Method for producing a technical object via an implicit and continuous digital simulation of a multi-scale complex physical phenomenon by deep learning

Abstract

The invention relates to a method for producing a technical object including a geometry comprising a plurality of surface points Ai forming a surface mesh, the geometry being configured to be subjected to at least one physical phenomenon, such as a flow of a fluid, a heat transfer, an electromagnetic field or a structural mechanical stress, comprising a phase of predicting physical values by a neural network. The neural network comprises a so-called surface encoder of the geometry and of at least one simulation condition in a representative vector, and a so-called volume decoder of the physical values over all or part of the points of the simulation space on the basis of the representative vector and of the parameters of the neural network obtained by deep learning, the physical values being correlated with one another.

Description

Process for producing a technical object via an implicit and continuous digital simulation of a complex multi-scale physical phenomenon by deep learning TECHNICAL FIELD OF THE INVENTION [1] The field of invention is that of digital simulation. [2] More specifically, the invention relates to a method for producing a technical object via an implicit and continuous digital simulation of a complex multi-scale physical phenomenon by automatic and deep learning. [3] The invention finds applications in particular in the context of an optimization of a shape of the technical object, by simulating a complex physical phenomenon of the fluidic type (e.g.: flow of a fluid, two-phase flow, particle tracking, etc.), of the structural type (e.g.: resistance of a structure) and / or of the thermal type (e.g.: conduction, convection, radiation). A particular application of the invention is multiphysics simulation, combining several types of phenomena, such as for example natural or forced convection around an object, etc. STATE OF THE ART [4] Techniques for simulating a physical phenomenon by discretizing a space by a continuous mesh of elements in which representative differential equations (for example partial differential equations) of the physical phenomenon are calculated are known from the prior art. The simulation is obtained when the calculation converges towards a result respecting predetermined boundary conditions as well as continuity of the physical data between two contiguous elements discretizing the space. This type of simulation is generally called CFD (acronym for "Computational Fluid Dynamics") when simulating a fluid flow, or FEA (acronym for "Finite Element Analysis") in the context of a calculation on a mechanical structure, for example of the strength of materials type. [5] The major drawback of these earlier techniques is that they are generally very computationally intensive, depending on the size of the mesh used to discretize the simulation space. For example, it is common to obtain a consistent simulation result after hundreds or even thousands of CPU (Computer Processor Unit) hours in parallel, making numerical simulation expensive in terms of calculation time and therefore energy consumption. Optimizing a parameter to be simulated, such as a dimension of a geometry, is then difficult to envisage without drastically limiting the number of simulations and therefore the quality of the optimization. [6] In order to improve the efficiency of simulation calculation time, numerical simulation techniques using automatic learning, also known as "Machine Learning", have been developed, implementing deep learning techniques, better known as "Deep Learning", in order to obtain a result by prediction compared to results previously calculated on similar problems. [7] For example, known from the prior art is the numerical simulation technique described in the article by Li et al. which was published in 2019 and which relates to a deep learning variational solver based on a domain decomposition method, called the deep domain decomposition method D3M (acronym for "deep decomposition method") for implementing parallel calculations in physical sub-domains of the simulation. [8] The major drawback of this technique is that it is based on a reconstruction of the simulation from the result of the simulation in each sub-domain. However, the intersection between each sub-domain poses significant difficulties in the simulation because the boundary conditions are only updated after reconstruction, thus resulting in an overconsumption of computing time. [9] In order to improve the efficiency of calculation and consequently the consumption of computing and energy resources, the Applicant has notably filed in this field, French patent application No. 1914967 which relates to a method of digital simulation by automatic learning, based on local predictions carried out sequentially from automatic learning of global simulations carried out previously.

[0010] However, there is a need to further improve the computational efficiency during the numerical simulation of a complex physical phenomenon in order to limit the use of computing resources to obtain a result from a numerical simulation of a physical phenomenon. The present invention falls within this general framework. The optimization of computing and energy resources is a challenge important in the design of a complex three-dimensional part subject to a physical phenomenon. This optimization can in particular be part of an optimization of the shape of the part subject to a physical phenomenon. This shape optimization can in particular be carried out with a view to obtaining better aerodynamics. STATEMENT OF THE INVENTION

[0011] To this end, the invention relates to a method for producing a technical object comprising a geometry comprising a plurality of surface points A tforming a surface mesh, said geometry being configured to be subjected to at least one physical phenomenon, such as a flow of a fluid, a heat transfer, an electromagnetic field or a structural mechanical stress.

[0012] According to the invention, the method comprises: - a phase, implemented by computer, of collecting a plurality of physical data sets in a so-called training database, the physical data being derived from measurements and / or digital simulations of a similar physical phenomenon carried out at least in the vicinity of a so-called training surface, said training surface being able to be linked or not to the geometry, and being distinct for each data set; - a phase, implemented by computer, of generating parameters of a neural network by deep learning on at least part of the data in the training database; - a phase, implemented by computer, of prediction by the neural network of physical values ​​for all or part of the points of a cloud of points of a simulation space configured at least in the vicinity of said geometry, as a function of at least one so-called simulation condition, the physical values ​​forming the simulation result; and - a phase of production of said object in which at least one dimension of the geometry of the object is established based on said simulation result,

[0013] In addition, the neural network comprises a so-called surface encoder of the geometry and at least one simulation condition in a representative vector and a so-called volume decoder of the physical values ​​on all or part of the points of the simulation space from the representative vector and the parameters of the network. of neurons obtained by deep learning, the physical values ​​being correlated with each other.

[0014] Thus, the neural network, as configured, makes it possible to obtain, compared to existing methods, a more coherent simulation result in a shorter time. Indeed, each node of the neural network contains coherent volume information, derived from measured or simulated physical data, making it possible to arrive at a simulation result in the 3D simulation space in very few iterations. The energy gain linked to obtaining a simulation result is thus notable compared to a conventional method, and is very useful, for example, in the context of optimizing a dimension of the geometry with regard to the constraint of the physical phenomenon(s).

[0015] Furthermore, thanks to the surface encoder and volume decoder configuration, the physical data can be continuously propagated through the neural network which is then adapted to perform calculations on point clouds inhomogeneous in distance or resolution.

[0016] It should be emphasized that a neural network configured to perform a numerical simulation classically works on a point cloud homogeneous in size, linked to a regular grid induced by the convolution support which is generally square.

[0017] In the present invention, since the point cloud can be inhomogeneous thanks to the configuration of the encoder and the decoder, the calculations performed are carried out much more quickly at equal resolution, offering the possibility of greatly refining certain delicate points of the simulation space, generally with a high gradient, such as for example the flow zone around an angle or a cavity. On the other hand, the low gradient zones do not need with the present configuration of the neural network to be very dense in calculation points, which is not the case with a conventional neural network where the resolution is generally homogeneous at all points of the simulation space.

[0018] In this description, a point cloud is understood to mean a set of points distributed in the simulation space and a mesh is understood to mean a point cloud whose connections are known.

[0019] In particular embodiments of the invention, the surface encoder comprises a plurality of sub-blocks, each sub-block processing the surface mesh according to a distinct sampling.

[0020] In particular implementations of the invention, the volume decoder comprises a plurality of sub-blocks, each sub-block processing the point cloud of the simulation space according to a distinct sampling.

[0021] In particular implementations of the invention, at least one so-called intra-scale residual connection is configured between a sub-block of the surface encoder comprising a surface mesh having a given sampling and a sub-block of the volume decoder comprising a point cloud having a similar sampling.

[0022] Thus, the loss of physical information is minimized during the different resamplings.

[0023] In particular embodiments of the invention, at least one so-called inter-scale residual connection is configured between two sub-blocks of the surface encoder and / or between two sub-blocks of the volume decoder, said two sub-blocks having distinct sampling.

[0024] Thus, the loss of physical information is minimized during encoding or decoding.

[0025] In particular embodiments of the invention, the decoder comprises two sub-blocks, one being dedicated to the calculation of the physical values ​​in the volume of the simulation space and the other being dedicated to the calculation of the physical values ​​on the surface of the geometry, the calculation being carried out jointly in the two sub-blocks.

[0026] In particular embodiments of the invention, the training database comprises a plurality of numerical simulation results, each result being associated with a training data set comprising: - a three-dimensional mesh of a surface representative of a geometry; - a set of points distributed in the simulation space; and - at least one boundary condition value.

[0027] In particular embodiments of the invention, when at least part of the data sets comprises physical data on the surface and in the volume, the deep learning is performed jointly on the physical data on the surface and in the volume.

[0028] In particular embodiments of the invention, during the prediction phase, an interpretation is carried out from the simulation values in all or part of the points of the cloud, to know the simulation values ​​at a distinct point of the cloud.

[0029] In particular embodiments of the invention, the physical data and / or physical values ​​are of types included among: - volume data; - surface data; - an overall coefficient; - an evolutionary curve in time of a volume data; - a time-evolving curve of a surface data; - an evolutionary curve in time of a global coefficient.

[0030] In particular embodiments of the invention, physical values ​​of different types are spatially and / or temporally correlated.

[0031] The invention also relates to a device comprising an element for storing the instructions for implementing the production method according to any of the preceding implementation modes.

[0032] Finally, the invention also relates to a neural computing device comprising: - a deep learning unit of physical data from measurements and / or numerical simulations of at least one physical phenomenon, such as a flow of a fluid, a heat transfer, an electromagnetic field or a structural mechanical constraint, carried out at least in the vicinity of a so-called training surface, said training surface being able to be distinct for each data set, the learning unit generating weights of a neural network; - a unit for predicting physical values ​​for all or part of the points of a point cloud of a simulation space configured at least in the vicinity of a geometry comprising a plurality of surface points A tforming a surface mesh, said geometry being capable of being subjected to at least one physical phenomenon, such as a flow of a fluid, a heat transfer, an electromagnetic field or a structural mechanical constraint, the prediction being carried out as a function of at least one so-called simulation condition, the physical values ​​forming a simulation result. BRIEF DESCRIPTION OF THE FIGURES

[0033] Other advantages, aims and particular characteristics of the present invention will emerge from the following non-limiting description of at least one particular embodiment of the devices and methods which are the subject of the present invention, with reference to the appended drawings, in which: - figure 1 is a block diagram illustrating a mode of implementation of the production method according to the invention - figure 2 is a schematic view of a simulation space comprising an example of a technical object subjected to a physical phenomenon to be simulated during the production method according to figure 1; - figure 3 is a schematic view of the organization of the sampling carried out by the neural network implemented by the production method according to figure 1; - Figure 4 is a schematic view of a neural computing device according to the invention. DETAILED DESCRIPTION OF THE INVENTION

[0034] This description is given without limitation, each characteristic of an embodiment being able to be combined with any other characteristic of any other embodiment in an advantageous manner.

[0035] Please note, from now on, that the figures are not to scale. Example of a particular implementation mode

[0036] Figure 1 illustrates a non-limiting example of an implementation mode of a method 100 for producing a technical object 200 such as for example that illustrated in Figure 2. The method 100 is implemented by a computer which comprises a unit for storing computer instructions for implementing the method 100. The instructions make it possible in particular to configure a processor of the computer for implementing the method 100.

[0037] The technical object 200 is here a hull of a boat subjected to at least one physical phenomenon, in this case to the flow of a fluid 205, namely water, around the hull 200. The simulated physical phenomenon(s) are generally governed by one or more partial differential equation(s) (PDE). The hull 200 comprises a geometry 210 comprising a plurality of surface points A t . Surface points A t . form a three-dimensional mesh representing the surface of the geometry 210. A simulation space 220, associated with the geometry 210, is determined around the geometry 210 in order to simulate the flow of the fluid around the shell 200.

[0038] It should be emphasized that depending on the configurations to be simulated, the simulation space 220 may be around, inside, through, or next to the geometry 210. Generally, the simulation space 220 is at least near the geometry 210 and may include points on the surface of the geometry 210.

[0039] The implementation method 100 comprises a first phase 110 of collecting a plurality of physical data sets in a training database.

[0040] The physical data sets come for example from measurements or a digital simulation, for example of the CFD type, carried out previously in a space at least close to one or more so-called training surfaces, similar or not to a surface presented by the geometry 210.

[0041] During the second phase 120, the implementation method 100 generates parameters of a neural network by deep learning on at least part of the data collected during the phase 110. This second phase 120 corresponds to a training phase to configure the neural network to make consistent predictions.

[0042] Once trained, the neural network predicts physical values ​​for all or part of the points of a point cloud included in the simulation space 220 during a third phase 130 of the method 100. It should be emphasized that the points of the cloud are generally distributed in a non-homogeneous manner in the simulation space 220, with a higher density in potentially high gradient areas. The physical values ​​thus obtained form a simulation result which is used in a fourth phase 140 of the method 100.

[0043] During this fourth phase 140, a production of the object 200, a dimension of which is established according to the simulation result, is carried out. The prediction of the physical values ​​carried out during the phase 130 makes it possible to refine the dimensions of the object 200, in particular with the aim of optimizing this dimension with regard to the simulated physical phenomenon(s). This optimization may in particular correspond to an optimization of the shape to reduce drag as much as possible, or even optimize the resistance / mass ratio of the object. many applications can be cited, particularly in the field of mechanics.

[0044] The core of the invention relates to the way in which the prediction of physical values ​​is carried out, which is in some way a lightweight numerical simulation compared to a classic numerical simulation of the CFD type, in particular with the aim of limiting the computing resources used to obtain a reliable result.

[0045] Compared to the state of the art, the neural network used for prediction advantageously includes a surface encoder and a volume decoder.

[0046] Each point of the point cloud of the simulation space 220 thus advantageously comprises data that is consistent with the points of the cloud located in the immediate vicinity. In other words, it is possible to propagate information about the physical phenomenon within the point cloud of the simulation space 220 in a very short period of computing time, while maintaining the continuity inherent in physical phenomena.

[0047] The weights of the neural network nodes thus include an implicit representation of the three-dimensional flow field, allowing this extremely rapid propagation of information in the mesh which is decorrelated from the neural network nodes.

[0048] It is worth noting that the neural network is usually trained on several simulations and / or measurements. Once trained, the neural network is used to infer in a few seconds physical quantities at different scales on a batch of data, i.e. the geometry and boundary conditions pair, which has never been seen before during training, without the need to readjust any parameters of the neural network.

[0049] Advantageously, the neural network decoder can be configured in two separate blocks, one being dedicated to the calculation in the volume of the simulation space, i.e. in the point cloud, and the other in the surface mesh, which allows to better specify these blocks to the physical configurations of the geometry and the space where the fluid flows, both reacting in a distinct way. A saving in computation time is also obtained by this advantageous configuration.

[0050] In order to generate a simulation result, the neural network was previously trained during training phase 120, on the physical data collected previously. Samples of data from traditional CFD techniques can thus be taken into account during training. to have a very good functional approximation of the underlying functional solution of the PDE.

[0051] The training is advantageously carried out with joint learning of the surface results of the geometry 210 and the volume results in the simulation space 220.

[0052] It should be noted that existing PINN (Physics-Informed Neural Networks) neural networks do not use a common training base for multiple simulations but must be retrained for each simulation. The neural network of the present invention, on the contrary, allows training to be shared for several simulations, allowing a significant time saving compared to PINNs.

[0053] As illustrated in Figure 3, the neural network comprises a plurality of sub-blocks 310 each processing meshes of the dataset with a distinct resampling in order to allow for a unified model that regresses at different scales (volumes, surfaces, global) to better propagate the information of the underlying physics within the points of the mesh.

[0054] The surface meshes 320, illustrated on the left of Figure 3, are sampled at different scales by the encoder 360, starting here from a matrix with 16k rows and 5 columns (including physical data such as speed in the three directions, pressure, temperature) until a representative vector 330 of dimension d6 is obtained. At each resampling, the 16k rows are divided by a predetermined number, for example equal to 2. This resampling allows a better understanding of the geometry 210. In other words, an encoding of the surface data is carried out in the representative vector 330.

[0055] During encoding, residual connections 335 between surface meshes having distinct samplings, i.e. between two sub-blocks 310, are made to better preserve the physical information.

[0056] The representative vector 330 is used to decode the physical values ​​on the volume points of the similarly resampled point clouds 340, by the decoder 370 illustrated on the right side of Figure 3.

[0057] Residual connections 345 between point clouds 340 having distinct samplings are also performed to improve the preservation of physical information during decoding.

[0058] It should be emphasized that the residual connections 335 and 345 may be between two consecutive sub-blocks, as illustrated in Figure 3, and / or between two non-consecutive sub-blocks.

[0059] Still with the aim of avoiding losing physical information in the encoding / decoding phases, a residual connection 350 is made between each surface mesh 320 and each point cloud 340 having the same sampling.

[0060] For example, for the highest sampling, the residual connection 350 is performed between the matrix of 16k rows and 5 columns of the surface mesh 320 and a point cloud 340 of 32k rows and 1 columns.

[0061] Thus, thanks to the residual connections 335 and 345, which are of the interscale type, and to the residual connections 350, which are of the intra-scale type, the physical information is well propagated between the surface mesh of the geometry 210 and the point cloud of the simulation space 220.

[0062] It should be emphasized that for the purpose of readability of Figure 3, the numbers of the residual connections 335, 345 and 350 have only been placed on a part of the connections drawn, as well as for the numbers 320 and 340 of the surface meshes and the volume point clouds. Each sub-block 310 here comprises an intra-scale residual connection 350 between the surface mesh 320 of said sub-block and the volume point cloud 340 of said sub-block. A residual connection 335 between the surface meshes 320 of two consecutive sub-blocks 310 is established. The same applies to the residual connections 345 between the point clouds 340.

[0063] This particular configuration of the neural network, comprising a surface encoder 360 and a volume decoder 370, correlated for each sampling level by residual connections 350 makes it possible to obtain a neural network capable of processing any type of point cloud and especially non-uniform point clouds, that is to say having a densification of points around points of interest such as, for example, salient angles of a geometry subjected to a flow.

[0064] A significant gain in computing time is thus obtained compared to conventional neural networks whose distribution of points is necessarily uniform because of their architecture based on a convolution support having a square shape. In other words, the neural network of the present invention allows to process simulations with a large number of parameters because the point cloud used can be densified only in areas with significant physical variations, and rarefied in areas of space with, on the contrary, low variations. This point cloud configuration is common in the field of classical CFD calculations but has never until now been used for neural networks because of their natural architecture.

[0065] Generally, simulation values ​​are of types included among: - volume data; - surface data; - an overall coefficient; - an evolutionary curve in time of a volume data; - a time-evolving curve of a surface data; - an evolutionary curve in time of a global coefficient.

[0066] In summary, traditional CFD methods aim to solve PDE equations at specific points using a particular mesh that discretizes the 2D / 3D space. The process of space meshing can be tedious, cumbersome, and time-consuming, as the resulting mesh must, among other things, adapt to the boundary conditions of the simulation.

[0067] In the present invention, an approximation of the underlying functional solution of the PDE is performed. Instead of requiring the neural network to output a fixed grid of physical variables, these variables are implicitly represented in the network weights. Thus, the network can be queried at any point in space. Therefore, this continuous network is able to learn a distribution of functions from 3D CFD data for which the points and the corresponding physical quantities are considered as samples of these functions. Furthermore, through backpropagation of the output through the network, it is possible to extract the different exact operators allowing the possibility of minimizing a physical residual loss.

[0068] The neural network thus configured makes it possible to obtain a simulation result in a short time, never achieved before, opening the way to better optimization of the technical object subject to physical constraints.

[0069] Figure 4 is a schematic view of a neural computing device 400 implementing steps 120 and 130 of the embodiment method 100.

[0070] The neural device 400 comprises for this purpose a unit 410 for deep learning of physical data from measurements and / or digital simulations of at least one physical phenomenon and a unit 420 for predicting physical values ​​for all or from the points of the point cloud of the simulation space 220, the physical values ​​forming the simulation result. Comparative data

[0071] The table, referenced [Table 1], makes it possible to compare the electrical consumption induced by a digital simulation process according to the invention, noted NN, in comparison with a digital simulation carried out by a classic CFD type technique.

[0072] [Table 1]

[0073] As can be seen, the digital simulation process implemented during the invention consumes very little electrical energy compared to a conventional CFD type calculation. Certainly, training the neural network has an energy cost that is not present for CFD, but this is largely offset by the fact that the energy cost for a simulation is extremely low.

[0074] Furthermore, the energy cost due to training is a fixed cost that can be shared with several simulations, which is particularly the case when seeking to optimize a shape of a geometry under defined boundary conditions (for example, given speed of the hull relative to the fluid, possibly a function of swell or current, etc.). The gain in terms of energy consumption being spectacular, the shape of the object produced can thus be better optimized by using a simulation process implementing a network of neurons as described in the example above, compared to a CFD type simulation.

[0075] Compared to the neural networks currently configured to perform numerical simulations, and in particular that described in the Applicant's international patent application, published under number WO 2022 / 229517, the present configuration of neural networks is capable of processing much more data with the same power consumption. In other words, the present invention makes it possible to be trained, with the same GPU (acronym for "Graphics Processor Unit"), on problems with more than 50 million mesh points, whereas the neural network of the previous patent application is limited to approximately 20 million points. This x2.5 factor in the number of points makes it possible to obtain a much finer resolution, with a robustness to tessellation of the order of 0.5%), thus significantly increasing the quality of the simulation results and consequently the optimization of the shape of the geometry under the constraint of the physical phenomenon.

[0076] The tables, referenced [Table 2] and [Table 3], illustrate the significant gain in terms of calculation error of the neural network described in the example above of an implementation mode according to the invention, referenced continuous model, compared to previous neural networks of the CNN type (English acronym for “Convolutional Neural Network”) or of the GAT type (English acronym for “Graph Attention Network”).

[0077] [Table 2] illustrates the absolute error on volume variables, while [Table 3] illustrates the absolute error on surface variables.

[0078] [Table 2]

[0079] [Table 3]

[0080] In [Table 3], WSS stands for Wall Shear Stress. ISP stands for Iso Static Pressure.

[0081] The data presented in Tables [Table 1], [Table 2] and [Table 3] are from simulations of a boat hull. The values ​​on the neural networks were averaged over 80 training simulations, 10 validation simulations and 10 test simulations.

[0082] It should also be noted that the data in [Table 3] correspond to a geometry + boundary conditions pair which does not correspond to any of the pairs used as input data when training the neural network.

Claims

Claims 1. Method for producing a technical object comprising a geometry comprising a plurality of surface points A t forming a surface mesh, said geometry being configured to be subjected to at least one physical phenomenon, such as a flow of a fluid, a heat transfer, an electromagnetic field or a structural mechanical stress, characterized in that said method comprises: • a phase, implemented by computer, of collecting a plurality of physical data sets in a so-called training database, the physical data being derived from measurements and / or digital simulations of a similar physical phenomenon carried out at least in the vicinity of a so-called training surface, said training surface being able to be linked or not to the geometry, and being distinct for each data set; • a phase, implemented by computer, of generating parameters of a neural network by deep learning on at least part of the data in the training database; • a phase, implemented by computer, of prediction by the neural network of physical values ​​for all or part of the points of a cloud of points of a simulation space configured at least in the vicinity of said geometry, as a function of at least one so-called simulation condition, the physical values ​​forming the simulation result; and • a phase of producing said object, at least one dimension of the geometry of the object being established as a function of said simulation result, in which the neural network comprises a so-called surface encoder of the geometry and at least one simulation condition in a representative vector and a so-called volume decoder of the physical values ​​on all or part of the points of the simulation space from the representative vector and the parameters of the neural network obtained by deep learning, the physical values ​​being correlated with each other.

2. Production method according to claim 1, in which the surface encoder comprises a plurality of sub-blocks, each sub-block processing the surface mesh according to a distinct sampling and the volume decoder comprises a plurality of sub-blocks, each sub-block processing the point cloud of the simulation space according to a distinct sampling.

3. Production method according to claim 2, in which at least one so-called intra-scale residual connection is configured between a sub-block of the surface encoder comprising a surface mesh having a given sampling and a sub-block of the volume decoder comprising a point cloud having a similar sampling.

4. Production method according to any one of claims 2 to 3, in which at least one so-called inter-scale residual connection is configured between two sub-blocks of the surface encoder and / or between two sub-blocks of the volume decoder, said two sub-blocks having distinct sampling.

5. Production method according to any one of claims 1 to 4, also comprising a block for interpreting at least one physical value on a point not included in the point cloud of the simulation space.

6. Production method according to any one of claims 1 to 5, in which the decoder comprises two sub-blocks, one being dedicated to the calculation of the physical values ​​in the volume of the simulation space and the other being dedicated to the calculation of the physical values ​​on the surface of the geometry, the calculation being carried out jointly in the two sub-blocks.

7. Production method according to any one of claims 1 to 6, in which the training database comprises a plurality of digital simulation results, each result being associated with a training data set comprising: • a three-dimensional mesh of a surface representative of a geometry; • a set of points distributed in the simulation space; and • at least one boundary condition value.

8. Production method according to any one of claims 1 to 7, wherein, when at least part of the data sets comprises physical data on the surface and in the volume, the deep learning is carried out jointly on the physical data on the surface and in the volume.

9. Production method according to any one of claims 1 to 8, in which, during the prediction phase, an interpretation is carried out from the simulation values ​​at all or part of the points of the cloud, to know the simulation values ​​at a distinct point of the cloud.

10. Production method according to any one of claims 1 to 9, in which the physical data and / or the physical values ​​are of types included among: • volume data; • surface data; • an overall coefficient; • an evolutionary curve in time of a volume data; • an evolutionary curve in time of a surface data; • an evolutionary curve in time of a global coefficient.

11. Production method according to any one of claims 1 to 10, in which the physical values ​​of different types are correlated spatially and / or temporally.

12. Device comprising an element for storing instructions for implementing a production method according to any one of claims 1 to 11.

13. Neural computing device, characterized in that it comprises: • a deep learning unit of physical data from measurements and / or digital simulations of at least one physical phenomenon, such as a flow of a fluid, a heat transfer, an electromagnetic field or a structural mechanical constraint, carried out at least in the vicinity of a so-called training surface, said training surface being able to be distinct for each data set, the learning unit generating weights of a neural network; • a unit for predicting physical values ​​for all or part of the points of a point cloud of a simulation space configured at least in the vicinity of a geometry comprising a plurality of surface points A tforming a surface mesh, said geometry being capable of being subjected to at least one physical phenomenon, such as a flow of a fluid, a heat transfer, an electromagnetic field or a structural mechanical constraint, the prediction being carried out as a function of at least one so-called simulation condition, the physical values ​​forming a simulation result.

Images