Method and device for additive manufacturing of a photopolymerized ceramic part

Abstract

The invention relates to a method for manufacturing a part (P) from photopolymerized ceramic material using an additive manufacturing machine, comprising the following steps: training (E2) a first neural network on the images (XAi), obtaining (E3) N third images (PRi) of the material from the test parts (P'), training (E4) a second neural network, correcting, from the first and second networks having been trained, one of the images (XAi) into a corrected image (Xopt(i)) of the setpoint, the machine manufacturing (E10) the part (P) by depositing and photopolymerizing the N photopolymerization layers (Ti) in the N superimposed planes (Hi) of ceramic material deposition according to the first N two-dimensional setpoint images (XAi), the fourth image (xi) of which has been replaced by the corrected image (Xopt(i)). Figure 2 for the abstract

Description

Title of the invention: Method and device for additive manufacturing of a photopolymerized ceramic part

[0001] The invention relates to a method for manufacturing a part from photopolymerized ceramic material using an additive manufacturing machine, and an additive manufacturing machine.

[0002] The field of the invention relates to high-pressure turbine blades of aeronautical turbomachinery.

[0003] High-pressure turbines are complex components that require an efficient cooling system to function properly.

[0004] The blades of these older generation turbines are known to exist; they are manufactured using the lost-wax casting process and incorporate a cooling circuit made with a ceramic core. Indeed, the cores of the older generation blades are produced by a ceramic injection molding process and therefore must have a demolding geometry.

[0005] Unlike the blades of the old generation, the new generation blades have ceramic cores made by additive manufacturing in order to allow a great freedom of geometries.

[0006] An additive manufacturing machine performs the deposition of several layers of a material to be photopolymerized according to slicing images of the three-dimensional geometry of the part to be manufactured.

[0007] These slicing images are used by the machine to obtain the geometry of the different layers to be printed successively.

[0008] However, the use of these slicing images by the additive manufacturing machine presents the following disadvantages.

[0009] Discretization into slicing images can introduce artifacts on the final part and inaccuracies in the geometry. The cause of these discrepancies between production and definition is the limitation of the resolution of the polymerization element of the machine (the light energy source and the projection system resulting in a specific pixel size), the interaction between the different layers, as well as the diffraction of light on the ceramic particles suspended in the bath to be polymerized.

[0010] An objective of the invention is to obtain a method for manufacturing a part made of photopolymerized ceramic material using an additive manufacturing machine, an additive manufacturing machine and a non-destructive testing machine such as a tomograph or other, which overcome the disadvantages mentioned above.

[0011] To this end, a first object of the invention is a method for manufacturing a part made of photopolymerized ceramic material using an additive manufacturing machine, which receives N initial two-dimensional geometric setpoint images of N slices of selective photopolymerization in N superimposed deposition planes of a photopolymerizable ceramic material, characterized in that the method comprises the following steps, implemented by at least one computer: training a first neural network on the first N two-dimensional setpoint images of the N photopolymerization slices, to learn to reconstruct the first N two-dimensional images into N second two-dimensional images, to minimize a first calculated difference between the N second two-dimensional images and the first N two-dimensional images,obtaining N third images of the material from photopolymerized ceramic test pieces having prescribed real geometric characteristics in the N superimposed planes, training a second neural network on the first N two-dimensional instruction images of the N photopolymerization slices, to learn the N third images, correction, from at least a part of the first trained neural network and from the second trained neural network, of at least one of the first N two-dimensional instruction images of the N photopolymerization slices, called the fourth image, in at least one determined plane among the N superimposed planes into at least a fifth corrected two-dimensional instruction image of the at least one determined slice of the photopolymerizable ceramic material in the at least one determined plane,The additive manufacturing machine produces the part from photopolymerized ceramic material by depositing and photopolymerizing N photopolymerization layers in N superimposed ceramic material deposition planes according to the first N two-dimensional reference images, of which at least a fourth two-dimensional reference image of at least one determined layer of the photopolymerizable ceramic material in at least one determined plane has been replaced by at least a fifth corrected two-dimensional reference image of at least one determined layer of the photopolymerizable ceramic material in at least one determined plane.

[0012] Thanks to the invention, the geometry of the part produced from the corrected image(s) perfectly conforms to the reference image. This resolves irregularities in the additive manufacturing machine, for example, in the case of LCM (lithography of ceramics), the diffraction of light on ceramic particles. The invention thus provides pre-processing of the sliced ​​images, which are to be sent to the additive manufacturing machine. This pre-processing allows for the advance compensation of errors related to the additive manufacturing process.

[0013] According to one embodiment of the invention, the process further comprises, for performing the correction: the formation of a third neural network, comprising in cascade the part of the first neural network that has been trained and the second neural network that has been trained, for said at least a fourth two-dimensional image of the input of at least a determined slice of the photopolymerizable ceramic material in at least one determined plane among the N superimposed planes, the optimization respectively of at least one random vector sent to the third neural network, by minimizing a second calculated difference between the at least a fourth two-dimensional image of the input of at least one determined slice of the photopolymerizable ceramic material in at least one determined plane and a sixth image, which was obtained by the third neural network from the at least one random vector, to obtain at least one optimized random vector, the training of a fourth neural network comprising the part of the first neural network having been trained, the application of the at least one optimized random vector to the fourth neural network,to obtain at least one fifth two-dimensional image corrected to the setpoint of at least one determined slice of the photopolymerizable ceramic material in at least one determined plane. According to one embodiment of the invention, the method further comprises, the manufacturing, using additive manufacturing machines, of test pieces made of photopolymerized ceramic material and having the prescribed actual geometric characteristics, the obtaining by tomography of three-dimensional tomographic representations of the test pieces, the projection of three-dimensional tomographic representations into the N superimposed planes to obtain the N third images of the material of the test pieces.

[0014] According to one embodiment of the invention, the part of the first neural network includes a first decoder of the first neural network.

[0015] According to one embodiment of the invention, the first neural network further comprises a first encoder, located upstream of the first decoder.

[0016] According to one embodiment of the invention, the first neural network comprises a first variational autoencoder, which comprises in cascade a first encoder, a variational calculation module and the first decoder.

[0017] According to one embodiment of the invention, the training of a first neural network is carried out on the first N two-dimensional images of the N photopolymerization slices applied to an input of the first encoder of the first neural network, to learn to reconstruct the first N two-dimensional images into the N second two-dimensional images, which are calculated by the first neural network from the first N two-dimensional images and which are present on an output of the first decoder of the first neural network, the N second two-dimensional images being calculated by the first neural network to minimize the first calculated difference between the N second two-dimensional images and the first N two-dimensional images.

[0018] According to one embodiment of the invention, the first difference is a first binary cross entropy calculated between the N second two-dimensional images and the N first two-dimensional images.

[0019] According to one embodiment of the invention, the second neural network includes a second auto-encoder, which in cascade includes a second encoder and a second decoder.

[0020] According to one embodiment of the invention, the training of the second neural network is carried out on the first N two-dimensional setpoint images of the N photopolymerization slices applied to an input of the second encoder of the second neural network, to learn the N third images at an output of the second encoder of the second neural network.

[0021] According to one embodiment of the invention, the third neural network comprises in cascade the variational calculation module, the first decoder having been trained and the second neural network having been trained.

[0022] According to an embodiment of the invention, the optimization is carried out for at least a fourth two-dimensional setpoint image of at least a determined slice of the photopolymerizable ceramic material in at least one determined plane among the N superimposed planes, to optimize respectively the at least one random vector sent to an input of the variational calculation module of the third neural network, by minimizing a second deviation calculated between the at least a fourth two-dimensional setpoint image of at least a determined slice of the photopolymerizable ceramic material in at least one determined plane and a sixth image, which was obtained at the output of the second encoder of the third neural network from the at least one random vector, to obtain the at least one optimized random vector at the input of the variational calculation module of the third neural network.

[0023] According to one embodiment of the invention, the fourth neural network comprises in cascade the variational calculation module and the first decoder having been trained.

[0024] According to one embodiment of the invention, the application of at least one optimized random vector is carried out at the input of the variational calculation module of the fourth neural network, to obtain on the output of the first decoder of the fourth neural network at least one fifth two-dimensional image corrected to the setpoint of at least one determined slice of the photopolymerizable ceramic material in at least one determined plane.

[0025] According to one embodiment of the invention, during the training of the first neural network, the first encoder learns a first multivariate normal distribution of each of the first N two-dimensional setpoint images of the N photopolymerization slices applied to the input of the first encoder of the first neural network.

[0026] According to one embodiment of the invention, the second deviation is equal to a structural similarity index measure calculated between at least a fourth two-dimensional setpoint image of at least a determined slice of the photopolymerizable ceramic material in at least one determined plane and the sixth image, which was obtained at an output of the third neural network from at least one random vector.

[0027] According to one embodiment of the invention, the second deviation is equal to a Sprensen-Dice index calculated between the at least a fourth two-dimensional setpoint image of the at least a determined slice of the photopolymerizable ceramic material in the at least one determined plane and the sixth image, which was obtained at an output of the third neural network from the at least one random vector.

[0028] According to one embodiment of the invention, the second deviation is equal to a Jaccard index calculated between the at least a fourth two-dimensional setpoint image of the at least a determined slice of the photopolymerizable ceramic material in the at least one determined plane and the sixth image, which was obtained at an output of the third neural network from the at least one random vector.

[0029] According to one embodiment of the invention, the training of the second neural network is carried out so that from the first N two-dimensional setpoint images of the N photopolymerization slices the second neural network calculates N seventh images, which are present on an output of the second neural network and which minimize a third calculated difference between the N seventh images and the N third images.

[0030] According to one embodiment of the invention, the third gap is a third binary cross entropy calculated between the N seventh images and the N third images.

[0031] A second object of the invention is an additive manufacturing machine for a part made of photopolymerized ceramic material, comprising at least one computer for implementing the manufacturing process as described above.

[0032] The invention will be better understood upon reading the following description, given solely by way of non-limiting example with reference to the figures below of the attached drawings.

[0033] [Fig-1] represents a modular synoptic diagram of a manufacturing device according to a method of embodiment of the invention.

[0034] [Fig.2] represents a flowchart of a manufacturing process according to a method of realization of the invention.

[0035] [Fig.3] represents a schematic perspective view of a manufacturing process additive layer by layer of a part.

[0036] [Fig.4] represents a modular block diagram of a first neural network used by the manufacturing process and device according to an embodiment of the invention.

[0037] [Fig.5] represents a modular block diagram of the first neural network used by the manufacturing process and device according to an embodiment of the invention.

[0038] [Fig.6] represents a modular block diagram of a second neural network used by the manufacturing process and device according to an embodiment of the invention.

[0039] [Fig.7] represents a modular block diagram of a third neural network used by the manufacturing process and device according to an embodiment of the invention.

[0040] [Fig.8] represents a modular block diagram of a fourth neural network used by the manufacturing process and device according to an embodiment of the invention.

[0041] [Fig.9] represents a modular block diagram of a convolutional layer of the encoder of the first neural network used by the manufacturing process and device according to an embodiment of the invention.

[0042] [Fig. 10] represents a modular synoptic diagram of a convolutional layer of the decoder of the first neural network used by the manufacturing process and device according to an embodiment of the invention.

[0043] [Fig. 11] represents a modular synoptic diagram of a convolutional layer of the encoder of the second neural network used by the manufacturing process and device according to an embodiment of the invention.

[0044] [Fig. 12] represents a modular synoptic diagram of a convolutional layer of the decoder of the second neural network used by the manufacturing process and device according to an embodiment of the invention.

[0045] An example of a method for correcting at least one of the setpoint images of photopolymerization wafers of a ceramic material of a machine, an additive manufacturing method of a part P using this correction method, a correction device 1000 for its implementation, a method for manufacturing the part P of photopolymerized ceramic material using an additive manufacturing machine 200, and an additive manufacturing machine 200 are described in more detail below with reference to figures 1 to 12.

[0046] In Figures 1 and 2, the additive manufacturing machine 200 and the additive manufacturing process enable the production of a real part P made of a ceramic material using the selective photopolymerization of this ceramic material. The additive manufacturing process is also called three-dimensional printing, or "3D printing." The additive manufacturing machine 200 is also called a three-dimensional printer, or "3D printer." An example of additive manufacturing is the production of ceramics by lithography (abbreviated in English: Lithography-based ceramic manufacturing (LCM)). The additive manufacturing machine 200 includes a deposition device 201 or dispenser 201 for depositing layers of a photopolymerizable ceramic material, which is largely liquid or fluid.This photopolymerizable ceramic material can be a photosensitive masterbatch loaded with ceramic particles, for example in a transparent vat. This masterbatch can consist of one (or more) photosensitive monomer and / or oligomer (for example in the form of a resin) loaded with ceramic particles, which can typically be micron-sized, and various additives (dispersants, solvents, photoinitiator). Selective photopolymerization radiation (R) allows the photopolymerizable ceramic material to be photopolymerized in certain areas forming a slice, thus solidifying the ceramic material in the areas of this slice.

[0047] As shown by way of example in [Fig. 3], each slice T; must correspond to a cross-sectional view of the actual part P in a plane H;. The additive manufacturing machine 200 is configured to print N slices 1) of the ceramic material having been photopolymerized successively in respectively N superimposed planes H;, for i being a natural number from 1 to N, where N is a natural number greater than or equal to 2, and for example greater than or equal to 10. The N slices T; of the ceramic material having been photopolymerized are connected to each other.

[0048] The additive manufacturing machine 200 comprises a printing platform 202 for printing, by means of the dispenser 201 and a selective photopolymerization radiation emission device 204 (or radiation projector 204), N slices T of ceramic material that have been successively photopolymerized on the printing platform 202 in N superimposed planes H, which are parallel to a base plane 203 of the printing platform 202, this base plane 203 being able to be horizontal. The projector 204 comprises an optical block having lenses and an optical correction device for the radiation R.

[0049] During a first step El of the additive manufacturing process implemented using the additive manufacturing machine 200, the computer 100 receives N initial two-dimensional XA images (also called XAj slice images) of the geometry of the N photopolymerization slices T in the N superimposed H planes of ceramic material deposition. These first N two-dimensional XA images may have been previously generated by another computer and received on an input interface of the computer 100. Or these first N two-dimensional XAi images may have been generated by the computer 100. These first N two-dimensional XA images may be recorded during step El in the permanent memory 101 of the computer 100.

[0050] For each slice T, the projector 204 projects the photopolymerization radiation R for an exposure time T (typically less than one second) according to the first two-dimensional geometric instruction image XA, which is an image of the areas where the ceramic material of the part to be manufactured must be located in the deposition plane H of this ceramic material. Thus, the ceramic material is photopolymerized only in the areas designated by the first two-dimensional geometric instruction image XA in the deposition plane H of this ceramic material, i.e., according to slice T. The unpolymerized ceramic material in the deposition plane H is removed. The polymerization of the photopolymerizable ceramic material traps the ceramic particles in the newly formed polymer network.

[0051] According to one embodiment of the invention, each of the first two-dimensional XAi images of the geometry of the N photopolymerization slices T is binary. Thus, each pixel of each of the first two-dimensional XAi images of the geometry of the N photopolymerization slices T is equal to either a bit 1 indicating the presence of photopolymerization, causing the photopolymerization of the photopolymerizable ceramic material at the location targeted by that pixel in slice 1), or a bit 0 indicating the absence of photopolymerization, causing the absence of photopolymerization of the material photopolymerizable ceramic at the location targeted by this pixel in the slice Tj. Of course, each of the pixels of the first two-dimensional XA images and of the corrected two-dimensional image Xopt (i) (or of the corrected two-dimensional images Xopt (i) ) could not be binary and could include other values ​​in the range ]0 ;1[ or the bit 0 or the bit 1.

[0052] Once the slice T; is printed in the plane H; according to the first two-dimensional image XA; of geometric setpoint, the additive manufacturing machine 200 moves the distributor 201 and the projector 204 of the selective photopolymerization radiation R relative to each other in the direction Z perpendicular to the plane H;, to move from a plane H; to the plane Hi+i directly superimposed on this plane H; and to print the next slice Ti+i in the plane Hi+i according to the next first two-dimensional image XAi+j of geometric setpoint corresponding to the next slice Ti+i.The first two-dimensional image XAi+, corresponding to the next slice Ti+i in the next plane Hi+i, may have areas of photopolymerized ceramic material that are different and / or offset and / or common with respect to the areas of photopolymerized ceramic material of the first two-dimensional image XA, corresponding to the slice Ti in the plane Hi, to create curvatures of the photopolymerized ceramic material with respect to the Z direction perpendicular to the plane Hi and to be connected to each other.

[0053] In each plane Hi, each slice Ti can have a thickness controlled by the distributor 201, the thickness being measured along the Z direction perpendicular to the plane Hi. Each slice Ti can have the same or a different thickness compared to the other slices Ti. In each plane Hi, each slice Ti can have a thickness typically between 25 µm and 100 µm.

[0054] The projector 204 can selectively illuminate the areas corresponding to the first two-dimensional image XA of the geometric design of the slice T, which is photopolymerized in the plane H, using a digital micromirror device. The printing is repeated as many times as there are slices T to be polymerized to form the final part P. A heat treatment can be performed to remove the binder and sinter the part P, in order to obtain a solid, durable, and high-density ceramic part P.

[0055] The actual part P to be manufactured by the additive printing machine 200 can be, for example, a turbine blade core A of an aeronautical turbomachine, or a high-pressure turbine blade A of an aeronautical turbomachine. Of course, the actual part P to be manufactured by the additive printing machine 200 could be another part of an aeronautical turbomachine. Part P could be a part making A component of an aircraft turbomachine, particularly an airplane turbomachine. The component could be, for example, a high-pressure turbine ring sector of the turbomachine, a turbine blade, particularly a low-pressure turbine blade, a turbine blade core, a turbomachine air distributor, a turbomachine combustion chamber, or other components. The actual component P to be manufactured could be a lost-wax ceramic casting core.

[0056] An example of a core of such a high-pressure turbine blade A of an aeronautical turbomachine as part P is shown in [Fig.3]. The blade core A has a body 51 of a refractory material having a leading edge 52 and a trailing edge 53, between which the body 51 extends along a first longitudinal direction AX, an extrados 54 and an intrados 55, between which the body 51 extends along a second direction EP of thickness, which is transverse (for example perpendicular) to the first direction AX, a blade foot 56 and a first blade tip surface 57, between which the body 51 extends along a third direction DR of height, transverse (for example perpendicular) to the first and second directions AX, EP, the blade foot 56 having the function of fixing on a longitudinal rotating hub of the turbine.

[0057] According to one embodiment of the invention, the direction Z perpendicular to the plane Hi extends from the trailing edge 53 to the leading edge 52, that is to say against the first longitudinal direction AX as shown in [Fig.3], or extends from the leading edge 52 to the trailing edge 53, that is to say in the first longitudinal direction AX.

[0058] The invention provides a method for correcting at least one image XAj, referred to as the image XAj to be corrected, from among the first N two-dimensional images XA of the N selective photopolymerization layers in the N superimposed planes H of the photopolymerizable ceramic material deposited by the additive manufacturing machine of the part P having the photopolymerizable ceramic material. The correction method is implemented by one (or more) computer 100. Of course, the correction method can be performed to correct several images XAj from among the first N two-dimensional images XA of the N selective photopolymerization layers in the N superimposed planes H of the photopolymerizable ceramic material deposited, or all of the first N two-dimensional images XA of the N selective photopolymerization layers in the N superimposed planes H of the photopolymerizable ceramic material deposited.The manufacturing process for part P made of photopolymerized ceramic material using the 200 additive manufacturing machine includes the correction process.

[0059] The computer 100 may be or include one or more computers, one or more processors, one or more microprocessors, one or more control circuits, one or more servers, one or more machines, or other components. The computer 100 may have been programmed by a computer program, including code instructions for implementing the method, when implemented on this computer 100. The computer 100 may include one or more permanent memories 101, one or more random access memories 102. The computer 100 may include one or more physical data input interfaces 103, one or more physical data output interfaces 104. This or these physical data input interfaces 103 may be or include one or more computer keyboards, one or more computer mice, one or more physical data communication ports, one or more touch screens, or others.This physical data output interface(s) 104 may be or include one or more physical data communication ports, one or more screens, or other features. A computer program may be stored and executed on the computer 100 and may include code instructions which, when executed on it, implement all or part of the method according to the invention.

[0060] In [Fig.4], the computer 100 includes a first neural network 1.

[0061] During a step E2 of the process, which is subsequent to the first step E1 and which is called the second step E2 in [Fig. 2], the computer 100 trains the first neural network 1 on the first N two-dimensional XA images of the N photopolymerization slices, to learn to reconstruct the first N two-dimensional XA images into N second two-dimensional XA images. The computer 100 trains the first neural network 1 so that each of the N second two-dimensional XA images learns to reconstruct each of the first N two-dimensional XA images.

[0062] According to an embodiment of the invention, shown in Figures 4, 5, and 9, the computer 100 trains, during step E2, the first neural network 1 on the first N two-dimensional XA images of the N photopolymerization slices, which are applied to the input 111 of the first neural network 1. The output 132 of the first neural network 1 provides each of the N second two-dimensional XA images that are calculated by the first neural network 1 from each of the first N two-dimensional XA images present at the input 111.

[0063] One embodiment of this learning is described below. The computer 100 trains the first neural network 1 to minimize a first calculated gap E between each of the N second two-dimensional images x;' and respectively each of the first N two-dimensional images XA. The first neural network 1 can thus include a first estimator 14 of the first difference E calculated between each of the N second two-dimensional images x;' and respectively each of the first N two-dimensional images XA. During the training of the first neural network 1 by the computer 100, the first estimator 14 modifies the control parameters of the first neural network 1 to minimize the first difference E calculated between each of the N second two-dimensional images x;' and respectively each of the first N two-dimensional images XA.

[0064] During a step E3 of the process, which is subsequent or prior to the second step E2 and which is called the third step E3, in [Fig.2], the calculator 100 obtains N third images PR; of the material of test parts P' of photopolymerized ceramic material and having actual geometric characteristics prescribed in the N superimposed planes H;.

[0065] In [Fig.6], the calculator 100 includes a second neural network 2.

[0066] During a step E4 of the process, which is subsequent to the third step E3 and which is called the fourth step E4 in [Fig. 2], the computer 100 trains the second neural network 2 on the first N two-dimensional XA images of the N photopolymerization slices, to learn the N third PR images. According to one embodiment of the invention, the computer 100 trains the second neural network 2 from the first N two-dimensional XA images of the N photopolymerization slices, which are applied to the input 211 of the second neural network 2, to learn the N third PR images at the output 222 of the second neural network 2.

[0067] An embodiment of this learning E4 is described below. The output 222 of the second neural network 2 provides each of the N seventh images PR;”, which are calculated by the second neural network 2 from each of the first N two-dimensional XA images of the N photopolymerization slices present at the input 211. The computer 100 trains the second neural network 2 so that each of the N seventh images PR;” learns to reconstruct each of the N third images PR;. Thus, the computer 100 trains the second neural network 2 to minimize a third gap E’ ’ calculated between each of the N seventh images PR;’ ’ and each of the N third images PR;. The first neural network 1 can thus include a second estimator 24 of the third gap E” calculated between each of the N seventh images PR;” and each of the N third images PR;.Thus, the second neural network 2 learns to minimize the reconstruction error E”. Thus, the second neural network 2 learns the effect of the manufacturing process. additive. During the training of the second neural network 2 by the computer 100, the second estimator 24 modifies the control parameters of the second neural network 2 to minimize the third gap E” calculated between each of the N seventh images PR / ' and respectively each of the N third images PR;.

[0068] According to an embodiment of the invention, during a step E8 of the process, which is subsequent to the fourth step E4 in Figures 2 and 8, the computer 100 corrects, from at least one part 13 of the first network 1 of neurons having been trained and from the second network 2 of neurons having been trained, at least one of the first N two-dimensional images XA; setpoint of the N slices T; of photopolymerization in at least one determined plane H; among the N superimposed planes H; in at least a fifth corrected two-dimensional image Xopt (i) setpoint of the at least one determined slice T; of the photopolymerizable ceramic material in the at least one determined plane H;.

[0069] According to an embodiment of the invention, during a step E8 of the process, which is subsequent to the fourth step E4 in Figures 2 and 8, the computer 100 corrects, from the first network 1 of neurons having been trained and from the second network 2 of neurons having been trained, at least one of the first N two-dimensional images XA; setpoint of the N slices T; of photopolymerization in at least one determined plane H; among the N superimposed planes H; in at least a fifth corrected two-dimensional image Xopt (i) setpoint of the at least one determined slice T; of the photopolymerizable ceramic material in the at least one determined plane H;.

[0070] This at least one of the first N two-dimensional XA images that is corrected is called at least a fourth two-dimensional XA image of at least one photopolymerization slice T in at least one determined plane H among the N superimposed planes H. Thus, it can be provided as the fourth two-dimensional XA image of at least one photopolymerization slice T in at least one determined plane H; one or more or all of the first N two-dimensional XA images of the N photopolymerization slices T in one or more or all of the N determined planes H. Thus, it can be provided as one or more or N fourth two-dimensional XA images of one, several, or all of the N photopolymerization slices T in one, several, or all of the N determined planes H.Thus, one or more or N fifth two-dimensional corrected images Xopt (i) can be provided from one, several or all of the. the N slices T; of photopolymerization in one or more or all of the N determined plane H;.

[0071] During a step E10 of the manufacturing process, which is subsequent to step E8 in [Fig.2], the additive manufacturing machine 200 manufactures the part P from photopolymerized ceramic material by depositing and photopolymerizing the N slices T; of photopolymerization in the N superimposed planes H; of deposition of the ceramic material following the first N two-dimensional reference images XA; of which at least a fourth two-dimensional reference image x; two-dimensional reference image of at least one determined slice T; of the photopolymerizable ceramic material in at least one determined plane H; has been replaced by at least a fifth corrected two-dimensional reference image Xopt (i) of at least one determined slice T;) of the photopolymerizable ceramic material in at least one determined plane H;.

[0072] The invention makes it possible to manufacture the parts P on the platform 202 using the additive manufacturing machine 200, maximizing their production. The invention reduces manufacturing time. The invention also makes it possible to print small cavities using the additive manufacturing machine 200.

[0073] The method according to the invention thus comprises an algorithm for processing the sliced ​​images XAj, enabling the advance correction of these sliced ​​images XAj, which are then sent to the additive manufacturing machine 200 to produce the actual part P. The method and device according to the invention thus guarantee improved dimensional quality in the additive manufacturing of the part P by the additive manufacturing machine 200. The method and device according to the invention allow for the advance modification, in the fifth corrected two-dimensional image Xopt (i), of the sliced ​​image(s) XAj, which are then sent to the additive manufacturing machine 200 to produce the actual part P, and allows for the advance correction of irregularities such as, for example, rounded corners and lines that are wider or narrower than those of the reference image X; in the determined slice(s) T;.The method and device according to the invention make it possible to anticipate deviations in shape and size during additive manufacturing by applying corrections to the sliced ​​image(s) XAj. The method and device according to the invention allow for better resolution of the geometries of ceramic parts P. The method and device according to the invention make it possible to print small cavities using the additive manufacturing machine 200. The method and device according to the invention make it possible to obtain sub-pixel precision (less than the size of a pixel). The method and device according to the invention make it possible to minimize the impact of diffraction of the selective photopolymerization radiation R on the ceramic particles. during additive manufacturing. The process and device according to the invention make it possible to learn about physical phenomena such as the diffraction of the R radiation of selective photopolymerization on ceramic particles during additive manufacturing and its impact on the manufactured part P. Once manufactured by the additive manufacturing machine 200, the dimensional error of the manufactured part P minimizes a dimensional error with respect to the image x; of reference.

[0074] According to one embodiment of the invention, part 13 of the first neural network 1 is a downstream part 13 of the first neural network 1. Part 13 of the first neural network 1 is downstream of an upstream part 11 of the first neural network 1, in the direction from the input 111 of the first neural network 1 to the output 132 of the first neural network 1. According to one embodiment of the invention, part 13 of the first neural network 1 includes a first decoder 13 of the first neural network 1. According to one embodiment of the invention, the first neural network 1 further includes a first encoder 11, located upstream of the first decoder 13. According to one embodiment of the invention, the first neural network 1 is convolutional.The first encoder 11 compresses each of the first N two-dimensional XA images of the N photopolymerization slices, which are applied to the input 111 of the first encoder 11 of the first neural network 1, into a first latent image IL, which has a dimension (number of pixels) less than the dimension (number of pixels) of the first two-dimensional XA image of the setpoint and which is present on the output 112 of this first encoder 11. The first decoder 13 decompresses each of the first vectors z of latent variables applied to the input 131 of the first decoder 13 of the first neural network 1 into the second two-dimensional image x;', which has a dimension (number of pixels) greater than the dimension (number of components) of the first vector z of latent variables and which is applied to the output 132 of the first decoder 13 of the first neural network 1.The output 132 of the first decoder 13 of the first neural network 1 has the same dimensions as the input 111 of the first encoder 11 of the first neural network 1. According to one embodiment of the invention, the first neural network 1 comprises a first variational autoencoder 10. According to another embodiment of the invention, the first variational autoencoder 10 comprises, in cascade, the first encoder 11, a variational computation module 12, and the first decoder 13. The variational computation module 12 transforms each first latent image IL, which is present at the output 112 of the first encoder 11 of the first neural network 1, into the first vector z of latent variables, which is applied to the input 131 of the first decoder 13 of the first neural network 1. These embodiments can be combined with each other.

[0075] According to an embodiment of the invention, shown in Figures 4, 5, and 9, the computer 100 trains, during step E2, the first neural network 1 on the first N two-dimensional XA images of the N photopolymerization slices, which are applied to the input 111 of the first encoder 11 of the first neural network 1. The output 132 of the first decoder 13 of the first neural network 1 provides each of the N second two-dimensional XA images that are calculated by the first neural network 1 from each of the first N two-dimensional XA images present at the input 111. The N second two-dimensional XA images are calculated by the first neural network 1 to minimize the first deviation E calculated between the N second two-dimensional XA images and the first N two-dimensional XA images.

[0076] According to an embodiment of the invention, in Figures 4, 5 and 9, the first encoder 11 comprises different neurons j, each having variable weights Wj and variable biases bj. These neurons are arranged in convolutional layers, also called "feature maps". Each neuron j of the first encoder 11 calculates the output y of that neuron j by multiplying the input aj of that neuron j by the weight Wj of that neuron j and adding the bias bj of that neuron j, then applying the activation function L of the first network 1 of neurons, according to the equation yj = L(aj.Wj+bj).

[0077] The variable weights Wj and variable biases bj are the control parameters 113 of the first encoder 11. The activation function L of the first neural network 1 can be a linear rectification (known as ReLU), or a sigmoid, or a hyperbolic tangent, or other.

[0078] According to an embodiment of the invention, in Figures 4, 5, 7, 8 and 10, the first decoder 13 comprises different neurons k, each having variable weights Wk and variable biases bk. Each neuron k of the first decoder 13 calculates the output yk of that neuron k by multiplying the input ak of that neuron k by the weight Wk of that neuron k and adding the bias bk of that neuron k, then applying the activation function L of the first network 1 of neurons, according to the equation yk = L(ak.Wk+bk).

[0079] The variable weights Wk and variable biases bk are the control parameters 133 of the first decoder 13. The activation function L of the first neural network 1 can be a linear rectification (known as ReLU), or a sigmoid, or a hyperbolic tangent, or other.

[0080] According to one embodiment of the invention, during a first substep E31 of step E3 of the process, the additive manufacturing machine 200 produces the test parts P' having the photopolymerized ceramic material and having the The actual geometric characteristics prescribed. The test pieces P' are made of the same material as piece P. The P' pieces are parts of the same family as piece P and have the same degree of complexity.

[0081] According to a first approach, these test parts P' are portions of part P, these portions serving as representative test specimens containing the geometric characteristics of the intended final parts P. These test parts P' can be several separate parts with different geometries. These parts are smaller than part P, which allows them to be manufactured on the build platform 202 by the additive manufacturing machine 200, maximizing their production, for example, by simultaneously manufacturing several separate parts with different geometries, while reducing manufacturing time. These test parts P' can have experimental geometries in order to predict how these experimental geometries will be printed by the additive manufacturing machine 200. These test parts P' can be portions of a high-pressure turbine blade from an aeronautical turbomachine in the example mentioned above.

[0082] According to a second approach, these test parts P' correspond to the actual parts P to be manufactured, i.e. they can be high pressure turbine blade cores of aeronautical turbomachine in the example mentioned above of high pressure turbine blades P of aeronautical turbomachine.

[0083] A strong interaction exists between the different manufacturing parameters of the additive manufacturing machine 200 (for example: energy of the selective photopolymerization radiation R emitted by the projector 204, exposure time T, geometric compensation or others).

[0084] In a second substep E32 of step E3 of the process, which is subsequent to the first substep E31, as shown in [Fig. 2], three-dimensional tomographic representations XT0M of the test pieces P' are obtained by means of an imaging device 205, which may be, for example, a tomograph (or other). The imaging device 205 could use, instead of a tomograph, a non-contact measuring device for a non-moldable geometry that is observable without overlap.

[0085] During a third substep E33 of step E3 of the process, which is subsequent to the second substep E32 and during a fourth substep E34 of step E3, which is subsequent to the third substep E33, in [Fig.2], the computer 100 projects the three-dimensional tomographic representations XTom to acquire N third images PR; of the material of the test pieces P' respectively in the N superimposed planes H;.

[0086] For example, to do this, during the third substep E33, in [Fig.2], the calculator 100 calculates from the three-dimensional tomographic representations XTOm is a three-dimensional volume V3D of the material of the test parts P'. Then, during the fourth substep E34, in [Fig. 2], the computer 100 segments the three-dimensional volume V3D of the material of the test parts P' to isolate the material and highlight the regions of interest. During the fourth substep E34, in [Fig. 2], the computer 100 projects the three-dimensional volume V3D of the material of the test parts P' onto the N superimposed planes H; to acquire the N third images PR; of the material of the test parts P'.

[0087] The tomographic imaging device 205 is a radiography device that transmits radiation, for example, X-rays. Each three-dimensional tomographic representation XT0M is a projection image taken from different viewing angles of the test piece P'. Each three-dimensional tomographic representation XT0M consists of voxels and can be composed of several two-dimensional images taken in image planes traversing the test piece P' and distinct from one another (parallel or non-parallel). The material of the test piece P' is semi-transparent to the radiation of the radiography device, in this case, X-rays.

[0088] According to one embodiment of the invention, the second neural network 2 comprises a second autoencoder 20. According to one embodiment of the invention, the second autoencoder 20 comprises, in cascade, a second encoder 21 and a second decoder 22. According to one embodiment of the invention, the second neural network 2 is convolutional. The second encoder 21 compresses each of the first N two-dimensional XA images of the N photopolymerization slices, which are applied to the input 211 of the second encoder 21 of the second neural network 2, into a second latent image IL2, which has a dimension (number of pixels) smaller than the dimension (number of pixels) of the first two-dimensional XA image and which is present at the output 212 of this second encoder 21.The second decoder 22 decompresses each of the second latent image IL2, which is present on the output 212 of this second encoder 21 and on the input 221 of this second decoder 22, into a seventh image PRi”, which has a dimension (number of pixels) greater than the dimension (number of components) of the second latent image IL2 and which is applied to the output 222 of the second decoder 22 of the second neural network 2. In the case of the second convolutional neural network 2, the second neural network 2 comprises a chain of convolutional layers which may be composed, according to one embodiment: of a kernel ranging from 3x3 to 5x5 pixels; of an activation function (which may be a linear rectification (known as ReLU), or a sigmoid, or a hyperbolic tangent, or other); of a number of convolutional layers (number of . Convolution is performed using convolution, followed by pooling layers. Each pooling layer can be a mean, maximum, or other type of pooling layer, and is characterized by its pooling window size and pooling function (maximum, minimum, mean, or other). Convolutional layers are used to apply filters to the image, and pooling layers are used to reduce its dimensionality.

[0089] According to an embodiment of the invention, in Figures 6, 7 and 11, the second encoder 21 comprises different neurons 1, each having variable weights Wi and variable biases bi. Each neuron 1 of the second encoder 21 calculates the output yi of that neuron 1 by multiplying the input ai of that neuron 1 by the weight Wi of that neuron 1 and adding the bias bi of that neuron 1, then applying the activation function L of the first network 1 of neurons, according to the equation yi = L(ai.Wi+bi).

[0090] The variable weights Wi and bi-variable biases are the control parameters 213 of the second encoder 21. The activation function L of the first neural network 1 can be a linear rectification (known as ReLU), or a sigmoid, or a hyperbolic tangent, or other.

[0091] According to an embodiment of the invention, in Figures 6, 7 and 12, the second decoder 22 comprises different neurons m, each having variable weights Wm and variable biases bm. Each neuron m of the second decoder 22 calculates the output ym of that neuron m by multiplying the input am of that neuron m by the weight Wm of that neuron m and adding the bias bm of that neuron m, then applying the activation function L of the first network 1 of neurons, according to the equation ym = L(am.Wm+bm).

[0092] The variable weights Wm and variable biases bm are the control parameters 223 of the second decoder 22. The activation function L of the first neural network 1 can be a linear rectification (known as ReLU), or a sigmoid, or a hyperbolic tangent, or other.

[0093] During a step E5 of the process, which is subsequent to the fourth step E4 and which is called the fifth step E5 in Figures 2 and 7, the computer 100 forms a third neural network 3 formed from the first neural network 1 trained during step E2 and from the second neural network 2 trained during step E2. According to one embodiment of the invention, the computer 100 forms during the fourth step E4 the third neural network 3 comprising in cascade the first decoder 13 of the first neural network 1 trained during step E2 and the second neural network 2 trained during step E2.

[0094] According to one embodiment of the invention, the computer 100 forms, during step E5 following step E4 and referred to as the fifth step E5 in Figures 2 and 7, the third neural network 3 comprising in cascade the part 13 of the first neural network 1 trained during step E2 and the second neural network 2 trained during step E4. In this third neural network 3, the second neural network 2 may comprise in cascade the second encoder 21 trained during step E4 and the second decoder 22 trained during step E4.According to one embodiment of the invention, the computer 100 forms, during step E5, the third neural network 3 comprising in cascade the variational calculus module 12, the first decoder 13 having been trained during step E2, and the second neural network 2 having been trained during step E4, and comprising in cascade the second encoder 21 having been trained during step E4 and the second decoder 22 having been trained during step E4. In the third neural network 3, the output 132 of the first neural network 1 having been trained during step E2, or the output 132 of the first decoder 13 of the first neural network 1 having been trained during step E2, is connected to the input 211 of the second neural network 2 having been trained during step E4.

[0095] According to one embodiment of the invention, during a step E6 of the process, which is subsequent to the fifth step E5 and which is referred to as the sixth step E6 in Figures 2 and 7, the computer 100 prescribes at least one (or more or all) fourth two-dimensional reference image x of at least one (or more or all) determined slice T of the photopolymerizable ceramic material in at least one (or more or all) determined plane H among the N superimposed planes H. The fourth two-dimensional reference image x may have been pre-recorded in the memory 101 of the computer 100. The fourth two-dimensional reference image x may be an image obtained by computer-aided design or other means. The fourth two-dimensional reference image x is a reference image.

[0096] According to an embodiment of the invention, the computer 100 optimizes during step E6 respectively at least one random vector V; sent to the third neural network 3, minimizing a second gap A; calculated between the at least one fourth two-dimensional setpoint image x; of the at least one determined slice T; of the photopolymerizable ceramic material in the at least one determined plane H; and a sixth image PRi', which was obtained by the third neural network 3 from the at least one random vector Vi, to obtain at least one optimized random vector Vopt (i).

[0097] According to one embodiment of the invention, the computer 100 applies the random vector V; to the input 121 of the third neural network 3 during step E6. The output 222 of the third neural network 3 provides the sixth image PR;', which was calculated by the third neural network 3 from the random vector V; present at the input 121 of the third neural network 3. The computer 100 optimizes, for each fourth two-dimensional image x; of the setpoint in the determined plane H; respectively, the random vector V; sent to the input 121 of the third neural network 3.

[0098] According to one embodiment of the invention, the computer 100 applies the random vector V; to the input 121 of the variational computation module 12 of the third neural network 3 during step E6. The output 222 of the second encoder 21 of the third neural network 3 provides the sixth image PR;', which was calculated by the third neural network 3 from the random vector V; present at the input 121 of the variational computation module 12 of the third neural network 3. The computer 100 optimizes, for each fourth two-dimensional input image x; in the determined plane H; respectively, the random vector V; sent to the input 121 of the variational computation module 12 of the third neural network 3.

[0099] According to one embodiment of the invention, the computer 100 optimizes, during step E6, the random vector V by minimizing the second error A, calculated between the fourth two-dimensional image x of the determined slice T of the photopolymerizable ceramic material in the determined plane H and the sixth image PR, thus obtaining an optimized random vector Vopt(i) at the input 121 of the third neural network 3. The third neural network 3 may thus include a third estimator 34 of the second error A, calculated between the fourth two-dimensional image xi of the determined slice T of the photopolymerizable ceramic material in the determined plane H and the sixth image PR. The sixth image PR is an image obtained by the third neural network 3.During optimization, the third estimator 34 modifies the random vector V, sent to input 121 of the third neural network 3, to minimize the second gap A, calculated between the fourth two-dimensional setpoint image x of the determined slice T of the photopolymerizable ceramic material in the determined plane H and the sixth image PR, and to determine the optimized random vector Vopt(i). This results in an optimization loop on the vector V. The calculator 100 can try several random vectors V for each fourth two-dimensional setpoint image x in the determined plane H. One, several, or all of the optimized random vectors Vopt(i) can be determined by the calculator 100 for one, several, or all of the fourth two-dimensional setpoint images x, i.e., for one of the or . several or all of the slices T; of the photopolymerizable ceramic material in one or more or all of the N superimposed planes H;. In the embodiment described above, this or these random vectors V; for each fourth two-dimensional image x; of setpoint in the determined plane H;, are sampled from the first latent image IL and the first vector z of latent variables.

[0100] During a step E7 of the process, which is subsequent to the second step E2 or the sixth step E6 and which is referred to as the seventh step E7 in Figures 2 and 8, the computer 100 forms a fourth neural network 4 from the first neural network 1 that was trained during step E2. The fourth neural network 4 comprises part 13 of the first neural network 1 that was trained. According to one embodiment of the invention, the computer 100 forms, during step E7, the fourth neural network 4 comprising the first decoder 13 that was trained during step E2.

[0101] According to one embodiment of the invention, the computer 100 forms during step E7 the fourth neural network 4 comprising in cascade the variational calculation module 12 and the first decoder 13 having been trained during step E2.

[0102] According to an embodiment of the invention, during step E8 of the process, which is subsequent to the seventh step E7 and which is called the eighth step E8 in Figures 2 and 8, the computer 100 applies at least one optimized random vector Vopt (i) to the fourth neural network 4, to obtain at least one fifth corrected two-dimensional image Xopt (i) of the control setpoint of at least one determined slice T; of the photopolymerizable ceramic material in at least one determined plane H;.

[0103] According to one embodiment of the invention, the calculator 100 applies the optimized random vector Vopt( ; , to the input 121 of the variational calculation module 12 of the fourth neural network 4. The fourth neural network 4 calculates on the output 132 of the first decoder 13 of the fourth neural network 3, from the optimized random vector (or vectors) Vopt( ) present at the input 121 of the variational calculation module 12 of the fourth neural network 4, the fifth corrected two-dimensional image Xopt( i ) of the determined slice (or slices) T; of the photopolymerizable ceramic material in the determined plane (or planes) H;.The fifth corrected two-dimensional image Xopt(i) of setpoint can be computed by the fourth neural network 4 for one, several, or all of the optimized random vectors Vopt(i), i.e., for one or more or all of the slices T; of the photopolymerizable ceramic material in one or more or all of the N superimposed planes H;. .

[0104] During a step E9 of the process, which is subsequent to the eighth step E8 and which is called the ninth step E9 in [Fig.2] or during the eighth step E8, the computer 100 replaces the image XAj to be corrected, i.e. the at least a fourth two-dimensional image x; of the photopolymerization in the at least one determined plane H; among the N superimposed planes H;, with the at least a fifth corrected two-dimensional image Xopt (i) of the photopolymerizable ceramic material in the at least one determined plane H;. .One, several, or all of the fifth corrected two-dimensional images Xopt(i) of the setpoint can be determined by the calculator 100 for one, several, or all of the fourth two-dimensional images x of the setpoint, that is, for one, several, or all of the slices T of the photopolymerizable ceramic material in one, several, or all of the N superimposed planes H. Steps E2 to E9 thus form a method for correcting at least one image XAj to be corrected by at least one fifth corrected two-dimensional image Xopt(i) of the setpoint of the slice T of the photopolymerizable ceramic material in the determined plane H.

[0105] According to an embodiment of the invention, in Figures 4 and 5, during training in the second step E2 of the first neural network 1, the first encoder 11 learns a first multivariate normal distribution N(p, o2)) of each of the first N two-dimensional control XA images of each of the N photopolymerization slices applied to the input 111 of the first encoder 11 of the first neural network 1.

[0106] According to one embodiment of the invention, the second deviation A; is equal to a structural similarity index measure SSIM calculated respectively between the fourth two-dimensional setpoint image x; of the determined slice T; of the photopolymerizable ceramic material in the determined plane H; and the sixth image PR;', which was obtained at the output 222 of the second encoder 21 of the third neural network 3 from the random vector V;.

[0107] According to one embodiment of the invention, the optimized random vector Vopt is equal to

[0108] Vopt ( ; ).= argmin f(V;) = SSIM(Xi - DAB(EAB(VDA(z)))).

[0109] According to one embodiment of the invention, the calculated structural similarity index measure SSIM is equal to SSIM(x, xb) according to the equation below:

[0110] where

[0111] px is a calculated average of the pixels of the fourth two-dimensional image x; of the determined slice T; of the photopolymerizable ceramic material in the determined plane H;,

[0112] ox is a calculated standard deviation of the pixels of the fourth two-dimensional image x; of the determined slice T; of the photopolymerizable ceramic material in the determined plane H;,

[0113] pxb is a calculated average of the pixels of the sixth image PR;',

[0114] oxb is a calculated standard deviation of the pixels of the sixth image PR;',

[0115] covxbx is the calculated covariance between the pixels of the fourth image x; two-dimensional setpoint of the determined slice T; of the photopolymerizable ceramic material in the determined plane H; and the pixels of the sixth image PR;', Ci = (fc1£) 2 ,c2 = (M) 2 and c3 = y

[0116] Ci, c2 and c3 are prescribed constants, L being the maximum value of a pixel. By default, ki=0.01 and k2=0.03.

[0117] According to another embodiment of the invention, the second deviation A; is equal to a Sprensen-Dice index s calculated respectively between the fourth two-dimensional setpoint image Xi of the determined slice T; of the photopolymerizable ceramic material in the determined plane H; and the sixth image PR;', which was obtained at the output 222 of the second encoder 21 of the third neural network 3 from the random vector V;.

[0118] According to one embodiment of the invention, the calculated Sprensen-Dice index s is equal to s according to the equation below:

[0119] For arbitrary finite sets X and Y, the index s is expressed by

[0120] v ™ 5 “ IXWYl

[0121] Here, IXI denotes the number of elements of X. Here, IXI denotes the number of elements of Y. Here, IX A Fl denotes the number of elements of XA F.

[0122] For example, the index s is equal to:

[0123] 2xE;£ 1(^*0 and PR^) s~

[0124] According to another embodiment of the invention, the second deviation A; is equal to a Jaccard index J calculated respectively between the fourth two-dimensional reference image Xi of the determined slice T; of the photopolymerizable ceramic material in the determined plane H; and the sixth image PR;', which was obtained at output 222 of the second encoder 21 of the third neural network 3 from the random vector (or vectors) V;.

[0125] According to one embodiment of the invention, the calculated Jaccard index s is equal to J(A, B) according to the equation below:

[0126] B) =

[0127] According to one embodiment of the invention, the first gap E is a first binary cross-entropy H(XA;, x;') calculated between each second two-dimensional image x;' and each first two-dimensional image XA;. Similarly, the third gap E” is a third binary cross-entropy H(PRi”, PRi) calculated between each seventh image PRi” and each third image PRi.

[0128] According to one embodiment of the invention, the first binary cross entropy H(XAi, xi') calculated between each second two-dimensional image xi' and each first two-dimensional image XAi; as well as the third binary cross entropy H(PRi”, PRi) calculated between each seventh image PRi” and each third image PRi are obtained according to the following principle.

[0129] By definition, the binary cross entropy H(X, Y) between an image X and an image Y measures the difference between the probability distribution of each pixel in the image X and the image Y. To obtain a probability value, the value of each pixel in the images X and Y is normalized by dividing the value of each pixel by the sum of each pixel in the respective images.

[0131] Where n = number of pixels in the image

[0132] : probability associated with the value of the X image

[0133] : probability associated with the value of the image Y. This formula and this definition for H(X, Y) are applied to calculate the first binary cross entropy H(XAi, xi') and the third binary cross entropy H(PRi”, PRi), replacing X and Y with the corresponding images mentioned above.

[0134] Of course, the embodiments, features, possibilities and examples described above can be combined with each other or selected independently of each other.

Claims

1. Demands A method for manufacturing a part (P) made of photopolymerized ceramic material using an additive manufacturing machine (200), which receives (E1) N first two-dimensional geometric setpoint images (XA;) of N slices (TO) of selective photopolymerization in N superimposed planes (H;) of deposition of a photopolymerizable ceramic material, characterized in that the method comprises the following steps, implemented by at least one computer (100): training (E2) a first neural network (1) on the first N two-dimensional setpoint images (XA;) of the N photopolymerization slices (T;), to learn to reconstruct the first N two-dimensional images (XAO) into N second two-dimensional images (x;'), to minimize a first deviation (E) calculated between the N second images (x;') two-dimensional and the first N images (two-dimensional XAO, obtaining (E3) N third images (PRi) of the material of test pieces (P') of photopolymerized ceramic material and having real geometric characteristics prescribed in the N superimposed planes (H;), training (E4) of a second network (2) of neurons on the first N images (two-dimensional XAO of instruction of the N slices (T;) of photopolymerization, to learn the N third images (PR;), correction, from at least a part (13) of the first network (1) of neurons having been trained and from the second network (2) of neurons having been trained, of at least one of the first N two-dimensional images (XA;) of instruction of the N slices (T;) of photopolymerization, called fourth image (x;), in at least one determined plane (H;) among the N superimposed planes (H;) in at least one fifth corrected two-dimensional image (Xopt (i)) of the setpoint of at least one determined slice (Tj) of the photopolymerizable ceramic material in at least one determined plane (H;), the additive manufacturing machine (200) manufacturing (E10) the part (P) of photopolymerized ceramic material by depositing and photopolymerizing the N photopolymerization slices (T;) in; the N superimposed planes (H;) of deposition of the ceramic material following the first N two-dimensional setpoint images (XA;), of which at least a fourth two-dimensional setpoint image (x;) of at least a determined slice (T;) of the photopolymerizable ceramic material in the at least a determined plane (H;) has been replaced by at least a fifth corrected two-dimensional setpoint image (Xopt (i)) of at least a determined slice (T;) of the photopolymerizable ceramic material in the at least a determined plane (H;).

2. The method according to claim 1, characterized in that it further comprises, for performing the correction: the training (E5) of a third neural network (3), comprising in cascade the part (13) of the first neural network (1) having been trained and the second neural network (2) having been trained, for said at least a fourth two-dimensional setpoint image (x;) of at least a determined slice (Tj) of the photopolymerizable ceramic material in at least one determined plane (H;) among the N superimposed planes (H;), the optimization (E6) respectively of at least one random vector (Vi) sent to the third neural network (3), by minimizing a second deviation (A;) calculated between the at least a fourth two-dimensional setpoint image (Xj) of at least a determined slice (T;) of the photopolymerizable ceramic material in at least one determined plane (H;) and a sixth image (PR;'), which was obtained by the third network (3) of neurons from at least one random vector (V;), to obtain at least one optimized random vector (Vopt (i)), the training (E7) of a fourth network (4) of neurons comprising the part (13) of the first network (1) of neurons having been trained, the application of at least one optimized random vector (Vopt (i)) to the fourth network (4) of neurons, to obtain at least one fifth corrected two-dimensional image (Xopt (i)) of the setpoint of at least one determined slice (Tj) of the photopolymerizable ceramic material in at least one determined plane (H;).;

3. A method according to any one of the preceding claims, characterized in that it further comprises the manufacture (E3, E31), by the additive manufacturing machine (200), of test pieces (P') of photopolymerized ceramic material and having the prescribed real geometric characteristics, the obtaining (E3, E32) by tomography of three-dimensional tomographic representations (XT0M) of the test pieces (P'), the projection (E3, E33, E34) of the three-dimensional tomographic representations (XTOm) in the N superimposed planes (H;) to obtain the N third images (PR;) of the material of the test pieces (P').

4. A method according to any one of the preceding claims, characterized in that part (13) of the first neural network (1) comprises a first decoder (13) of the first neural network (1).

5. Method according to claim 4, characterized in that the first neural network (1) further comprises a first encoder (11), located upstream of the first decoder (13).

6. A method according to claim 5, characterized in that the training (E2) of a first neural network (1) is carried out on the first N two-dimensional setpoint images (XA;) of the N photopolymerization slices (Tj) applied to an input (111) of the first encoder (11) of the first neural network (1), to learn to reconstruct the first N two-dimensional images (XA;) into the N second two-dimensional images (x;'), which are calculated by the first neural network (1) from the first N two-dimensional images (XA;) and which are present on an output (132) of the first decoder (13) of the first neural network (1), the N second two-dimensional images (x;') being calculated by the first neural network (1) to minimize the first gap (E) calculated between the N second two-dimensional images (x;') and the first N two-dimensional images (XAj).

7. A method according to any one of the preceding claims, characterized in that the first gap (E) is a first binary cross entropy (H(XA;), X;')) calculated between the N second two-dimensional images (x;') and the first N two-dimensional images (XA;).

8. A method according to any one of the preceding claims, characterized in that the second neural network (2) comprises a second auto-encoder (20), which comprises in cascade a second encoder (21) and a second decoder (22).

9. Method according to claim 8, characterized in that the training (E4) of the second neural network (2) is carried out on the first N two-dimensional setpoint images (XA;) of the N photopolymerization slices (T;) applied to an input (211) of the second encoder (21) of the second neural network (2), to learn the N third images (PR,) at an output (222) of the second encoder (21) of the second neural network (2).

10. A method according to any one of the preceding claims, when they depend at least on claim 2, characterized in that the second gap (Ai) is equal to a structural similarity index (SSIM) measure calculated between the at least a fourth two-dimensional setpoint image (x;) of the at least a determined slice (Tj) of the photopolymerizable ceramic material in the at least a determined plane (H;) and the sixth image (PR;'), which was obtained at an output (222) of the third neural network (3) from the at least a random vector (V;).

11. A method according to any one of the preceding claims, when they depend at least on claim 2, characterized in that the second gap (Ai) is equal to a Sprensen-Dice index calculated between the at least a fourth two-dimensional setpoint image (x;) of the at least a determined slice (Tj) of the photopolymerizable ceramic material in the at least a determined plane (H;) and the sixth image (PR;'), which was obtained at an output (222) of the third neural network (3) from the at least a random vector (V;).

12. A method according to any one of the preceding claims, when they depend at least on claim 2, characterized in that the second gap (Ai) is equal to a Jaccard index calculated between the at least a fourth two-dimensional setpoint image (x;) of the at least a determined slice (Tj) of the photopolymerizable ceramic material in the at least a determined plane (H;) and the sixth image (PR;'), which was obtained at an output (222) of the third neural network (3) from the at least a random vector (V;).

13. A method according to any one of the preceding claims, characterized in that the training (E4) of the second neural network (2) is carried out so that from the first N two-dimensional setpoint images (XA;) of the N photopolymerization slices (Tj) the second neural network (2) calculates N seventh images (PR;”), which are present on an output (222) of the second neural network (2) and which minimize a third gap (E”) calculated between the N seventh images (PR;”) and the N third images (PRi).

14. A method according to claim 13, characterized in that the third gap (E”) is a third binary cross entropy (H(PR i”, PR;)) calculated between the N seventh images (PR;”) and the N third images (PR;).

15. Additive manufacturing machine (200) for a part (P) made of photopolymerized ceramic material, comprising at least one computer (100) for implementing the manufacturing process according to any one of the preceding claims.

Images