Condition determination device, method for determining condition, and program
The condition determination device uses Bayesian optimization to adapt experimental data for production systems, resolving scale-up issues by efficiently determining optimal conditions through iterative learning.
Patent Information
- Application Number
- JP2023082131
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2023-05-18
- Publication Date
- 2024-11-28
AI Technical Summary
Existing manufacturing processes face scale-up issues where optimal conditions in experimental systems do not translate to optimal results in production systems, leading to wasted time, manpower, and materials due to repeated trial and error.
A condition determination device and method using Bayesian optimization to determine trial conditions for a production system based on data from an experimental system, with a control unit adjusting conditions until predetermined termination criteria are met, employing a surrogate model to predict optimal conditions efficiently.
Effectively addresses scale-up problems by determining optimal production conditions at a lower cost, minimizing trial and error, and ensuring performance criteria are met.
Smart Images

Figure 2024165717000001_ABST
Abstract
Description
[Technical field]
[0001] The present disclosure relates to a condition determination device, a condition determination method, and a program. [Background technology]
[0002] In recent years, the use of machine learning technology has been expanding in various industrial fields. For example, in manufacturing plants and systems, there are attempts to set process conditions such as pressure and temperature based on the results of experimental plants and simulations using machine learning technology.
[0003] Optimization techniques can be used to set process conditions. For example, JP 2022-185927 A describes the use of constrained optimization techniques using Bayesian optimization for industrial product development or manufacturing process development. [Prior art documents] [Patent documents]
[0004] [Patent Document 1] JP 2022-185927 A Summary of the Invention [Problem to be solved by the invention]
[0005] On the other hand, when a manufacturing plant or system is operated under process conditions derived from an experimental plant or system or a simulation result, it is known that there are cases where the expected performance cannot be achieved, that is, the so-called scale-up problem. As a result, even if the production conditions are derived using an experimental plant or a simulation, the test runs are repeated many times on-site through trial and error, which may waste the time, manpower, and materials required for this.
[0006] One objective of the present disclosure is to provide a technique for determining conditions for effectively solving scale-up problems. [Means for solving the problem]
[0007] One aspect of the present disclosure relates to a condition determination device having a trial condition determination unit that determines trial conditions for a second system based on first data regarding a first system and second data regarding a second system, and a control unit that causes the trial condition determination unit to determine further trial conditions for the second system when a trial result for the second system attempted under the trial conditions does not satisfy a predetermined termination condition. Effect of the Invention
[0008] According to the present disclosure, a condition determination technique for effectively solving the scale-up problem can be provided. [Brief description of the drawings]
[0009] [Figure 1] FIG. 1 is a schematic diagram illustrating a condition determining device according to an embodiment of the present disclosure. [Diagram 2] FIG. 2 is a block diagram illustrating a hardware configuration of a condition determination device according to an embodiment of the present disclosure. [Diagram 3] FIG. 3 is a schematic diagram illustrating a condition determination process according to an embodiment of the present disclosure. [Figure 4] FIG. 4 is a block diagram illustrating a functional configuration of a condition determination device according to an embodiment of the present disclosure. [Diagram 5] FIG. 5 is a schematic diagram illustrating Bayesian optimization according to one embodiment of the present disclosure. [Figure 6] FIG. 6 is a flowchart illustrating a condition determination process according to an embodiment of the present disclosure. [Figure 7] FIG. 7 is a schematic diagram illustrating a condition determination process according to another embodiment of the present disclosure. [Figure 8] FIG. 8 is a schematic diagram illustrating a condition determination process according to another embodiment of the present disclosure. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0010] Hereinafter, embodiments of the present disclosure will be described with reference to the drawings.
[0011] In the following embodiment, a condition determining device is disclosed that determines conditions for a manufacturing process or the like by utilizing two or more different systems, such as an experimental system and a production system.
[0012] 1, for a manufacturing process in a manufacturing plant or the like, the condition determining device 100 determines conditions to be used in the production system 30 by utilizing an experiment system 20 for experiments of the manufacturing process and a production system 30 in which the manufacturing process itself is actually carried out in the manufacturing plant. For example, in the case of a chemical manufacturer, the condition determining device 100 can determine conditions such as temperature, pressure, flow rate, and concentration in unit operations such as reaction and separation.
[0013] It may be desirable to determine such conditions so as to maximize an index showing a specific performance, such as a yield rate. Typically, various conditions are examined in a small-scale system, such as an experimental system 20, in which the trial costs are lower than those of the production system 30, and the results of the examination are often set in a large-scale system, such as the production system 30 in a factory.
[0014] For example, in the optimization of factory conditions, the experimental data may be laboratory data, simulation data, etc., and the production data may be data of an actual factory. In this case, the conditions may be the temperature, pressure, flow rate, etc. of the equipment, and the results may be sales, profits, yield rate, refinement, production time, failure rate, etc. In addition, the experimental data may be data of an actual factory for product A, and the production data may be data of an actual factory for product B. In this case, the conditions may be part size, part position, part accuracy, etc., and the results may be sales, profits, yield rate, refinement, production time, failure rate, etc. In addition, in the optimization of new material conditions, the experimental data may be simulation data, and the production data may be laboratory data. In this case, the conditions may be the type of atoms, composition ratio, etc., and the results may be thermal conductivity, resistivity, energy, etc. In addition, in the optimization of drug conditions, the experimental data may be simulation data, and the production data may be laboratory data. In this case, the conditions may be the type of atoms, composition ratio, etc., and the results may be the effect of the drug.
[0015] On the other hand, as is known as the scale-up problem, optimal conditions in the experimental system 20 do not necessarily lead to optimal results in the production system 30. For this reason, the condition determination device 100 according to the following embodiment executes a condition determination process, which will be described in detail below, in order to determine optimal conditions in the production system 30 at the lowest possible cost.
[0016] Here, the condition determination device 100 may be realized by a computing device such as a server, a personal computer (PC), a smartphone, a tablet, etc., and may have a hardware configuration such as that shown in Fig. 2. That is, the condition determination device 100 has a drive device 101, a storage device 102, a memory device 103, a processor 104, a user interface (UI) device 105, and a communication device 106, which are interconnected via a bus B.
[0017] Programs or instructions for implementing various functions and processes in the condition determination device 100 may be stored in a removable storage medium such as a CD-ROM (Compact Disk-Read Only Memory) or a flash memory. When the storage medium is set in the drive device 101, the programs or instructions are installed from the storage medium via the drive device 101 into the storage device 102 or the memory device 103. However, the programs or instructions do not necessarily have to be installed from the storage medium, and may be downloaded from any external device via a network or the like.
[0018] The storage device 102 is realized by a hard disk drive or the like, and stores installed programs or instructions as well as files, data, etc. used in executing the programs or instructions.
[0019] The memory device 103 is realized by a random access memory, a static memory, or the like, and when a program or an instruction is activated, reads and stores the program or instruction, data, or the like from the storage device 102. The storage device 102, the memory device 103, and the removable storage medium may be collectively referred to as a non-transitory storage medium.
[0020] The processor 104 may be realized by one or more CPUs (Central Processing Units), GPUs (Graphics Processing Units), processing circuitry, etc., which may be composed of one or more processor cores, and performs various functions and processing of the condition determination device 100 in accordance with programs, instructions, data such as parameters necessary to execute the programs or instructions, etc. stored in the memory device 103.
[0021] The user interface (UI) device 105 may be composed of input devices such as a keyboard, a mouse, a camera, a microphone, etc., output devices such as a display, a speaker, a headset, a printer, etc., and input / output devices such as a touch panel, and realizes an interface between a user and the condition determination device 100. For example, a user operates the condition determination device 100 by operating a GUI (Graphical User Interface) displayed on a display or a touch panel using a keyboard, a mouse, etc.
[0022] The communication device 106 is realized by various communication circuits that execute wired and / or wireless communication processing with an external device, the Internet, a LAN (Local Area Network), a cellular network, or other communication network.
[0023] However, the above-described hardware configuration is merely an example, and the condition determination device 100 according to the present disclosure may be realized by any other appropriate hardware configuration.
[0024] [Condition determination device] Next, a condition determining device 100 according to an embodiment of the present disclosure will be described. Fig. 3 is a schematic diagram showing a condition determining process according to an embodiment of the present disclosure. As shown in Fig. 3, in an experimental system 20, various conditions (e.g., temperature, pressure, flow rate, concentration, etc.) are tried, and the results (e.g., yield rate, etc.) tried under each condition are collected, and a data set D of experimental system data consisting of pairs of conditions and results is generated. S is provided to the condition determining device 100. The condition determining device 100 initially receives the acquired experimental data set D S Based on the above, the condition determination device 100 determines trial conditions for a trial in the production system 30. When the trial conditions are acquired from the condition determination device 100, the production system 30 executes work under the acquired trial conditions. If the trial results under the trial conditions satisfy a predetermined termination condition, for example, if the yield rate achieved under the trial conditions is equal to or greater than a predetermined threshold, the condition determination process is terminated. On the other hand, if the yield rate achieved under the trial conditions is less than the predetermined threshold, the condition determination device 100 determines the experimental system data set DS and the trial results by the production system 30. The condition determination device 100 causes the production system 30 to execute the work under the trial conditions, and collects the trial conditions and trial results acquired in the production system 30 to obtain a production system data set D L Next, the condition determination device 100 obtains the experimental data set D S , Production Data Set D L The next trial conditions are determined based on the above, and the production system 30 is caused to execute the work under the trial conditions. The above-mentioned process is repeated until the trial results under the trial conditions satisfy a predetermined end condition.
[0025] Fig. 4 is a block diagram showing a functional configuration of a condition determination device 100 according to an embodiment of the present disclosure. As shown in Fig. 4, the condition determination device 100 has a trial condition determination unit 110 and a control unit 120. Each functional unit of the trial condition determination unit 110 and the control unit 120 may be realized by a processor 104 executing a computer program stored in a memory device 103 of the condition determination device 100.
[0026] The trial condition determination unit 110 determines trial conditions for the second system based on first data related to the first system and second data related to the second system. Here, the first system and the second system are realized, for example, by an experimental system 20 and a production system 30, respectively, and the first system may be executable with lower trial cost and accuracy than the second system. Specifically, the trial condition determination unit 110 determines trial conditions for the second system based on an experimental system data set D related to the experimental system 20. S and production dataset D for production system 30 L Based on the above, trial conditions for the production system 30 may be determined. Here, the production system data set D L may be composed of pairs of trial conditions determined by the trial condition determination unit 110 and trial results obtained in the production system 30 executed under the respective trial conditions. Note that, initially, the trial condition determination unit 110 generates an experimental system data set D S The trial conditions for the production system 30 may be determined based only on the above.
[0027] For example, the trial condition determination unit 110 may determine the next trial condition by using Bayesian optimization as described below. That is, the trial condition determination unit 110 may regard the production system 30 as a black-box function and sequentially search for trial conditions by using a surrogate model for the black-box function.
[0028] According to Bayesian optimization, the trial condition determination unit 110 performs 1) approximation of the production system 30 by a surrogate model, and 2) design of an acquisition function from the surrogate model. For example, as shown in FIG. 5, the trial condition determination unit 110 sets a surrogate model that approximates the production system 30. As the surrogate model, for example, Gaussian process regression can be used without being limited. n Given, the output y i =f(x i ) joint probability p(y1,y2,...,y n If the input x follows an n-dimensional Gaussian (normal) distribution, then the function f(·) is said to follow a Gaussian process. i can be a scalar or a vector, but for simplicity, it is assumed to be a scalar here.
[0029] Here, it is assumed that the problem is a maximization problem that searches for an input that maximizes the output. However, the Bayesian optimization according to the present disclosure is not limited to this. For example, it can be considered as minimization by inverting the sign. Also, if you want to approach a certain value, you can consider the difference between the output and that specific value, or the square or absolute value of the difference, and minimize that difference.
[0030] When Gaussian process regression is used as the surrogate model, the trial condition determination unit 110 generates an experimental system data set D from the experimental system 20 and the production system 30. S and production dataset D L and the experimental data set D S and production dataset D L and the predictive distribution of the Gaussian process based on
number
number
number
number
number
number
number
number
[0031] Also, for a positive definite kernel, for a set X,
number
number
number
number
number
[0032] In this case, the following kernel trick is known to hold true.
number
number
[0033] In this way, the experimental data set D S The predictive distribution of a Gaussian process as a surrogate model based on
number
number
number
[0034] The trial condition determination unit 110 determines the trial conditions to be sampled next from the acquisition function derived in this manner, and transmits the determined trial conditions to the production system 30. Upon acquiring the trial conditions, the production system 30 executes the work under the acquired trial conditions, and acquires results such as the yield rate.
[0035] That is, trial condition determination unit 110 determines trial conditions for production system 30 based on data related to experimental system 20 and data related to production system 30. Note that, initially, trial condition determination unit 110 may determine trial conditions for production system 30 based only on data related to experimental system 20.
[0036] Specifically, the trial condition determination unit 110 uses Bayesian optimization to determine the experimental data set D S and production dataset D L For example, the trial condition determining section 110 may model the data generation process of the experimental system 20 as follows.
number
[0037] That is, the parameter W of experimental system 20 S has mean 0 and variance λ S 2 I is a Gaussian distribution, and the noise in the output of experimental system 20 has mean 0 and variance σ S 2 According to this, the output of the experimental system 20 is the parameter W S It is assumed that noise is added to the matrix product of and feature vector φ(x). Furthermore, the trial condition determination unit 110 may model the data generation process of the production system 30 as follows.
number
[0038] The following experimental data set D S and production dataset D L
number
number
number
[0039] First, the joint distribution is
number
[0040] Now, to make the formula easier to read,
number
number
number
[0041] When the trial result for the second system, which is tried under the trial condition, does not satisfy a predetermined end condition, the control unit 120 causes the trial condition determination unit 110 to determine further trial conditions for the second system. Specifically, when the result (e.g., yield rate) of the work performed in the production system 30 according to the trial condition determined by the trial condition determination unit 110 does not reach or exceed a predetermined threshold, the control unit 120 may instruct the trial condition determination unit 120 to determine the next trial condition for the production system 30. On the other hand, when the work result is equal to or exceeds the predetermined threshold, the control unit 120 determines that the trial condition is satisfactory and ends the condition determination process. That is, until the work result under the trial condition determined by the trial condition determination unit 110 satisfies the predetermined end condition, the control unit 120 causes the trial condition determination unit 110 to sequentially determine trial conditions and causes the production system 30 to try the work under the sequentially determined trial conditions. Here, the threshold value is set as the satisfying condition, which is merely an example, and may be, for example, the number of trials, and is not limited to this example.
[0042] In the above-described embodiment, the experimental system data and / or the production system data are composed of pairs of trial conditions and trial results, but the experimental system data and / or the production system data according to the present disclosure is not limited to this, and may include, for example, a prediction model that predicts trial results from the trial conditions of the experimental system 20 and / or the production system 30. Such a prediction model enables the trial condition determination unit 110 to predict trial results from the trial conditions for the experimental system 20 and / or the production system 30, and the experimental system data set D S and / or production dataset D L There may be no need to obtain
[0043] [Condition determination process] Next, a condition determination process according to an embodiment of the present disclosure will be described. The condition determination process is executed by the condition determination device 100, and more specifically, may be realized by one or more processors 104 of the condition determination device 100 executing one or more programs or instructions stored in one or more memory devices 103. Fig. 6 is a flowchart showing the condition determination process according to an embodiment of the present disclosure.
[0044] In step S101, the condition determining device 100 acquires data related to the experimental system 20. Specifically, the condition determining device 100 acquires an experimental system data set D, which is composed of pairs of conditions and results indicating experimental results in the experimental system 20. S Get the.
[0045] In step S102, the condition determining device 100 determines trial conditions for the production system 30 based on the data related to the experimental system 20 and the data related to the production system 30. Note that, initially, the condition determining device 100 may determine trial conditions for the production system 30 based only on the data related to the experimental system 20.
[0046] Specifically, the condition determination device 100 uses Bayesian optimization to S and production dataset D L For example, the condition determining device 100 may model the data generation process of the experimental system 20 as follows.
number
[0047] That is, the parameter W of experimental system 20 S has mean 0 and variance λ S 2 I is a Gaussian distribution, and the noise in the output of experimental system 20 has mean 0 and variance σ S 2 According to this, the output of the experimental system 20 is the parameter W S It is assumed that noise is added to the product of the matrix of the feature vector φ(x) and the feature vector φ(x). Furthermore, the condition determining device 100 may model the data generation process of the production system 30 as follows.
number
[0048] The following experimental data set D S and production dataset D L
number
number
number
[0049] First, the joint distribution is
number
[0050] Now, to make the formula easier to read,
number
number
number
[0051] In step S103, the condition determination device 100 acquires trial results of the production system 30 that has been tried under the trial conditions. Specifically, the condition determination device 100 acquires trial results for the trial conditions from the production system 30 that has been tried under the trial conditions determined in step S103.
[0052] In step S104, the condition determining device 100 determines whether a predetermined end condition is satisfied. The predetermined end condition may be, for example, that the trial result exceeds a predetermined evaluation criterion, such as that the yield rate as the trial result is equal to or greater than a predetermined threshold. Using a threshold as the satisfied condition is just an example, and may be, for example, the number of trials, and is not limited to this example. If the predetermined end condition is satisfied (S104: YES), the condition determining device 100 ends the condition determination process. On the other hand, if the predetermined end condition is not satisfied (S104: NO), the condition determining device 100 returns to step S102 and repeats steps S102 to S104.
[0053] In the above-described embodiment, the condition determining device 100 uses the experimental data set D S and production dataset D LHowever, the determination of the trial conditions according to the present disclosure is not necessarily limited to this.
[0054] In one embodiment, the control unit 120 may cause the trial condition determination unit 110 to acquire first data related to the first systems from a plurality of first systems. Specifically, the control unit 120 may acquire an experimental system data set D from a plurality of experimental systems 20_1, . . . , 20_N as shown in FIG. S1 ,···,D SN The experimental data set D S1 ,···,D SN In other words, the trial condition determination unit 110 may determine the trial conditions for the production system 30 based on the experimental system data set D S1 ,···,D SN The next trial condition may be determined by performing Bayesian optimization based on the experimental data set D. S1 ,···,D SN The trial condition determination unit 110 may derive an acquisition function (for example, PI) from the Gaussian process, and determine the sampling point that maximizes the acquisition function as the trial condition for the production system 30.
[0055] Furthermore, in one embodiment, control unit 120 may cause trial condition determination unit 110 to acquire first data related to a first system from a plurality of first systems, cause trial condition determination unit 110 to acquire second data related to a second system from a second system, and cause trial condition determination unit 110 to determine further trial conditions for the second system based on the first data and the second data, as shown in Fig. 8. That is, the embodiments described with reference to Figs. 3 and 7 may be used in combination.
[0056] In the above embodiment, the trial condition determination unit 110 determines the trial conditions using a Gaussian process as a surrogate model, but the present disclosure is not limited to this, and a model such as Physics-Informed Neural Networks (deep learning based on physical laws) may be used. This is expected to enable prediction according to the law even in an extrapolation area with no data.
[0057] In the above embodiment, the trial results such as the yield rate are evaluated, but multiple evaluation targets may be used. For example, multiple other indices such as yield, quality, cost, and safety may be optimized.
[0058] According to the above-described embodiment, the condition determining device 100 can determine optimal conditions in the production system 30 at the lowest possible cost, and can effectively solve the scale-up problem.
[0059] (Appendix 1) a trial condition determination unit that determines trial conditions for the second system based on first data regarding the first system and second data regarding the second system; a control unit that causes the trial condition determination unit to determine further trial conditions for the second system when a trial result for the second system tried under the trial conditions does not satisfy a predetermined termination condition; A condition determination device having the following. (Appendix 2) 2. The condition determination device of claim 1, wherein the first system is executable with lower trial cost and accuracy than the second system. (Appendix 3) The condition determination device according to claim 1 or 2, wherein the control unit causes the trial condition determination unit to determine the further trial conditions based further on the trial results. (Appendix 4) 4. The condition determining device according to any one of claims 1 to 3, wherein the control unit causes the trial condition determining unit to obtain the first data from a plurality of first systems. (Appendix 5) 5. The condition determining device according to any one of claims 1 to 4, wherein the first data includes trial conditions and trial results of the first system. (Appendix 6) 6. The condition determining device according to any one of appendices 1 to 5, wherein the first data includes a prediction model that predicts a trial result from trial conditions of the first system. (Appendix 7) determining trial conditions for the second system based on first data for the first system and second data for the second system; determining further trial conditions for the second system when a trial result for the second system tried under the trial conditions does not satisfy a predetermined termination condition; 23. A computer implemented method for determining a condition, comprising: (Appendix 8) determining trial conditions for the second system based on first data for the first system and second data for the second system; determining further trial conditions for the second system when a trial result for the second system tried under the trial conditions does not satisfy a predetermined termination condition; A program that causes a computer to execute the following.
[0060] Although the examples of the present disclosure have been described in detail above, the present disclosure is not limited to the specific embodiments described above, and various modifications and variations are possible within the scope of the gist of the present disclosure described in the claims. [Explanation of symbols]
[0061] 20 Experimental 30 Production 100 Condition determination device 110 Trial condition determination unit 120 Control section
Claims
1. a trial condition determination unit that determines trial conditions for the second system based on first data regarding the first system and second data regarding the second system; a control unit that causes the trial condition determination unit to determine further trial conditions for the second system when a trial result for the second system tried under the trial conditions does not satisfy a predetermined termination condition; A condition determination device having the following.
2. 2. The condition determination apparatus of claim 1, wherein the first system is executable with lower trial cost and accuracy than the second system.
3. The condition determination device according to claim 1 , wherein the control unit causes the trial condition determination unit to determine the further trial condition further based on the trial result.
4. The condition determining device according to claim 1 , wherein the control unit causes the trial condition determining unit to obtain the first data from a plurality of first systems.
5. The condition determination device according to claim 1 , wherein the first data includes trial conditions and trial results of the first system.
6. The condition determination device according to claim 1 , wherein the first data includes a prediction model that predicts a trial result from a trial condition of the first system.
7. determining trial conditions for the second system based on first data for the first system and second data for the second system; determining further trial conditions for the second system when a trial result for the second system tried under the trial conditions does not satisfy a predetermined termination condition; 23. A computer implemented method for determining a condition, comprising:
8. determining trial conditions for the second system based on first data for the first system and second data for the second system; determining further trial conditions for the second system when a trial result for the second system tried under the trial conditions does not satisfy a predetermined termination condition; A program that causes a computer to execute the following.
Citation Information
Patent Citations
Evaluation device, evaluation method, and program
JP2022185927A
