Learning method of a learning device, design method of a design pattern, and manufacturing method of a laminate

Abstract

To easily derive an appropriate lamination pattern in consideration of a constraint condition.SOLUTION: A learning method of a learning device allows a learning device to learn a retrieval model for retrieving a lamination pattern of a laminate formed by laminating fiber sheets as a one direction material with one fiber orientation direction using the fiber sheets, wherein the retrieval model includes a constraint condition relating on lamination of the fiber sheets and is a learning model using a measure function and a value function, and allows the learning device to execute the steps of acquiring an initial lamination pattern that is the lamination pattern in an initial state, and allowing the learning device to learn the retrieval model with the initial lamination pattern as an input so that the initial lamination pattern becomes the lamination pattern satisfying the constraint condition.SELECTED DRAWING: Figure 5

Description

Technical Field

[0001] The present disclosure relates to a learning method for a learning device, a design method for a design pattern, a manufacturing method for a laminate, and a design device for a design pattern.

Background Art

[0002] Conventionally, a design method for a composite material laminate structure for designing a laminate structure of a composite material laminate, which is a laminate, has been known (see, for example, Patent Document 1). In this design method, using a relational expression as a prediction model for predicting a laminate structure from physical property values, the laminate structure is calculated with the physical property values as inputs.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] In the laminate structure, there are constraint conditions such as the lamination order in the lamination direction. Examples of the constraint conditions include a condition that the difference in the orientation angle formed by the fiber orientation directions of two adjacent layers is 45° or less. In the design method of Patent Document 1, based on physical property values, using a relational expression, it is said that laminate structure information, which is information on a laminate structure considering the above constraint conditions, is calculated.

[0005] However, in Patent Document 1, when using physical property values as input values, if an attempt is made to calculate laminate structure information considering the above constraint conditions, the solution cannot be predicted, and there is a possibility that the laminate structure information cannot be obtained. Further, in Patent Document 1, even when the solution converges, there is a possibility that the calculation time will be long, and as the number of laminates increases, the calculation time increases, making it difficult to suppress an increase in calculation cost.

[0006] Therefore, an object of the present disclosure is to provide a learning method for a learning device, a design method for a design pattern, a manufacturing method for a laminate, and a design device for a design pattern, which can easily derive an appropriate lamination pattern considering constraint conditions.

Means for Solving the Problem

[0007] A learning method for a learning device according to the present disclosure is a learning method for a learning device in which a learning device learns a search model for searching for a lamination pattern of a laminate formed by laminating a fiber sheet as a unidirectional material in which the fiber orientation direction is unidirectional. The search model includes constraint conditions related to the lamination of the fiber sheet and is a learning model using a policy function and a value function. The learning device is caused to execute a step of acquiring an initial lamination pattern that is the lamination pattern in the initial state, and a step of learning the search model so that the lamination pattern satisfies the constraint conditions with the initial lamination pattern as an input.

[0008] A design method for a design pattern according to the present disclosure is a design method for a design pattern in which a design pattern that is a lamination pattern of the laminate satisfying the constraint conditions is designed using the search model learned by the learning method of the learning device described above, using a design device. The design device is caused to execute a step of deriving a candidate pattern that is a candidate lamination pattern of the laminate by a predetermined algorithm, and a step of inputting the candidate pattern of the laminate into the search model and deriving a design pattern that is a lamination pattern satisfying the constraint conditions.

[0009] A manufacturing method for a laminate according to the present disclosure includes a step of laminating the fiber sheet based on the design pattern designed by the design method for the design pattern described above, and a step of integrating the laminated fiber sheets to form the laminate.

[0010] The design device of the design pattern of the present disclosure is a design device for designing a design pattern that is a stacking pattern of the laminate satisfying the above constraints, using the search model learned by the learning method of the above learning device, and includes a step of deriving a candidate pattern that is a candidate stacking pattern of the laminate by a predetermined algorithm, and a step of inputting the candidate pattern of the laminate into the search model and deriving a design pattern that is the stacking pattern satisfying the above constraints, and is provided with a control unit that executes the steps.

Advantages of the Invention

[0011] According to the present disclosure, an appropriate stacking pattern considering the constraints can be derived quickly.

Brief Description of the Drawings

[0012]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Modes for Carrying Out the Invention

[0013] Hereinafter, embodiments according to the present disclosure will be described in detail with reference to the drawings. Note that the present invention is not limited by these embodiments. In addition, the constituent elements in the following embodiments include those that can be replaced and are easy for those skilled in the art, or those that are substantially the same. Furthermore, the constituent elements described below can be combined as appropriate, and when there are multiple embodiments, it is also possible to combine each embodiment.

[0014] [This embodiment] The learning method of the learning device, the design method of the design pattern, the manufacturing method of the laminate, and the design device of the design pattern according to this embodiment relate to the lamination pattern P of the laminate 1 formed by laminating fiber sheets S. FIG. 1 is an explanatory diagram regarding the lamination pattern. FIG. 2 is a diagram regarding the design device of the design pattern according to this embodiment. FIG. 3 is an explanatory diagram regarding the search model. FIG. 4 is a flowchart regarding the learning method of the design device according to this embodiment. FIG. 5 is a flowchart regarding the learning of the search model. FIG. 6 is a flowchart regarding the design method of the design pattern according to this embodiment. FIG. 7 is a flowchart regarding the manufacturing method of the laminate according to this embodiment.

[0015] (Laminate) In FIG. 1, the upper figure shows the lamination pattern P of the laminate 1 that does not satisfy the constraint conditions, and the lower figure shows the lamination pattern P of the laminate 1 that satisfies the constraint conditions. As shown in FIG. 1, the laminate 1 is formed by laminating a plurality of fiber sheets S in the lamination direction and integrally molding them. The fiber sheet S is, for example, a prepreg in which reinforcing fibers are impregnated with resin, and is a unidirectional material in which the fiber orientation direction is one direction. Note that the fiber sheet S is not particularly limited to a prepreg, and may be a dry reinforcing fiber sheet that has not been impregnated with resin. The laminate 1 has a predetermined lamination pattern P in the lamination direction. The lamination pattern P is the arrangement order in the lamination direction of the orientation angles formed by the reference direction and the orientation direction of each fiber sheet S.

[0016] In the upper diagram of FIG. 1, the lamination pattern P of the laminate 1 has an orientation angle of (0, 45, 90, 0, 45, 90, -45, 45, 0, 0, 0, 0). Here, constraints may be imposed on the lamination pattern P according to the required characteristics. The constraints include a first condition that the difference in the orientation angles of adjacent fiber sheets S in the lamination direction is 45° or less, and a second condition that the continuity of the fiber sheets S having the same orientation direction in the lamination direction is 3 layers or less. In the upper diagram of FIG. 1, in the adjacent fiber sheets S surrounded by the dotted line C1, the first condition is not satisfied, and in the aligned fiber sheets S surrounded by the dotted line C2, the second condition is not satisfied. On the other hand, in the lower diagram of FIG. 1, the lamination pattern P of the laminate 1 has an orientation angle of (0, 0, 45, 90, 45, 90, -45, 0, 45, 0, 0, 0), and satisfies the above constraints. Note that the first condition of the constraints is a condition that the difference in the orientation angles is 45° or less, but the orientation angle is not particularly limited and may be any orientation angle. Also, the second condition of the constraints is a condition that the continuity of the fiber sheets S is 3 layers or less, but the number of layers is not particularly limited and may be any number of layers. Furthermore, the constraints may be conditions according to the characteristics required for the laminate 1 and are not particularly limited to the first condition and the second condition.

[0017] (Design device for design pattern) The design device 10 for the design pattern is a device that designs a design pattern that is the lamination pattern P of the laminate 1 that satisfies the constraints. Also, the design device 10 functions as a learning device for learning a search model M for searching for the lamination pattern P. In this embodiment, the design device 10 and the learning device are integrated, but are not particularly limited and may be configured separately.

[0018] As shown in FIG. 2, the design device 10 includes a control unit 15 and a storage unit 16.

[0019] The storage unit 16 stores programs and data. Further, the storage unit 16 may also function as a work area for temporarily storing the processing results of the control unit 15. The storage unit 16 may include any storage device such as a semiconductor storage device and a magnetic storage device.

[0020] The storage unit 16 includes, as a program, a search model M. The search model M is a learning model for searching for the stacking pattern P of the laminate 1. The search model M is a learning model that combines Monte Carlo tree search and deep reinforcement learning and includes the above-described constraint conditions. In the search model M, a policy function and a value function are used, and the stacking pattern P that satisfies the constraint conditions is searched by swapping the orientation angles of the adjacent fiber sheets S in the stacking direction. Further, this search model M is a learning model that can be learned without teacher data. The storage unit 16 stores, as data, stacking data D1 related to the stacking pattern P and constraint data D2 related to the constraint conditions. The stacking data D1 is information related to the stacking pattern P shown in FIG. 1 and includes the stacking data D1 used as an input and the stacking data D1 obtained as an output. The constraint data D2 is information related to the above-described first condition and second condition.

[0021] The control unit 15 executes programs and exchanges data with the storage unit 16. The control unit 15 includes, for example, an integrated circuit such as a CPU (Central Processing Unit). Specifically, the control unit 15 executes learning of the search model M using the stacking data D1 and the constraint data D2 stored in the storage unit 16. Further, the control unit 15 derives the stacking pattern P of the laminate 1 that satisfies the constraint conditions using the learned search model M.

[0022] Here, referring to FIG. 3, the derivation of the lamination pattern P of the laminate 1 that satisfies the constraint conditions using the search model M by the control unit 15 will be described. The upper lamination pattern P in FIG. 3 corresponds to the upper lamination pattern P in FIG. 1 and is a lamination pattern P that does not satisfy the constraint conditions. The lower lamination pattern P in FIG. 3 corresponds to the lower lamination pattern P in FIG. 1 and is a lamination pattern P that satisfies the constraint conditions. In FIG. 3, when a lamination pattern P that does not satisfy the constraint conditions is input, the orientation angles of the fiber sheets S adjacent in the lamination direction are swapped by the search model M that combines Monte Carlo tree search and deep reinforcement learning. In FIG. 3, a learned search model M is used. Specifically, the learned policy function and value function included in the search model M are used. Therefore, the search model M performs a search regarding the swapping of the orientation angles of the fiber sheets S, and selects the swapping of the orientation angles of the fiber sheets S that results in a policy with a high reward from the search results of the performed search. Then, when the search model M satisfies the constraint conditions by swapping the orientation angles of the fiber sheets S, it derives the lamination pattern P that satisfies the constraint conditions as a solution.

[0023] As shown in FIG. 3, the number of laminations of the lamination pattern P used in the search model M is, for example, 12 layers. In the present embodiment, the number of laminations of the lamination pattern P is less than the number of laminations of the laminate 1. Note that the number of laminations of the lamination pattern P may be the same as the number of laminations of the laminate 1.

[0024] (Learning method of the design device) Next, referring to FIGS. 4 and 5, a learning method of the design device 10 in which the above-described design device 10 learns the search model M will be described.

[0025] As shown in FIG. 4, in the learning method of the design device (learning device), first, the control unit 15 of the design device 10 executes a step of acquiring an initial stacking pattern which is a stacking pattern P in the initial state (step S11). As the initial stacking pattern, for example, it is a stacking pattern that satisfies the constraint conditions. Subsequently, the control unit 15 executes step S12 of training the search model M using the initial stacking pattern. That is, in step S12, since the search model M is a model that can be trained without teacher data, the training of the search model M can be executed using the initial stacking pattern. After this, the control unit 15 evaluates whether the performance of the trained search model M is satisfactory (step S13). In step S13, the performance of the search model M is evaluated by inputting the stacking pattern P serving as the evaluation input to the trained search model M and evaluating whether the output stacking pattern P satisfies the constraint conditions. In step S13, when the control unit 15 determines that the performance of the search model M is satisfactory (step S13: Yes), the evaluation of the search model M ends. On the other hand, in step S13, when the control unit 15 determines that the performance of the search model M is not satisfactory (step S13: No), it proceeds to step S12 and executes the training of the search model M again.

[0026] Next, with reference to FIG. 5, step S12 regarding the learning of the search model M will be specifically described. In the learning of the search model M, when the control unit 15 acquires an initial laminated pattern that satisfies the constraint conditions in step S11, it swaps the layers of a part of the fiber sheets S in the initial laminated pattern to generate a plurality of initial laminated patterns that do not satisfy the constraint conditions (step S21). In step S21, the layers of adjacent fiber sheets may be swapped, or the layers of the fiber sheets may be randomly swapped, and it is not particularly limited. Subsequently, the control unit 15 selects one initial laminated pattern as an input from among the plurality of initial laminated patterns that do not satisfy the constraint conditions (step S22). The control unit 15 sets the input initial laminated pattern as the state of the laminated pattern P (step S23). Subsequently, the control unit 15 determines whether or not a predetermined number of episode steps has been reached (step S24). Here, the number of episode steps is a series of flows from step S26 to step S28 described later and then back to step S23 starting from step S23, and is the number of repetitions of this episode step. And the predetermined number of episode steps is a numerically defined value in advance. When the control unit 15 determines that the predetermined number of episode steps has been reached (step S24: Yes), it proceeds to step S29 described later. On the other hand, when the control unit 15 determines that the predetermined number of episode steps has not been reached (step S24: No), it proceeds to step S25.

[0027] In step S25, the control unit 15 determines whether the laminated pattern P set as the state in step S23 satisfies the constraint conditions (step S25). If the control unit 15 determines that the laminated pattern P satisfies the constraint conditions (step S25: Yes), it proceeds to step S29, which will be described later. On the other hand, if the control unit 15 determines that the laminated pattern P does not satisfy the constraint conditions (step S25: No), it uses the search model M to execute a search based on the laminated pattern P set as the state in step S23 (step S26). In step S26, in the laminated pattern P, the search is executed by swapping the layers of adjacent fiber sheets S. Then, the control unit 15 obtains the search result, which is the result of the executed search, in step S26. After that, the control unit 15 determines whether the number of searches has reached a predetermined number (step S27). If the control unit 15 determines that the number of searches has not reached the predetermined number (step S27: No), it returns to step S26 to re-execute the search. In this way, step S26 is repeatedly executed until the predetermined number of searches is reached.

[0028] In step S27, if the control unit 15 determines that the number of searches has reached the predetermined number (step S27: Yes), it selects and executes an action of selecting one laminated pattern P after swapping using the information of a plurality of search results corresponding to the number of searches (step S28), and then proceeds to step S23. In step S23, the control unit 15 sets the state of the laminated pattern P selected in step S28. Note that by executing the action in step S28, the control unit 15 increments the number of episode steps (increases by “+1”).

[0029] In step S24, if the control unit 15 determines that the number of steps of a predetermined episode has been reached, or in step S25, if the control unit 15 determines that the stacked pattern P satisfies the constraint conditions, the process proceeds to step S29. In step S29, the control unit 15 updates the policy function and the value function included in the search model M. In step S29, the policy function and the value function are updated based on the information obtained by executing all the episode steps. Specifically, in step S29, when searching for the stacked pattern P by the search model M, the policy function and the value function are updated such that the smaller the number of replacements between the layers of the fiber sheet S, the higher the reward. Then, the control unit 15 determines whether the learning of the search model M has ended (step S30). In step S30, if the control unit 15 determines that the learning has ended (step S30: Yes), the process proceeds to step S13. On the other hand, in step S30, if the control unit 15 determines that the learning has not ended (step S30: No), the process proceeds to step S22 to continue the learning of the search model M. In step S30, the end of the learning of the search model M may be determined based on, for example, whether all of the plurality of initial stacked patterns generated in step S21 or a predetermined number of them have been executed. The determination of the end of the learning of the search model M is not particularly limited to the above, and any determination may be made.

[0030] Note that in the above learning method, by executing step S25, when the control unit 15 determines that the stacked pattern P satisfies the constraint conditions, the process proceeds to step S29. However, the method is not particularly limited to this, and step S25 may be omitted, and even when the stacked pattern P satisfies the constraint conditions, the search by the search model M may be repeatedly executed until the number of steps of a predetermined episode is reached.

[0031] (Design method of design pattern) Next, with reference to FIG. 6, a method for designing a design pattern using the above-described design apparatus 10 will be described. The design pattern is a stacking pattern P of the laminate 1 that satisfies the constraint conditions. In FIG. 6, a case where the design pattern of the laminate 1 and the stacking pattern P used in the search model M have the same number of stacking layers will be described.

[0032] As shown in FIG. 6, in the method for designing a design pattern, first, the control unit 15 of the design apparatus 10 executes a step S31 of deriving a candidate pattern that is a stacking pattern P that is a candidate for the laminate 1 by a predetermined algorithm. The predetermined algorithm used in step S31 is, for example, a genetic algorithm. The candidate pattern is a stacking pattern regarding all the stackings of the laminate 1 derived by a predetermined algorithm. In step S31, the derived candidate pattern is used as the stacking pattern P to be input to the search model M. Subsequently, the control unit 15 inputs the candidate pattern to the search model M (step S32). The control unit 15 sets the input candidate pattern as the state of the stacking pattern P (step S33). Subsequently, the control unit 15 executes steps S33 to S38 using the learned search model M. Since steps S33 to S38 are the same steps as steps S23 to S28 in FIG. 5, the description thereof will be omitted. When the control unit 15 determines in step S34 that a predetermined number of episode steps has been reached, or when it determines in step S35 that the stacking pattern P satisfies the constraint conditions, the control unit 15 outputs the stacking pattern P obtained as the search result as the design pattern (step S39). After executing step S39, the control unit 15 ends the processing related to the design method.

[0033] Here, the number of stacked layers of the design pattern of the laminate 1 may be greater than the number of stacked layers of the stacked pattern P used in the search model M. In this case, as an example, after executing step S31 of deriving the candidate pattern of the laminate 1, the control unit 15 executes a step of extracting, as an extraction pattern, a stacked pattern P that does not satisfy the constraint conditions among the candidate patterns of the laminate 1. Then, in step S32, the control unit 15 inputs the extraction pattern into the search model M. As another example, after executing step S31 of deriving the candidate pattern of the laminate 1, in step S32, the control unit 15 sets a plurality of stacked patterns P side by side in the stacking direction so as to cover the number of stacked layers of the laminate 1. At this time, the control unit 15 sets them so that a part of the plurality of stacked patterns P arranged overlap. As an example of overlapping the plurality of stacked patterns P, the candidate pattern is divided into block units by the number of stacked layers of the stacked pattern P used in the search model M, optimized so that each stacked pattern P satisfies the constraint conditions, and then, across the interface between the blocks, the stacked pattern P used in the search model M is set and optimized so that the stacked pattern P across the interface satisfies the constraint conditions.

[0034] (Method for manufacturing a laminate) Next, with reference to FIG. 7, a method for manufacturing a laminate based on the stacked pattern P designed by the above-described design device 10 will be described.

[0035] As shown in FIG. 7, in the method for manufacturing the laminate 1, first, a step S41 of laminating the fiber sheet S is executed based on the design pattern derived by the design device 10. In step S41, for example, the fiber sheet S is laminated using an automatic laminating device that automatically laminates the fiber sheet S. That is, in step S41, the design pattern derived by the design device 10 is input to the automatic laminating device, and the fiber sheet S is laminated so as to have a predetermined design pattern. Subsequently, in the method for manufacturing the laminate 1, a step S42 of integrating the laminated fiber sheets S to form the laminate 1 is executed. In step S42, when the fiber sheet S is a reinforced fiber sheet impregnated with a thermosetting resin, the resin is thermally cured by heating, and the plurality of laminated fiber sheets S are integrated to form the laminate 1.

[0036] As described above, the learning method of the design device 10 (learning device), the design method of the design pattern, the manufacturing method of the laminate, and the design device 10 of the design pattern described in this embodiment are understood as follows, for example.

[0037] The learning method of the learning device (design device 10) according to the first aspect is a learning method of the learning device (design device 10) in which the learning device (design device 10) learns a search model M for searching for a lamination pattern P of a laminate 1 formed by laminating a fiber sheet S as a unidirectional material in which the fiber orientation direction is one direction. The search model M includes constraint conditions related to the lamination of the fiber sheet S and is a learning model using a policy function and a value function. The learning device is caused to execute a step S11 of obtaining an initial lamination pattern that is the lamination pattern P in the initial state, and a step S12 of learning the search model so that the lamination pattern P that satisfies the constraint conditions is obtained with the initial lamination pattern as an input.

[0038] According to this configuration, since the lamination pattern P can be used as an input to the search model M, it is possible to easily derive the search for the lamination pattern P that satisfies the constraint conditions by the search model M as compared with the case where physical property values are used as inputs.

[0039] As a second aspect, the search model M is a learning model that combines Monte Carlo tree search and deep reinforcement learning.

[0040] According to this configuration, efficient search by the search model M can be executed.

[0041] As a third aspect, the step S12 of training the search model M includes: a step S11 of obtaining the laminated pattern P that satisfies the constraint condition as the initial laminated pattern; a step S21 of generating a laminated pattern P that does not satisfy the constraint condition by swapping some of the layers of the fiber sheets S in the laminated pattern P that satisfies the constraint condition; and a step S22 of selecting the laminated pattern P that does not satisfy the constraint condition as an input of the initial laminated pattern.

[0042] According to this configuration, since a laminated pattern P that does not satisfy the constraint condition can be generated using the laminated pattern P that satisfies the constraint condition, the amount of data used for training the search model M can be reduced.

[0043] As a fourth aspect, the step S12 of training the search model M performs the search by swapping adjacent layers of the fiber sheets S when searching for the laminated pattern P by the search model M.

[0044] According to this configuration, since the search rule for the laminated pattern P can be made a simple rule, the laminated pattern P that satisfies the constraint condition can be efficiently searched by the search based on the simple rule.

[0045] As a fifth aspect, in the step S12 of training the search model M, the less the number of swaps of the layers of the fiber sheets S during the search for the laminated pattern P by the search model M, the higher the reward in the value function for the training.

[0046] According to this configuration, it is possible to train a search model M that can quickly derive an appropriate lamination pattern P considering the constraint conditions.

[0047] As a sixth aspect, the constraint conditions include at least one of a condition regarding the continuity of the fiber sheets in which the orientation directions are the same direction in the lamination direction, and a condition regarding the difference in the orientation angles formed by the orientation directions of adjacent fiber sheets in the lamination direction.

[0048] According to this configuration, it is possible to form a laminate 1 having a lamination pattern P with the required characteristics (strength, rigidity, generation and progression of damage).

[0049] As a seventh aspect, the condition regarding the continuity of the fiber sheet S is a condition in which the continuity of the fiber sheet S is three layers or less, and the condition regarding the difference in the orientation angles is a condition in which the difference in the orientation angles is 45° or less.

[0050] According to this configuration, the characteristics required for the laminate can be appropriately adjusted.

[0051] As an eighth aspect, the number of laminations in the lamination pattern P used in the search model M is less than the number of laminations of the laminate.

[0052] According to this configuration, since the number of laminations of the lamination pattern P used in the search model M can be made smaller than the number of laminations of the laminate 1, the learning load of the search model M can be reduced.

[0053] The design method of the design pattern according to the ninth aspect is a design pattern design method for designing, using a design apparatus, a design pattern that is a stacking pattern of the laminate satisfying the constraint conditions, by using the search model learned by the learning method of the learning apparatus described above, and the design apparatus is caused to execute: a step S31 of deriving, by a predetermined algorithm, a candidate pattern that is a candidate stacking pattern of the laminate; and steps S32 to S39 of inputting the candidate pattern of the laminate into the search model and deriving a design pattern that is a stacking pattern of the laminate satisfying the constraint conditions.

[0054] According to this configuration, by using the search model M with the candidate pattern as the input, an appropriate design pattern considering the constraint conditions can be easily derived.

[0055] As a tenth aspect, when the number of stacking layers of the design pattern of the laminate 1 is larger than the number of stacking layers in the stacking pattern used in the search model, after the execution of step S31 of deriving the candidate pattern of the laminate, a step of extracting, as an extraction pattern, a stacking pattern that does not satisfy the constraint conditions among the derived candidate patterns of the laminate is executed, and in steps S32 to S39 of deriving the design pattern that satisfies the constraint conditions, the extraction pattern is input into the search model.

[0056] According to this configuration, by setting the stacking pattern P of the search model M for the extraction pattern that is a part of the candidate pattern of the laminate 1 and does not satisfy the constraint conditions, a design pattern of the laminate 1 that satisfies the constraint conditions can be derived.

[0057] As an eleventh aspect, when the number of stacking layers of the laminate 1 is larger than the number of stacking layers in the stacking pattern P used in the search model M, in steps S32 to S39 of deriving the design pattern P that satisfies the constraint conditions, a plurality of the stacking patterns P are arranged so as to cover the candidate pattern, and are set so that a part of the plurality of arranged stacking patterns P overlaps.

[0058] According to this configuration, even when the number of stacked layers of the stacked pattern P of the search model M is small, the stacked pattern P of the search model M can be set for all the stacked numbers of the candidate patterns of the stacked body 1. Therefore, it is possible to obtain a design pattern of the stacked body 1 that satisfies the constraint conditions.

[0059] As a twelfth aspect, in steps S32 to S39 of deriving the design pattern that satisfies the constraint conditions, when searching for the stacked pattern P using the search model M, the layers of the adjacent fiber sheets S are swapped and the search is executed.

[0060] According to this configuration, since the search rule for the stacked pattern P can be made a simple rule, it is possible to efficiently search for the stacked pattern P that satisfies the constraint conditions by searching based on the simple rule.

[0061] The method for manufacturing the stacked body 1 according to the thirteenth aspect includes a step S41 of stacking the fiber sheets S based on the design pattern designed by the above-described design method of the design pattern, and a step S42 of integrating the stacked fiber sheets S to form the stacked body 1.

[0062] According to this configuration, it is possible to manufacture the stacked body 1 having a stacked pattern P that satisfies the constraint conditions.

[0063] The design device 10 for the design pattern according to the fourteenth aspect is a design device 10 for designing a design pattern that is the stacked pattern P of the stacked body 1 that satisfies the constraint conditions, using the search model M learned by the learning method of the above-described learning device (design device 10). The design device 10 includes a control unit 15 that executes a step S31 of deriving a candidate pattern that is the stacked pattern P that is a candidate for the stacked body 1 by a predetermined algorithm, and steps S32 to S39 of inputting the candidate pattern of the stacked body 1 into the search model M and deriving a design pattern that is the stacked pattern P that satisfies the constraint conditions.

[0064] According to this configuration, by using the search model M with the candidate pattern P as the input, an appropriate design pattern considering the constraint conditions can be easily derived.

Explanation of symbols

[0065] 1 laminate 10 design device 15 control unit 16 memory unit S fiber sheet P lamination pattern M search model D1 lamination data D2 constraint data

Claims

1. A learning method for a learning device in which the learning device learns a search model for searching for a stacking pattern of a laminate formed by stacking fiber sheets as a unidirectional material in which the fiber orientation direction is unidirectional, The search model includes constraint conditions related to the stacking of the fiber sheets and is a learning model using a policy function and a value function, In the learning device, obtaining an initial stacking pattern that is the stacking pattern in the initial state; using the initial stacking pattern as an input to train the search model so as to obtain a stacking pattern that satisfies the constraint conditions, The step of training the search model, obtaining a stacking pattern that satisfies the constraint conditions as the initial stacking pattern; generating a stacking pattern that does not satisfy the constraint conditions by swapping layers of some of the fiber sheets in the stacking pattern that satisfies the constraint conditions; selecting the stacking pattern that does not satisfy the constraint conditions as an input of the initial stacking pattern, a learning method for a learning device.

2. The learning method for a learning device according to claim 1, wherein the search model is a learning model that combines Monte Carlo tree search and deep reinforcement learning.

3. The step of training the search model, The learning method for a learning device according to claim 1 or 2, wherein when searching for the stacking pattern by the search model, swapping between adjacent layers of the fiber sheets and executing the search.

4. The step of training the search model, The learning method for a learning device according to claim 3, wherein the less the number of swaps between layers of the fiber sheets during the search for the stacking pattern by the search model, the higher the reward in the value function.

5. The constraint condition includes at least one of a condition related to the continuity of the fiber sheet in which the orientation direction is the same direction in the stacking direction and a condition related to the difference in the orientation angle formed by the orientation directions of the adjacent fiber sheets in the stacking direction. The learning method of the learning device according to any one of claims 1 to 4.

6. The condition related to the continuity of the fiber sheet is a condition in which the continuity of the fiber sheet is three layers or less, The condition related to the difference in the orientation angle is a condition in which the difference in the orientation angle is 45° or less. The learning method of the learning device according to claim 5.

7. The number of layers in the stacking pattern used in the search model is less than the number of layers in the stacked body. The learning method of the learning device according to any one of claims 1 to 6.

8. A learning method of a learning device in which the learning device learns a search model for searching for a stacking pattern of a stacked body formed by stacking fiber sheets using a fiber sheet as a unidirectional material in which the orientation direction of the fibers is one direction, The search model includes constraint conditions related to the stacking of the fiber sheets and is a learning model using a policy function and a value function, In the learning device, Obtaining an initial stacking pattern which is the stacking pattern in the initial state, Using the initial stacking pattern as an input, the search model is learned so as to be a stacking pattern that satisfies the constraint condition. The step of learning the search model is The learning is such that the less the number of replacements of the layers of the fiber sheets during the search for the stacking pattern by the search model, the higher the reward in the value function. The learning method of the learning device. A learning method for a learning device in which the learning device learns a search model for searching for a lamination pattern of a laminate formed by laminating a fiber sheet as a unidirectional material in which the fiber orientation direction is unidirectional, The search model includes constraint conditions related to the lamination of the fiber sheet and is a learning model using a policy function and a value function, In the learning device, Obtaining an initial lamination pattern that is the lamination pattern in the initial state; Using the initial lamination pattern as an input, the search model is learned so as to obtain the lamination pattern that satisfies the constraint conditions, A learning method for a learning device, wherein the number of laminations in the lamination pattern used in the search model is less than the number of laminations of the laminate.

10. A design method for a design pattern, in which a design pattern that is a lamination pattern of the laminate satisfying the constraint conditions is designed using a design device, using the search model learned by the learning method of the learning device according to any one of Claims 1 to 9, In the design device, Deriving a candidate pattern that is a candidate lamination pattern of the laminate by a predetermined algorithm; Inputting the candidate pattern of the laminate into the search model and deriving a design pattern that is the lamination pattern satisfying the constraint conditions.

11. When the number of laminations of the design pattern of the laminate is larger than the number of laminations in the lamination pattern used in the search model, After executing the step of deriving the candidate pattern of the laminate, among the derived candidate patterns of the laminate, extracting the lamination pattern that does not satisfy the constraint conditions as an extraction pattern, The design method of the design pattern according to claim 10, wherein in the step of deriving the design pattern that satisfies the constraint condition, the extracted pattern is input into the search model.

12. When the number of layers of the laminate is larger than the number of layers in the lamination pattern used in the search model, The design method of the design pattern according to claim 10, wherein in the step of deriving the design pattern that satisfies the constraint condition, a plurality of the lamination patterns are arranged side by side so as to cover the candidate pattern of the laminate, and a part of the plurality of arranged lamination patterns is set to overlap.

13. The step of deriving the design pattern that satisfies the constraint condition is The design method of the design pattern according to any one of claims 10 to 12, wherein during the search for the lamination pattern by the search model, the layers of the adjacent fiber sheets are swapped and the search is executed.

14. A method for manufacturing a laminate, comprising: laminating the fiber sheets based on the design pattern designed by the design method of the design pattern according to any one of claims 10 to 13; and integrating the laminated fiber sheets to form the laminate.

Images