Turbomachine blade manufacturing process using a neural network

Abstract

The present invention relates to a method for manufacturing a turbomachine blade, comprising the steps of: Generation (S1), by a neural network and from a reference value of a first parameter (P1) characteristic of the blade, of representative data for a set of candidate geometries for the blade, the data comprising, for each candidate geometry, a value of the first parameter (P1) satisfied by the candidate geometry and satisfying a proximity condition with the reference value; from the representative data generated by the neural network, determination (S2) of a blade geometry that best satisfies a value of a second parameter (P2) characteristic of the blade, the second parameter (P2) characteristic of the blade being different from the first parameter (P1) characteristic of the blade; manufacturing (S3) the turbomachine blade according to the determined blade geometry. Figure for the abstract: Fig. 1

Description

Title of the invention: Method for manufacturing turbomachine blades using a neural network technical field

[0001] This disclosure relates to the general field of methods for assisting in the design of profiles of turbomachine parts, in particular parts having an aerodynamic profile such as turbomachine blades. STATE OF THE ART

[0002] The manufacture of turbomachine parts, and in particular the manufacture of new parts with an aerodynamic profile, such as turbine or compressor blades, is a long and complex process.

[0003] When the turbomachine is in operation, the blades are subjected to a flow of gas in a flow channel. The aerodynamic profile of the blades, defined by their geometry, directly affects the performance of the turbomachine. Thus, shape optimization is a very active research area due to strong industrial interest.

[0004] The geometry of the blade must comply with operational and structural constraints; for example, it must allow the manufactured part to withstand a certain rotational speed and resist significant stresses and high temperatures. The geometry must also satisfy installation constraints; for example, the maximum permissible height of the blade must comply with the dimensions of the flow channel where it is fixed or rotates.

[0005] The objective of a design step is to provide a geometry that meets required performance criteria, while respecting these different technical constraints, so that it can be manufactured.

[0006] In general, existing design processes make it possible to obtain a single blade geometry, verifying a chosen performance criterion, and satisfying technical constraints.

[0007] To optimize the geometry, it is necessary to define the representation of the geometry, for example, explicit, implicit, parameterized, and to choose the optimization algorithm. Indeed, obtaining the optimal geometry depends on a parameterization of the blade geometry and the chosen optimization algorithm.

[0008] A classic example of an optimization algorithm is a method based on a gradient calculation, representing the sensitivity of the performance criterion to be optimized with respect to the parameterization of the blade geometry. By modifying the parameterization Through iteration, we can obtain an optimal geometry, that is to say, one corresponding to an optimum of the performance criterion.

[0009] However, this method is very costly because at each iteration it requires solving the relevant physical equations through complex computational fluid dynamics (CFD) calculations and performing numerous gradient calculations. It is also uncertain because it is impossible to predict the number of iterations required to obtain the optimal geometry, which depends heavily on the initial conditions. Thus, to obtain a variety of blade geometries that meet the technical constraints and satisfy the performance criterion, it is necessary to repeat the optimization process several times starting from several different initial conditions.

[0010] Other gradient-free optimization methods have also been proposed, for example, reinforcement learning algorithms. However, these optimization methods are conditioned by the chosen performance criterion and do not allow for the rapid generation of a variety of geometries satisfying different performance criteria. Thus, before the blade can be manufactured, numerous iterations are necessary between the different design offices.

[0011] Bayesian inference-based optimization methods have also been proposed. These methods allow for obtaining a variety of geometries, in the form of a distribution, for a chosen performance criterion. However, they require resampling when the performance criterion is changed or when a technical constraint is modified. Since the convergence of the optimization process can be slow, this solution is not satisfactory for meeting industrial challenges.

[0012] Thus, there are no efficient and rapid methods for generating a wide variety of blade profiles that meet technical constraints, making it possible to determine a blade profile that satisfies performance criteria. In particular, there is no turbomachine blade design process that allows for the direct generation of a set of geometries compatible with technical constraints, with a predefined number of calculation steps, and therefore a limited cost, in order to manufacture a turbomachine blade determined according to physical criteria. Description of the invention

[0013] One aim of the proposed method is to facilitate the design and manufacture of a new turbomachine blade geometry that meets geometric constraints and performance criteria.

[0014] This objective is achieved by a method for manufacturing a turbomachine blade, comprising the following steps: - generation, by a neural network and from a reference value of a first characteristic parameter of the dawn, of data representative of a set of candidate geometries for the dawn, the data including, for each candidate geometry, a value of the first parameter satisfied by the candidate geometry and verifying a proximity condition with the reference value, - from the representative data generated by the neural network, determination of a dawn geometry that best satisfies a value of a second characteristic parameter of the dawn, the second characteristic parameter of the dawn being different from the first characteristic parameter of the dawn, - manufacturing of the turbomachine blade according to the determined blade geometry.

[0015] The neural network of the process according to this disclosure makes it possible to rapidly generate turbomachine blade profiles that meet predefined technical constraints, by conditioning on the value of the first parameter characterizing the set of candidate geometries from which a geometry is selected to be subsequently manufactured. The selected geometry satisfies a value of the second parameter. The process overcomes limitations of existing processes by allowing rapid access to representative data of a set of candidate geometries for the turbomachine blade, conditioned by a plurality of parameters corresponding to performance criteria and technical constraints.

[0016] The computational cost of the proposed method for generating several geometries that meet a set of constraints, corresponding to the verification of the proximity condition with the reference value of the first parameter, is lower than that of existing methods. Indeed, the use of such a neural network avoids the need for geometric optimization processes requiring numerous computational fluid dynamics (CFD) simulations and gradient calculations. This allows for faster exploration of candidate geometries and simplifies the determination of the geometry to be manufactured, based on the value of the second parameter. The manufacturing process is less biased because the determined geometry is chosen from a wide variety of candidate geometries and satisfies the value of the second parameter.

[0017] The invention is advantageously complemented by the following features, taken individually or in any of their technically possible combinations: - the second parameter is a performance criterion, chosen in particular from: - a compression ratio, and - fluidodynamic efficiency; - the first parameter is a geometric constraint of the turbomachine blade, specifically chosen from: - a length, - a rotational speed, - pressure, - a mass flow rate; - the neural network implements a diffusion model to generate representative data; - the neural network generates the data from N respective reference values ​​of N first characteristic parameters of the dawn which are different, the representative data comprising, for each candidate geometry, N respective values ​​of the N first parameters which are satisfied by the candidate geometry and which satisfy respective proximity conditions with the N reference values; - the representative data include, for each candidate geometry, a latent representation of the candidate geometry, the latent representation being of smaller dimension than representative data of a mesh of the candidate geometry, and in which the determination of the geometry of the dawn includes the determination of representative data of the mesh of the geometry of the dawn from the latent representation; - the determination of representative data of the mesh of the geometry of the blade includes an operation inverse to a principal component analysis, the inverse operation being applied to the latent representation of the geometry of the blade; - The determination of representative data of the blade geometry mesh is implemented by an auto-encoder. DESCRIPTION OF THE FIGURES

[0018] Other features, objectives and advantages of the invention will become apparent from the following description, which is purely illustrative and not limiting, and which should be read in conjunction with the accompanying drawings on which:

[0019] Fig. 1 is a flowchart of steps of a manufacturing process according to an embodiment.

[0020] Fig. 2 is a schematic representation of distributions of values ​​of three characteristic parameters of the dawn, obtained from the set of candidate geometries generated by the neural network, in an embodiment where the neural network includes an unconditioned generative model.

[0021] Fig. 3, Fig. 4 and Fig. 5 are schematic representations of distributions of values ​​of three characteristic parameters of the blade, obtained from the set of candidate geometries generated by the neural network, in several embodiments.

[0022] Fig. 6 is a schematic representation of one embodiment of the manufacturing process.

[0023] Fig. 7 is a schematic representation of the steps to obtain a training database to train the neural network, in an example embodiment.

[0024] The [Fig.8] is a schematic representation of sub-steps of a geometric mesh transformation step in the example of implementation of the [Fig.6].

[0025] Fig. 9 is a schematic representation of the learning steps of the neural network, in an embodiment where the neural network implements a diffusion model.

[0026] Throughout the figures, similar elements bear identical references. DETAILED DESCRIPTION OF THE INVENTION

[0027] New turbomachine blade geometries are, in a manner known per se, generally manufactured after obtaining a geometry by computer-aided design. Generally, a design device comprises a processor, enabling the execution of computer program instructions, and memory, which enables the storage of the computer program. The memory also allows for data storage, for example, the storage of databases representing turbomachine blade geometries.

[0028] The computer program includes instructions for implementing a design process as described below. In particular, it may contain the characteristic data of a neural network trained to generate representative data of turbomachine blade geometries. The neural network is represented by a succession of layers performing operations on weighted inputs. The set of weight values ​​can be stored in memory and accessed by the computer program to enable the implementation of the design process.

[0029] Figure 1 schematically illustrates the steps of the manufacturing process according to one embodiment. This process comprises the following steps.

[0030] The manufacturing process first includes design steps for a blade geometry. The first SI step of the process is the generation, by the neural network and from a reference value of a first characteristic parameter of the blade, of data representative of a set of candidate geometries for the turbomachine blade, the data comprising, for each candidate geometry, a value of the first parameter satisfied by the candidate geometry and verifying a proximity condition with the reference value.

[0031] Generally, and in a manner known per se, a turbomachine blade geometry is associated with a plurality of characteristic parameters.

[0032] Characteristic parameters can characterize the shape, that is, the three-dimensional geometry of the turbomachine blade. For example, a characteristic parameter can correspond to a volume of the blade or to a length characterizing a dimension of the blade. Since the blade extends between a root and a tip, such a length can be a height of the blade, that is, the distance between the root and the tip of the blade, or a thickness of the blade, characterizing the size of the blade within a flow channel of the turbomachine. In the case where the geometry of the blade is generated by a parameterized generating curve, the characteristic parameters can include parameters of the generating curve.

[0033] The characteristic parameters can characterize constraints related to the use of the blade, for example, correspond to quantities representative of its operation. In particular, the parameters can characterize flow conditions to which the blade is subjected in use, for example the mass flow rate of a gas flow in the flow channel, a temperature or pressure value at a point on the blade, or even a rotational speed of a rotor part of the turbomachine on which the blade is fixed.

[0034] Furthermore, the characteristic parameters can also characterize the expected performance of the blade during its use, that is, correspond to a performance criterion of the blade. The performance criteria may concern, for example, a ratio between compression and expansion, also called the compression ratio, or a fluidodynamic efficiency.

[0035] More specifically, the performance criteria can be calculated from the motion field induced by the geometry of the blade. For a chosen geometry of the turbomachine blade, a given inlet condition of the gas flow, i.e. fixed values ​​of pressure, velocity and temperature of the incident gas flow, a rotor operating condition, for example the rotational speed, and a geometry of the flow channel can be varied independently.

[0036] The geometry of the blade deflects the gas flow within the duct, inducing acceleration or deceleration. The resulting velocity variations generate a decrease or increase in pressure. Numerical simulations can be used to determine the motion field induced by the blade geometry. Performance criteria of interest for addressing industrial challenges, such as compression ratio or efficiency, can then be associated with a given geometry.

[0037] The compression ratio corresponds to the pressure difference upstream and downstream of the blade, normalized relative to the upstream pressure. The efficiency depends on the change in flow entropy between the upstream and downstream sides of the blade caused by fluid dynamic losses.

[0038] Thus, when designing a new turbomachine blade geometry, numerical simulations can be used to estimate the value of the blade performance criteria under fixed flow conditions.

[0039] In order to implement the process according to this disclosure, it is necessary to select at least one first characteristic parameter and one second characteristic parameter, from among the various parameters characterizing a blade geometry.

[0040] A reference value is chosen for the first parameter. This reference value allows for conditioning the generation by the neural network of a set of candidate geometries for the turbomachine blade. More precisely, "conditioning" means that the reference value is provided as input to the neural network and has an impact on the neural network's output. Here, the reference value constrains the neural network, in that the value of the first parameter satisfied by each generated candidate geometry must satisfy a proximity condition with the reference value.

[0041] More specifically, since each turbomachine blade geometry is associated with a value for the first parameter, it is possible to obtain the distribution of the values ​​of the first parameter for the candidate geometries generated by the neural network.

[0042] The proximity condition can correspond to a standard deviation of the distribution of values ​​of the first parameter thus obtained, for example. The set of candidate geometry of the blade will thus have a value of the first parameter within an interval centered on the reference value with a certain probability.

[0043] Typically, the first parameter corresponds to a geometric constraint of the blade.

[0044] Preferably, the conditioning of the generation by the neural network of data representing the set of candidate geometries for the turbomachine blade is done according to a plurality of parameters, representing both geometric constraints and expected performance criteria. For example, this allows us to guarantee that the geometry determined during a second step of the process allows the dawn to meet expected performance criteria.

[0045] For example, the neural network can generate representative data for a set of candidate geometries from N respective reference values ​​of N first characteristic parameters of the turbomachine blade. The data include, for each candidate geometry, N respective values ​​of the N first parameters that are satisfied by the candidate geometry and that satisfy respective proximity conditions with the N reference values.

[0046] More generally, the neural network can be conditioned by a policing function depending on the N respective reference values ​​of the first N characteristic parameters representing geometric constraints and performance criteria of the turbomachine blade. For example, the policing function can be a linear combination of normalized values ​​of the N respective reference values.

[0047] The second step S2 of the process is the determination of a blade geometry from the representative data generated by the neural network during the first step SL. The determined geometry best satisfies a value of a second characteristic parameter of the blade. The second characteristic parameter of the blade is different from the first characteristic parameter of the blade.

[0048] By "at best" means that the geometry is determined so as to be associated with a value of the second parameter equal to or close to a chosen value, within a certain interval, for example a range of 10% around the chosen value, preferably a range of 1% relative to the chosen value.

[0049] Preferably, the data generated by the neural network includes, for each candidate geometry, a value for the second parameter. Thus, for the set of candidate geometries, it is possible to easily obtain a distribution of the values ​​of the second parameter.

[0050] It is possible to determine the geometry of the blade so that it has a value of the second characteristic parameter satisfying an optimality criterion.

[0051] The optimality criterion may be, for example, an extreme value of the second parameter within the obtained distribution, for example, the maximum. Generally, the second parameter may correspond to a performance criterion for the blade. The optimality criterion may be having a value of the second parameter greater than an expected performance threshold, for example, one imposed by a standard.

[0052] The determination of the geometry of the blade can be done by an optimization algorithm, for example gradient method, to determine among the set of candidate geometries a geometry which satisfies the value of the second parameter.

[0053] The third step S3 of the process is the fabrication of the turbomachine blade, according to the geometry determined in the second step S2. The turbomachine blade It can be manufactured using various manufacturing methods. As is known, the blade can be produced by machining, molding, or additive manufacturing, and can incorporate composite materials.

[0054] By way of illustration, consider a manufacturing process for a turbomachine blade that depends on three parameters of interest. The first parameter represents a mass flow rate, corresponding to flow conditions. The second parameter represents a compression ratio, and the third parameter represents fluid dynamic efficiency, and these correspond to performance criteria.

[0055] Figures 2 to 5 show distributions of values ​​for the three parameters, obtained from the parameter values ​​characterizing the set of candidate geometries generated by the neural network.

[0056] More specifically, the distributions in black lines correspond to distributions calculated from a reference database and are identical in Figures 2 to 5. The shaded distributions correspond to distributions of parameter values ​​obtained using different proximity conditions for data generation by the neural network.

[0057] As is known per se, the representative turbomachine blade geometry data from the reference (test) database are not used during the training of the neural network. However, the reference database is representative of an experimental design used to generate the complete set of representative data.

[0058] Figure 2 shows parameter distributions for candidate geometries generated by the neural network in the absence of conditioning on the value of the first parameter. When the generation of the set of candidate geometries is not conditioned by imposing no proximity condition with the reference value of the first parameter, or more generally of the other parameters associated with the geometry of the turbomachine blade, the distributions of values ​​for the set of candidate geometries are similar to those obtained from the geometries in the reference database.

[0059] A similarity between the distributions obtained by the neural network and those obtained from the geometries of the reference database (test) demonstrates correct operation of the neural network.

[0060] However, it is advantageous to condition the neural network according to the reference value chosen for the first parameter. Since the neural network is trained to generate data including, for each candidate geometry, the value of the first parameter satisfied by the candidate geometry, the distribution of values ​​of the first parameter for all generated geometries can easily be obtained. As illustrated in [Fig. 3], a distribution of the values ​​of the first parameter centered on the reference value. The determination of a geometry can be done by selecting a candidate geometry from a subset of candidate geometries having a compression ratio value and an efficiency value equal to (or greater than) respective thresholds, for example.

[0061] By modifying the proximity condition with the reference value, it is also possible to modify the intensity of the conditioning, so as to reduce the impact of the reference value of the first parameter on the generation of candidate geometries. This is illustrated in [Fig. 4]. It can be seen that the standard deviation of the distribution of the first parameter is still centered on the chosen reference value, but that the standard deviation of the distribution is larger. This also modifies the distributions for the values ​​of the other characteristic parameters.

[0062] As explained previously, it is also possible to condition the generation of candidate geometries according to a plurality of parameters.

[0063] In the embodiment illustrated in [Fig. 5], the generation of candidate geometries by the neural network is conditioned on reference values ​​for the first and third parameters. The calculated distributions of the values ​​of the first and third parameters are centered on the chosen reference value, as expected. The distribution of the second parameter is advantageously obtained directly from the data generated by the neural network.

[0064] We can then determine (step S2) the geometry of the blade as the candidate geometry presenting the value corresponding to the maximum compression ratio, and then manufacture (step S3) a turbomachine blade presenting this geometry.

[0065] Figure 6 schematically represents steps S1, S2 and S3 in an embodiment of the process depending only on a first parameter PI and a second parameter P2. The first input parameter PI (for example, the mass flow rate) is used to condition the generation of turbomachine blade geometries by the neural network.

[0066] In this embodiment, the representative data for the set of candidate geometries contain, for each candidate geometry, a value for the first parameter, a value for the second parameter, and a mesh. The proximity condition is that the values ​​of the first parameter satisfied by the candidate geometries follow a standardized distribution around the reference value D. A distribution for the values ​​of the second parameter P2 can be obtained from the set of candidate geometries generated by the neural network in step SL. For example, the second parameter P2 is a performance criterion, in particular the fluid dynamic efficiency of the geometry. In the second step S2, the candidate geometry associated with the maximum value E* of the second parameter P2 is determined. The geometry of the turbomachine blade thus satisfies the performance criterion. represented by an optimality criterion on the second parameter P2, while best satisfying the geometric constraint (flow conditions) represented by the first parameter PI.

[0067] As can be seen from the above, the flexibility of the neural network allows the generation of data to be conditioned with respect to one or more parameters, the value of the conditioning to be changed by modifying the reference value given as input for the first parameter, and advantageously, the intensity of the conditioning to be modified by changing the proximity condition with the reference value. Advantageously, the neural network does not need to be retrained when the flow conditions to which the blade is subjected, for example, are changed. This saves time during the design stages.

[0068] It is also particularly advantageous to quickly obtain different geometries that satisfy the technical constraints for different performance criteria. Indeed, the choice of a geometry results from a compromise between the various performance criteria and generally requires the intervention of an expert, which introduces a significant bias. This bias is avoided in the described process.

[0069] Neural network architecture

[0070] According to this disclosure, the neural network can be trained from a training database comprising results from computational fluid dynamics (CFD) simulations.

[0071] The generation of representative data for candidate geometries for the turbomachine blade, conditioned by the reference value of the first parameter, is based on deep learning generative models. Thus, the neural network of the process is capable of generating new blade geometries similar to the blade geometries in the training database.

[0072] Advantageously, the generative model makes it possible to generate a plurality of turbomachine blade geometries, which can satisfy different criteria, without requiring retraining of the neural network.

[0073] In one embodiment, the neural network implements a diffusion model to generate the data.

[0074] Implementing a diffusion model as a generative model has proven to be more efficient in terms of the quality of the generated blade geometries and numerical efficiency during the learning and inference stages of the generative model.

[0075] The data in the training database are assigned a membership class, corresponding, for example, to ranges of values ​​for the different parameters. It is possible to condition the diffusion model on the membership class of the blade geometries in the training database. This This allows the generation of data of a specific class, and in particular turbomachine blade geometries satisfying performance criteria and / or geometric constraints.

[0076] Preferably, the diffusion model can be trained to achieve a compromise between a conditioned and unconditioned diffusion model, using a class-free method. In particular, it is possible to condition with respect to a scalar value, rather than with respect to a class. Specifically, it is possible to condition with respect to an expected performance criterion value, or to satisfy a geometric constraint.

[0077] In an alternative embodiment, the generative model of the neural network can be chosen from among other solutions capable of generating a set of candidate geometries satisfying a proximity condition with the reference value, or in other words of reproducing and sampling a conditioned distribution.

[0078] More specifically, the diffusion model is trained to reproduce the distribution of the representative dataset of turbomachine blade geometry. For example, the distance between the generated distribution and the reference distribution can be minimized via an optimization process based on calculating the gradients of a cost function. However, in the field of deep learning, the optimum achieved may be a local rather than a global optimum, so the optimization process does not guarantee the shape of the resulting distribution.

[0079] For example, the neural network can be a conditioned variational autoencoder (“Variational Autoencoders” or VAE).

[0080] Alternatively, the neural network can be a conditioned generative adversarial neural network (“Generative Adversarial Neural Networks” or GAN).

[0081] Alternatively, the neural network can be a conditioned normalizing flows network.

[0082] Since the generative model is conditioned by parameters, and in particular by the reference value of the first parameter, it is possible to impose a distribution to be obtained for some of the parameters, and to observe the impact that the conditioning has on the distribution of other parameters, as in figures 2 to 5.

[0083] Learning database

[0084] With reference to [Fig.7], the neural network generation model is trained from an initial training database Mi. The training database comprises a set of turbomachine blade geometries, and for each blade geometry, the values ​​of the associated parameters.

[0085] Generally, a database is generated after defining an experimental plan taking into account ranges of values ​​representing different flow conditions, and ranges of values ​​representing an permissible geometric variability of the turbomachine blade.

[0086] For example, the geometry of a turbomachine blade is characterized by an original mesh. The original mesh comprises a plurality of geometric mesh elements, typically the spatial coordinates of nodes in the original mesh, edges connecting the nodes of the original mesh, and polygons encompassing the edges to form the three-dimensional geometry of the turbomachine blade. Typically, the polygons are triangles or quadrilaterals.

[0087] As explained previously, for each turbomachine blade geometry, it is possible to calculate a plurality of parameters representative of the blade geometry, characterizing, for example, the flows induced by the blade geometry under the imposed conditions. For example, numerical fluid dynamics simulations can be performed by a simulator, allowing a value for several parameters to be associated with certain geometric mesh elements. For example, the training database can contain a surface pressure value for each polygon forming the mesh.

[0088] From the set of values ​​calculated by the numerical simulations, it is thus possible to calculate a plurality of parameter values ​​representative of the performance of the blade geometry under the flow conditions.

[0089] In a complementary or alternative manner, each numerical simulation can be associated with a plurality of parameters representative of the conditions of use of the blade, for example flow conditions.

[0090] In the initial training database Mi, a turbomachine blade geometry can thus be associated with a plurality of characteristic parameters. As explained previously, the characteristic parameters can represent, for example, a geometric constraint, an operating condition, or a simulated performance.

[0091] In a preferred embodiment, the training of the generative model of the neural network can be carried out from a reduced training database Mr, comprising, for each blade geometry, a latent representation of the geometry. A latent representation is understood to be a reduced-dimensional vector associated with the geometry of the turbomachine blade.

[0092] For each geometry, the latent representation is of lower dimension than the representative mesh data. Preferably, the representation vector is a vector of dimension n, with n less than 100. Typically, one can choose n = 30 or n = 60. As explained later, the dimension n must be sufficient to allow a decompression algorithm is used to reconstruct the mesh. The dimension n can be chosen using statistical indicators such as the explained variance or the Frobenius norm. Preferably, the dimension n is chosen to accurately represent (according to criteria based on the aforementioned indicators) a majority of the training dataset, for example, at least 99% of the data in the training dataset. Preferably, the latent representation vector is in bijection with the initial mesh of the turbomachine blade.

[0093] An example of obtaining the reduced training database Mr from the initial training database Mi is schematically illustrated in [Fig.7]. In this embodiment, the acquisition comprises three successive steps: a geometric transformation step of the mesh Bl, a dimensionality reduction step B2, and a data concatenation and normalization step B3.

[0094] During the geometric transformation step of the mesh Bl, each initial mesh of the initial training database Mi is transformed into a reference mesh stored in an intermediate database Mf.

[0095] During the dimension reduction steps B2 and concatenation and normalization B3, the reference meshes of the intermediate database Mf obtained in step Bl are transformed into training vectors and stored in the reduced training database Mr.

[0096] More specifically, during the dimension reduction step B2, the reference meshes of the intermediate database Mf obtained in step Bl are transformed into latent representations of the mesh (also called latent representation vectors).

[0097] Finally, during the data concatenation and normalization step B3, the latent representation vectors can be concatenated with the parameter values ​​associated with the geometry of the turbomachine blade from the initial training database Mi. The concatenated vectors can be normalized ("scaling") to obtain homogeneous training vectors suitable for training the neural network.

[0098] The training vectors are stored in the reduced training database Mr.

[0099] In a preferred embodiment, the geometric transformation step of the mesh Bl comprises several intermediate steps B11, B12 schematically illustrated in [Fig.8].

[0100] An initial mesh of the initial training database Mi is first transformed into a deformed mesh Mm, by a mesh adaptation operation B11 (or "mesh morphing" according to Anglo-Saxon terminology).

[0101] During the mesh adaptation step B11, the geometric coordinates of the initial mesh elements and the associated parameters obtained by numerical simulation are, for example, transferred to the corresponding element of the deformed mesh Mm.

[0102] The deformed meshes Mm can then be interpolated onto a reference manifold corresponding to a common mesh G in two dimensions, in order to obtain homogeneous meshes Mf. The interpolation can be performed by projection. This ensures that all the deformed meshes Mm have the same number of geometric elements and are associated with the same coordinates.

[0103] For example, the common mesh G, also called the "reference manifold", can be a square of unit area, discretized according to a Cartesian grid. More precisely, each side of the square is divided into X segments of identical length 1 / X, so that the Cartesian mesh defined by the Cartesian grid comprises X2 cells of square shape and area {1 / X}².

[0104] The coordinates of the nodes of the initial mesh Mi and the values ​​of the simulated parameters are then interpolated on the nodes of the Cartesian mesh G. This gives, for each initial mesh Mi, a homogeneous mesh Mf (or "iso-manifold" representation) of the geometry of the turbomachine blade.

[0105] It will be understood that obtaining the reduced training database is not limited to this embodiment. In general, any approach that allows each blade geometry in the initial database to be represented as a vector of predefined fixed dimensions is suitable.

[0106] Alternatively, the latent representation vector can be obtained by randomly selecting a fixed and predetermined number of points from the mesh representing the geometry of each blade in the initial database.

[0107] The dimension reduction step B2 can be implemented by different technical solutions.

[0108] In one embodiment, the dimensionality reduction step B2 is performed by a variational autoencoder neural network (VAE). Such a neural network can be capable of generating a latent representation vector from the initial mesh Mi, the deformed mesh Mm, or preferably the homogeneous mesh Mf corresponding to the iso-manifold representation.

[0109] In a preferred embodiment, the dimension reduction step B2 is carried out by a principal component analysis (PCA) reduction method of the initial mesh Mi, the deformed mesh Mm or preferably the homogeneous mesh Mf.

[0110] For example, the latent representation vector may include PCA mode coefficients. During the data concatenation step B3, data representing geometric constraints may be concatenated with the data representing the geometry of the turbomachine blade to obtain the drive vector. Geometric constraints may be implicitly imposed during the dimensionality reduction step B2.

[0111] For example, a turbomachine blade height or a turbomachine blade root orientation can be integrated into the modal representation of the geometry when the latent representation vector is obtained by principal component analysis. Indeed, if all the blades in the initial database comply with a geometric constraint, for example, the height, then the modes of the PCA database allow for obtaining a linear combination of blade geometries that satisfy the geometric constraint.

[0112] The determination of representative data for the blade geometry mesh can be implemented using an autoencoder. For example, when the latent representation vector is obtained by a variational autoencoder neural network, a normalization of the reduced representation can be performed to ensure that the candidate geometry complies with the geometric constraints.

[0113] More generally, the reduced training database Mr can be obtained by a dimensionality compression algorithm. The compression process then includes a dimensionality decompression algorithm, allowing a discretized blade geometry to be obtained from the latent representation vector. Such a dimensionality decompression algorithm can be obtained iteratively, in order to allow for a correct reconstruction of the blade geometry.

[0114] The reduced training database Mr containing all the training vectors thus obtained can then be used to train the neural network to generate new turbomachine blade profiles.

[0115] Neural network training

[0116] It is advantageous to use a reduced training database to train the neural network. Indeed, this allows the neural network to learn essential features of the blade geometry more simply, using reduced-dimensional data.

[0117] Once trained, the neural network is capable of generating a set of candidate geometries for the turbomachine blade, as a function of at least one input parameter, and in particular of the reference value of the first parameter.

[0118] In an embodiment where the generative model of the neural network is trained from the reduced training database Mr, the network of neurons generates data including, for each candidate geometry, a latent representation of the candidate geometry.

[0119] In a manner known per se, an example of a diffusion model learning process is schematically illustrated in [Fig.9].

[0120] The diffusion model of the neural network is trained iteratively. The training dataset Mr corresponds to an initial distribution. A noise-making algorithm is applied to each latent representation vector of the training dataset Mr, in order to obtain a noisy dataset Mb. The noise-making algorithm is applied as long as the noisy dataset Mb does not satisfy a reference distribution. The noisy dataset Mb satisfying the reference distribution is given as input to the diffusion model. The diffusion model also takes as input the first parameter PL. The first parameter PI allows the conditioning of the neural network, for example, according to the flow conditions or desired performance criteria. The diffusion model provides as output a dataset Mc corresponding to the candidate geometries of turbomachine blades.As long as the distribution associated with the Mc dataset output of the diffusion model does not correspond to the initial distribution (before noise), or to a proximity condition with the first parameter PI, the diffusion model is updated, i.e., corrected. The neural network training ends when the Mc dataset satisfies the initial distribution.

[0121] In this embodiment, step S2 of determining the geometry of the blade includes determining data representative of a mesh of the part from the latent representation.

[0122] For example, the process according to this disclosure may include a step of decoding the latent representation vector corresponding to the sample of the distribution of values ​​of the second parameter that best satisfies the value of the second parameter, or an optimality criterion. Implementing the decoding step makes it possible to obtain a three-dimensional mesh characterizing the geometry of the turbomachine blade from the latent representation corresponding to the geometry. Indeed, it is possible to reconstruct the mesh characterizing the geometry of the blade by reversing the process of obtaining the latent representation vector described above.

[0123] The determination of representative data of the mesh of the geometry of the blade may include an operation inverse to a principal component analysis, the inverse operation being applied to the latent representation of the geometry of the blade.

[0124] For example, in the embodiment where the latent representation is obtained by principal component analysis, the determination of the mesh may include the inverse operation, applied to the latent representation of the geometry satisfying the value of the second parameter. The number of modes kept in the decomposition, and therefore the dimension n of the latent representation vector, can be obtained by iteration, in order to obtain a correct reconstruction of the mesh.

[0125] Reconstruction of the reference mesh Mf or of the initial mesh Mi by an inverse transformation operation may be possible.

[0126] Preferably, the operations carried out during the interpolation steps B12 and mesh adaptation B11 are invertible, so that the geometry in the form of a three-dimensional mesh, structured or unstructured, of the turbomachine blade is obtained directly from the determined latent representation vector.

[0127] In the embodiment where the latent representation vectors are obtained by a variational auto-encoder neural network, the determination of the mesh at step S2 can be implemented by an auto-encoder.

[0128] More specifically, a regularization can be used to solve an inverse problem and obtain the initial mesh corresponding to the geometry of the turbomachine blade.

[0129] More generally, the decoding step can be implemented by the dimensionality decompression algorithm.

[0130] This disclosure also relates to a computer program product, comprising instructions for implementing the S1 and S2 design steps of the manufacturing process described above, when the program product is executed by a computer. The computer program may include instructions for running the trained neural network, based on the intended inputs. It may include instructions for determining the blade geometry. If several candidate geometries satisfy the value of the second parameter, the computer program may provide decision support for an expert. For example, the computer program may calculate characteristic parameter distributions of the candidate blade geometries that best satisfy the value of the second parameter, and command the display of the distributions on a human-machine interface, for example, a screen.Alternatively, the instruction can randomly determine the blade geometry from among the candidate geometries that best satisfy the value of the second parameter. If none of the candidate geometries satisfy the value of the second parameter, the instruction can determine the geometry with the closest value.

[0131] The initial training database Mi and / or the reduced training database Mr can be stored in memory accessible by the computer program.

[0132] The weights of the neural network after training can be stored in accessible memory.

Claims

Demands

1. A method for manufacturing a turbomachine blade, comprising the steps of: - Generation (S1), by a neural network and from a reference value of a first parameter (P1) characteristic of the blade, of data representative of a set of candidate geometries for the blade, the data comprising, for each candidate geometry, a value of the first parameter (P1) satisfied by the candidate geometry and satisfying a proximity condition with the reference value, - from the representative data generated by the neural network, determination (S2) of a geometry of the blade best satisfying a value of a second parameter (P2) characteristic of the blade, the second parameter (P2) characteristic of the blade being different from the first parameter (P1) characteristic of the blade, - manufacturing (S3) of the turbomachine blade according to the geometry of the blade determined.

2. A method according to claim 1, wherein the second parameter (P2) is a performance criterion, in particular chosen from: - a compression ratio, and - a fluidodynamic efficiency.

3. A method according to any one of claims 1 and 2, wherein the first parameter (PI) is a geometric constraint of the turbomachine blade, and in particular selected from: - a length, - a rotational speed, - a pressure, - a mass flow rate.

4. A method according to any one of claims 1 to 3, wherein the neural network implements a diffusion model to generate representative data.

5. A method according to any one of claims 1 to 4, wherein the neural network generates representative data from N respective reference values ​​of the first N characteristic parameters of the dawn which are different, the representative data including, for each candidate geometry, N respective values ​​of the first N parameters which are satisfied by the candidate geometry and which satisfy respective proximity conditions with the N reference values.

6. A method according to any one of claims 1 to 5, wherein the representative data comprise, for each candidate geometry, a latent representation of the candidate geometry, the latent representation being of smaller dimension than representative data of a mesh of the candidate geometry, and wherein the determination of the blade geometry comprises the determination of representative data of the mesh of the blade geometry from the latent representation.

7. A method according to claim 6, wherein the determination of the representative data of the blade geometry mesh comprises an operation inverse to a principal component analysis, the inverse operation being applied to the latent representation of the blade geometry.

8. Method according to claim 6, wherein the determination of representative data of the blade geometry mesh is implemented by an auto-encoder.