNet neural network architecture.

7. Method according to any one of the preceding claims, wherein the field of stresses comprising a value of the stresses for one or more regions of the input geometric representation is obtained by means of interpolation techniques to align the output data of finite element simulations with the geometric representation of each brake caliper.

8. Method according to any one of claims 4 to 7, wherein the geometric transformations applied to the coordinates of the vertices comprise rotations and translations to make uniform the orientation of the brake calipers in space.

9. Method according to any one of the preceding claims, wherein the artificial intelligence model is trained using an adapted loss function for regression problems.

10. Method according to any one of the preceding claims, wherein the artificial intelligence model is trained using a neural network architecture comprising a diffusion layer which utilizes the heat equation to propagate information through the brake caliper surface.

11. A computer-implemented method for predicting physical features of a brake caliper from an input geometricrepresentation thereof, for example a geometric mesh (100), said method comprising performing the following steps on one or more electronic processors:- receiving or interrogating a trained artificial intelligence model (1) as obtained by the method according to any one of claims 1 to 10;- predicting the physical features of the brake caliper using the trained artificial intelligence model, said predicting comprising:i) using the input geometric representation as the input data for the trained artificial intelligence model (1); ii) providing as output data from the trained artificial intelligence model a prediction of at least one physical feature of the brake caliper, for example a field of stresses.

12. Computer-implemented method for predicting physical features of a brake caliper according to claim 11, said method not including any step of performing a finite element analysis method (FEM) for generating a field of stresses or an energy integral.

13. A system for training an artificial intelligence or machine learning model based on neural networks intended to predict physical features of a brake caliper from a geometric representation thereof, for example a geometric mesh, said system comprising:- an input interface to access:training data comprising a plurality of different representative training instances for input geometric representations of brake calipers, each training instance comprising:a geometric representation of a respective brake caliper;local information relating to the geometry of each brake caliper;data defining an artificial intelligence model configured to receive a respective geometric representation as input and generate an output of the artificial intelligence model;- a processor module configured to train the artificial intelligence model using the training data to obtain a trained artificial intelligence model intended to predict the physical properties of an input geometric representation of brake caliper provided as input data to the artificial intelligence model, said training being performed with input data comprising the received local information and output data relating to physical features of the brake caliper comprising at least one of the following:- a field of stresses due to the pressure of a brake fluid on the brake caliper and to the friction forcegenerated between pad and brake disc, said field of stresses being obtained for each brake caliper from finite element simulations and comprising a value of the stresses for one or more regions of the input geometric representation, for example for each vertex of the input geometric mesh;- integral of the elastic deformation energy of the brake caliper.

14. A system for predicting physical features of a brake caliper from an input geometric representation thereof, said system comprising:- an input interface for accessing data defining a trained artificial intelligence model as obtained by the method according to any one of claims from 1 to 10;- a processor module configured to predict the physical features of the brake caliper using the trained artificial intelligence model, said prediction comprising:i) using the input geometric representation as input data for the trained artificial intelligence model;ii) providing as output data from the trained artificial intelligence model a prediction of at least one physical feature of the brake caliper.

Images