Estimation model generation device, estimation device, estimation model generation method, estimation method, estimation model generation program, and estimation program

Abstract

The purpose of the present invention is to provide an estimation model generation device capable of estimating a crystal transition feature.　An estimation model generation device according to the present invention is characterized by comprising: a training data input unit for receiving training data which is a set of condition data including crystal transition data, which is a feature amount representing a crystal transition feature, and at least one type of data selected from one or more items of process data, which is a feature amount representing a crystal transition method for performing crystal transition in a crystal transition device, one or more items of starting material data, which is a feature amount representing the starting material of the crystal transition, and one or more items of device data, which is a feature amount representing the crystal transition device, and; and an estimation model generation unit for generating, by machine learning using the training data, an estimation model that outputs the crystal transition data upon receiving the condition data as an input.

Description

Estimation Model Generation Device and Estimation Device, Estimation Model Generation Method and Estimation Method, and Estimation Model Generation Program and Estimation Program

[0001] The present invention relates to an estimation model generation device and an estimation device for estimating the characteristics of crystal transition performed by a crystal transition device, an estimation model generation method and an estimation method, and an estimation model generation program and an estimation program.

[0002] Generally, it has been reported that in many compounds, regardless of whether they are organic compounds or inorganic compounds, there are multiple crystal forms with different crystal structures, which is called crystal polymorphism. The multiple crystal forms existing in crystal polymorphism not only show different patterns in analyses such as X-ray diffraction and infrared spectroscopic analysis, but also have different physical properties such as melting point and solubility. Generally, the more energetically stable the crystal form under defined conditions, the higher the melting point and the lower the solubility tend to be, and usually the crystal form with the highest melting point and the lowest solubility is called the stable form.

[0003] When a compound having crystal polymorphism is used particularly in pharmaceutical applications, etc., it is important to select an optimal crystal form such as the stable form. As an optimal crystal form in such a compound having crystal polymorphism, for example, Form II crystal among Form I crystal and Form II crystal of reduced coenzyme Q10 is known (Patent Document 1). Further, in a crystal transition method capable of performing crystal transition to an optimal crystal form by obtaining a product containing the optimal crystal form (another crystal form) from a raw material containing one crystal form in a compound having crystal polymorphism, a method for efficiently producing the optimal crystal form is required. As such a method, for example, a method is known in which a mixture of Form I crystal and Form II crystal of reduced coenzyme Q10 is heated to 32°C or higher in the presence of 0.001 to 50 parts by weight of a solvent based on 100 parts by weight of the total amount of crystals to increase the ratio of Form II crystal (Patent Document 2).

[0004] International Publication No. 2012 / 176842 International Publication No. 2020 / 067275

[0005] On the other hand, in crystal transition methods, which involve obtaining a product containing the optimal crystal form (another crystal form) from a raw material containing one crystal form to achieve a crystal transition to the optimal crystal form, the crystal transition characteristics of the product, such as the crystal transition rate, crystal orientation, and oxidation stability, are influenced by various conditions such as the crystal transition method and raw materials. Therefore, there is a need to develop methods that can estimate the crystal transition characteristics, such as the crystal transition rate, according to various conditions such as the crystal transition method and raw materials, and that can identify the various conditions necessary to achieve the desired crystal transition characteristics, such as the crystal transition rate.

[0006] The present invention has been made in view of these points, and its object is to provide an estimation model generation apparatus and estimation apparatus, an estimation model generation method and estimation method, and an estimation model generation program and estimation program that can estimate the characteristics of crystal transitions.

[0007] To solve the above problems, the present invention provides an estimation model generation device comprising: a training data input unit that accepts training data which is a pair of condition data including one or more data selected from one or more types of process data which are feature quantities that represent a crystal transition method in which a crystal transition is performed in a crystal transition apparatus, one or more types of raw material data which are feature quantities that represent the raw materials for the crystal transition, and one or more types of apparatus data which are feature quantities that represent the crystal transition apparatus, and crystal transition data which are feature quantities that represent the characteristics of the crystal transition; and an estimation model generation unit that uses the training data to generate an estimation model by machine learning that outputs the crystal transition data when the condition data is input.

[0008] Furthermore, the estimation device of the present invention includes a condition data input unit that accepts input of condition data including one or more types of data selected from one or more types of process data which are feature quantities that represent a crystal transition method in which a crystal transition is performed in a crystal transition apparatus, one or more types of raw material data which are feature quantities that represent the raw materials for the crystal transition, and one or more types of apparatus data which are feature quantities that represent the crystal transition apparatus; and an estimation unit that obtains estimated values ​​of crystal transition data which are feature quantities that represent the characteristics of the crystal transition from the condition data by using an estimation model, wherein the estimation model is an estimation model generated by machine learning using training data which is a pair of the condition data and the crystal transition data, so as to output the crystal transition data when the condition data is input.

[0009] Furthermore, the present invention provides a method for generating an estimation model, comprising the steps of: preparing training data which is a pair of condition data including one or more data selected from one or more types of process data which are feature quantities that represent a crystal transition method performed in a crystal transition apparatus, one or more types of raw material data which are feature quantities that represent the raw materials for the crystal transition, and one or more types of apparatus data which are feature quantities that represent the crystal transition apparatus, and crystal transition data which are feature quantities that represent the characteristics of the crystal transition; and generating an estimation model by machine learning using the training data which outputs the crystal transition data when the condition data is input.

[0010] Furthermore, the estimation method of the present invention comprises the steps of preparing condition data including one or more data selected from one or more process data which are feature quantities that represent a crystal transition method in which a crystal transition is performed in a crystal transition apparatus, one or more raw material data which are feature quantities that represent the raw materials for the crystal transition, and one or more apparatus data which are feature quantities that represent the crystal transition apparatus; and obtaining estimated values ​​of crystal transition data which are feature quantities that represent the characteristics of the crystal transition from the condition data by using an estimation model, wherein the estimation model is an estimation model generated by machine learning using training data which is a set of the condition data and the crystal transition data, so as to output the crystal transition data when the condition data is input.

[0011] Furthermore, the estimation model generation program of the present invention is characterized by causing an information processing device to execute the estimation model generation method described above.

[0012] Furthermore, the estimation program of the present invention is characterized by causing an information processing device to execute the above estimation method.

[0013] Furthermore, the present invention's estimation model generation method is characterized by comprising the steps of: preparing training data which is a pair of condition data including one or more types of data selected from one or more types of process data which are feature quantities that represent a crystal transition method in which a crystal transition is performed in a crystal transition apparatus, one or more types of raw material data which are feature quantities that represent the raw materials for the crystal transition, and one or more types of equipment data which are feature quantities that represent the crystal transition apparatus, and crystal transition data which are feature quantities that represent the characteristics of the crystal transition; and generating an estimation model by machine learning using the training data which outputs the remaining types of data of the condition data when analysis data including some types of data of the condition data and the crystal transition data is input.

[0014] Furthermore, the estimation method of the present invention comprises the steps of preparing some types of condition data, which include two or more types of data selected from one or more types of process data, which are feature quantities representing a crystal transition method in which a crystal transition is performed in a crystal transition apparatus; one or more types of raw material data, which are feature quantities representing the raw materials for the crystal transition; and one or more types of apparatus data, which are feature quantities representing the crystal transition apparatus, and crystal transition data, which are feature quantities representing the characteristics of the crystal transition; and obtaining estimated values ​​of the remaining types of condition data from analysis data including some types of condition data and the crystal transition data by using an estimation model, wherein the estimation model is an estimation model generated by machine learning using training data which is a pair of the condition data and the crystal transition data, so as to output the remaining types of condition data when analysis data including some types of condition data and the crystal transition data is input. This specification encompasses the disclosures of Japanese Patent Application No. 2025-015212, which forms the basis of the priority claim of this application.

[0015] According to the present invention, the characteristics of the crystal transition can be estimated.

[0016] This is a diagram illustrating the configuration of the estimation model generation device according to the first embodiment. This is a schematic process diagram of the estimation model generation method according to the first embodiment. This is a diagram illustrating the configuration of the estimation device according to the first embodiment. This is a schematic process diagram of the estimation method according to the first embodiment. This is a diagram illustrating the configuration of the estimation model generation device for inverse analysis according to the second embodiment. This is a schematic process diagram of the estimation model generation method for inverse analysis according to the second embodiment. This is a diagram illustrating the configuration of the estimation device for inverse analysis according to the second embodiment. This is a schematic process diagram of the estimation method for inverse analysis according to the second embodiment. This is a graph showing the feature importance of conditional data in the estimation model of a random forest. This is a graph showing the feature importance of conditional data in the estimation model of a decision tree.

[0017] The following describes embodiments relating to the estimation model generation apparatus and estimation apparatus, estimation model generation method and estimation method, and estimation model generation program and estimation program of the present invention.

[0018] [First Embodiment] (Estimation Model Generation Apparatus and Estimation Model Generation Method) Figure 1 is a schematic diagram illustrating the configuration of the estimation model generation apparatus according to the first embodiment. Figure 2 is a schematic process diagram of the estimation model generation method according to the first embodiment.

[0019] As shown in Figure 1, the estimation model generation device according to the first embodiment includes an input unit including a training data input unit, a processing unit including an estimation model generation unit, an output unit including an estimation model output unit, and a storage unit storing the estimation model generation program according to the first embodiment. The estimation model generation device is composed of, for example, an information processing device such as a desktop computer. The information processing device, although not shown, includes a processing unit (computer), a storage device (e.g., ROM, RAM, HDD, etc.), an input device, an output device, and an input / output interface (I / F), etc. The estimation model generation program causes the information processing device to execute the estimation model generation method by making the estimation model generation device function.

[0020] In the estimation model generation method according to the first embodiment, an estimation model generation device is used to generate an estimation model for estimating the characteristics of a crystal transition from condition data when a crystal transition is performed in a crystal transition apparatus. Specifically, in the estimation model generation method, first, as shown in Figure 2, multiple sets of training data are prepared, which are pairs of condition data including one or more data selected from one or more types of data that represent the crystal transition method performed in a crystal transition apparatus, one or more types of data that represent the raw materials for the crystal transition, and one or more types of equipment data that represent the crystal transition apparatus, and crystal transition data that represents the characteristics of the crystal transition (training data preparation step).

[0021] A crystal transition apparatus is a device that performs crystal transitions. Here, "crystal transition" refers to the transition from one crystalline form to another in polymorphic organic or inorganic compounds. Specific examples of crystal transitions are not particularly limited as long as they are transitions from one crystalline form to another, but for example, the transition from the metastable Form I crystal (one crystalline form) of reduced coenzyme Q10 (hereinafter sometimes abbreviated as "Q10") to the stable Form II crystal (another crystalline form), as well as the transition from the metastable crystal (one crystalline form) to the stable crystal (another crystalline form) of ethyl=(R)-2-[4-(6-chloro-2-quinoxalyloxy)phenoxy]propionic acid, as cited in Japanese Patent Publication No. 2001-114707, etc., and MMH Smets et al., Cryst. Growth Des. 2015, 15, 10, Examples include the transitions between α-type crystals (one crystal form) and β-type crystals (another crystal form) of DL-norleucine, as listed in 5157-5167, and the transitions between α-type crystals (one crystal form) and γ-type crystals (another crystal form). Form I crystals have a melting point of around 48°C and are a Q10 crystal form that exhibits characteristic peaks at diffraction angles (2θ±0.2°) of 3.1°, 18.7°, 19.0°, 20.2°, and 23.0° in powder X-ray diffraction using Cu-Kα. Form II crystals have a melting point of around 52°C and are a Q10 crystal that exhibits characteristic peaks at diffraction angles (2θ±0.2°) of 11.5°, 18.2°, 19.3°, 22.3°, 23.0°, and 33.3° in powder X-ray diffraction using Cu-Kα.

[0022] The crystal transition method is not particularly limited as long as it is a method that can perform a crystal transition by obtaining a product containing another crystal form from a raw material containing one crystal form. Specific examples of crystal transition methods include, for example, when the crystal transition is a transition from a metastable crystal (one crystal form) to a stable crystal (another crystal form), a solid-phase transition method in which a raw material (solid) containing one crystal form as the main raw material is subjected to heating and / or shearing in its original state (e.g., in powder form) without dissolving it in a solvent, thereby transitioning one crystal form to another and obtaining a product (solid) containing the other crystal form; a melt-mediated transition method in which a less stable crystal (one crystal form) in the main raw material melts upon heating, and then a more stable crystal (another crystal form) crystallizes continuously; and a solution-mediated transition method in which a less stable crystal (one crystal form) in the main raw material dissolves in a solvent, and then primary nuclei of the more stable crystal (another crystal form) are formed and grow, leading to crystallization. In chocolate, melt-mediated transition is used in tempering operations to transform type III or type IV crystals into type V crystals. Furthermore, in organic or inorganic substances where solid-phase transitions do not readily occur due to steric hindrance, solution-mediated transitions often proceed, and therefore solution-mediated transitions are frequently used. As crystal transition apparatuses, for example, when the crystal transition method is solid-phase transition, melt-mediated transition, or solution-mediated transition, a jacketed stirring reaction vessel having a tank for initiating the crystal transition reaction, a stirring device including a stirring blade for stirring the tank, a jacket for heating the tank, and a chiller for adjusting the tank temperature, as well as a stirring granulator, stirring dryer, and other stirring mechanisms and temperature control devices (for example, devices with a heat exchange mechanism in contact with the powder, such as an internal coil in addition to the jacket), are used. The processing time in solid-phase transition is not particularly limited and should be set appropriately considering the amount of raw material and the desired crystal transition rate. The processing atmosphere in the solid-phase transition method is not particularly limited; it can be under reduced pressure, increased pressure, or atmospheric pressure, as long as Q10 does not completely melt. The heating temperature in the solid-phase transition method is preferably as high as possible and at a temperature where Q10 does not completely melt, specifically in the range of 45°C to 50°C.Means of shearing in the solid-phase transition method include, for example, stirring devices including stirring blades to stir the inside of the tank where the crystal transition reaction occurs, as well as millstones, mortars, ball mills, bead mills, etc. The type of stirring blade is not particularly limited, but examples include anchor blades, screw blades, helical ribbon blades, wide paddle blades, multi-stage inclined paddle blades, three-way receding blades, and other stirring blades that have surfaces facing each other in close proximity to the wall. In the solid-phase transition method, the raw material may be subjected to either heating or shearing, or a combination of heating and shearing may be performed.

[0023] Conditional data is not particularly limited as long as it includes one or more types of data selected from one or more types of process data, one or more types of raw material data, and one or more types of equipment data. Process data is not particularly limited as long as it is a feature quantity that represents the crystal transition method performed in the crystal transition apparatus. For example, if the crystal transition method is a solid-phase transition method, a melt-mediated transition method, a solution-mediated transition method, etc., and the crystal transition apparatus is a jacketed stirred reaction vessel, a stirred granulator, a stirred dryer, or a reaction apparatus having a stirring mechanism and temperature control function, then examples of process data include processing time, composition of the processing atmosphere (gas phase inside the vessel), pressure inside the reaction vessel, temperature inside the vessel (heating temperature), jacket temperature (temperature of the heating medium), jacket flow rate (flow rate of the heating medium), chiller temperature (temperature of the chiller's cooling medium), and vibration spectrum, rotation speed, torque, and power consumption of the stirring blades, temperature and spectrum of the raw materials, vibration spectrum of the stirring motor, and data on human operation history. These individual data may include post-processing data such as the mean (e.g., mean over processing time), the sum of the means, the variance of the means, the variance (e.g., variance over processing time), the mean of the variances, the sum of the variances, and the median. For example, the sum of the mean and variance of the power consumption per unit time of the stirring blade over processing time, the mean and sum of the variances of the power consumption per unit time of the stirring blade over processing time, and the mean and spectral change of the raw material temperature over processing time may be used. Examples of the jacket temperature of the reaction vessel include the jacket inlet temperature, jacket outlet temperature, and jacket temperature difference (temperature difference between the jacket inlet and outlet temperatures). Furthermore, the temperature control device for the crystal transition apparatus is not limited to the jacket; for example, a device with a heat exchange mechanism in contact with the powder, such as an internal coil, may also be used.

[0024] The raw material data is not particularly limited as long as it is a feature that represents the raw material undergoing a crystal transition. Examples of raw material data include, for example, if the crystal transition is a transition from a metastable crystal to a stable crystal, data representing the density of the raw material, the hardness of the raw material, the orientation of the crystal of the raw material, the electrical properties of the raw material, the Young's modulus of the raw material, data representing the taste of the raw material, powder property data of the raw material (e.g., particle size, bulk density, angle of repose, etc.), data representing the purity of the raw material (e.g., content of auxiliary materials, water content, etc.), and optical property data of the raw material (e.g., refractive index, absorptivity, reflectivity, transmittance, characteristic values ​​related to diffraction, characteristic values ​​related to scattering, etc.).

[0025] The apparatus data is not particularly limited as long as it is a feature that represents the crystal transition apparatus. For example, if the crystal transition method is a solid-phase transition method, a melt-mediated transition method, a solution-mediated transition method, etc., and the crystal transition apparatus is a jacketed stirred reaction vessel, a stirred granulator, a stirred dryer, or a reaction apparatus with other stirring mechanisms and temperature control functions, then the apparatus data may include, for example, the size of the reaction vessel tank, the shape of the reaction vessel tank, the volume of the reaction vessel tank, the type of impeller, the size of the impeller, the shape of the impeller, and the clearance of the impeller.

[0026] The conditional data may include one or more types of environmental data in addition to the one or more types of data described above. Examples of environmental data include data on the temperature and humidity outside the crystal transition apparatus.

[0027] Crystal transition data are not particularly limited as long as they are feature quantities that express the characteristics of the crystal transition. Here, "characteristics of the crystal transition" refers to the characteristics of a product containing other crystal forms obtained from a raw material containing one crystal form in the crystal transition method, or the change in characteristics when the raw material changes into a product. Examples of crystal transition data include, for example, if the crystal transition is a transition from a metastable crystal to a stable crystal, the crystal transition rate in the product (increase in the crystal transition rate over processing time), data expressing the orientation of the crystals in the product, data expressing the oxidation stability of the product, data expressing reactivity, melting point, data expressing taste, data expressing deliquility, and data expressing absorbability to the human body, powder property data of the product (e.g., particle size, bulk density, angle of repose, etc.), optical property data of the product (e.g., refractive index, absorptivity, reflectivity, transmittance, characteristic values ​​related to diffraction, characteristic values ​​related to scattering, etc.), mechanical property data of the product (e.g., hardness, Young's modulus, shear modulus, tensile strength, bending strength, etc.), and electrical property data of the product.

[0028] The training data is a pair of condition data and crystal transition data, and is not particularly limited as long as it is recognized that there is a certain relationship, such as a correlation (hereinafter sometimes referred to as "correlation, etc."), between the condition data and the crystal transition data. For example, if the crystal transition is a transition from a metastable crystal to a stable crystal, the crystal transition method is a solid-phase transition method, and the crystal transition apparatus is a jacketed stirred reaction vessel, then the training data may be a pair of condition data including one or more types of data selected from the average temperature inside the vessel during processing time, the average temperature of the raw materials, the average temperature of the jacket inlet, the average temperature of the jacket outlet, the average temperature difference of the jacket, and the average temperature of the chiller, the average power consumption of the impeller during processing time, the sum and variance of the average power consumption per unit time of the impeller during processing time, and the average and sum of the variance of the power consumption per unit time of the impeller during processing time, and crystal transition data which is the increase in the crystal transition rate during processing time.

[0029] Next, as shown in Figure 2, multiple sets of training data are input to the estimation model generation device via the training data input unit (training data input step). In this case, the training data input unit of the estimation model generation device receives the input of multiple sets of training data and stores them in the storage unit.

[0030] Next, as shown in Figure 2, the estimation model generation unit of the estimation model generation device generates an estimation model using machine learning, which outputs the crystal transition data when the above condition data is input, using multiple sets of training data stored in the memory unit (estimation model generation step). The estimation model generation unit then stores the estimation model in the memory unit.

[0031] The machine learning algorithm used to generate the estimation model is not particularly limited as long as it can generate an estimation model capable of estimating crystal transition data. Examples include random forests, decision trees, support vector machines (SVMs), and neural networks, with random forests being preferred because they can generate estimation models capable of estimating crystal transition data with high accuracy.

[0032] Next, as shown in Figure 2, the estimated model output unit of the estimated model generation device outputs the estimated model stored in the memory unit to the estimation device described later (estimated model output step). In the estimation device described later, the estimated model input unit receives the outputted estimated model as input and stores it in the memory unit.

[0033] (Estimation device and estimation method) Figure 3 is a schematic diagram illustrating the configuration of the estimation device according to the first embodiment. Figure 4 is a schematic process diagram of the estimation method according to the first embodiment.

[0034] As shown in Figure 3, the estimation device according to the first embodiment includes an input unit including an estimation model input unit and a condition data input unit, a processing unit including an estimation unit, an output unit including an estimation result output unit, and a storage unit storing the estimation program and estimation model according to the first embodiment. The estimation device is composed of, for example, an information processing device. The information processing device is the same as the information processing device described in the above section (Estimation Model Generation Device and Estimation Model Generation Method). The estimation program causes the estimation method to be executed by making the estimation device function in the information processing device.

[0035] In the estimation method according to the first embodiment, an estimation device is used to estimate the characteristics of a crystal transition from condition data when a crystal transition is performed in a crystal transition apparatus. Specifically, in the estimation method, first, as shown in Figure 4, condition data is prepared that includes one or more types of data selected from one or more types of process data which are characteristic quantities that represent a crystal transition method performed in a crystal transition apparatus, one or more types of raw material data which are characteristic quantities that represent the raw materials for the crystal transition, and one or more types of apparatus data which are characteristic quantities that represent the crystal transition apparatus (condition data preparation step). The crystal transition apparatus, crystal transition, crystal transition method, and condition data are the same as the crystal transition apparatus, crystal transition, crystal transition method, and condition data described in the above section (Estimation Model Generation Apparatus and Estimation Model Generation Method).

[0036] Next, as shown in Figure 4, the condition data is input to the estimation device via the condition data input unit (condition data input step). In this case, the condition data input unit of the estimation device receives the condition data and stores it in the storage unit.

[0037] Next, as shown in Figure 4, the estimation unit of the estimation device uses the estimation model stored in the memory unit to obtain estimated values ​​of crystal transition data, which are feature quantities that represent the characteristics of the crystal transition, from the condition data stored in the memory unit (estimation step). The estimation unit then stores the estimated values ​​of the crystal transition data in the memory unit. The crystal transition data is the same as the crystal transition data described in the section on (estimation model generation device and estimation model generation method) above.

[0038] Next, as shown in Figure 4, the estimation result output unit of the estimation device outputs the estimated value of the crystal transition data stored in the memory unit (estimation result output step). In this way, the characteristics of the crystal transition are estimated from the condition data when performing a crystal transition in the crystal transition apparatus using the estimation device.

[0039] (Effects) According to the estimation model generation apparatus and estimation apparatus of the first embodiment, the characteristics of a crystal transition can be estimated from the condition data when a crystal transition is performed in a crystal transition apparatus. Similarly, according to the estimation model generation method and estimation method of the first embodiment, the characteristics of a crystal transition can be estimated from the condition data when a crystal transition is performed in a crystal transition apparatus.

[0040] [Second Embodiment] (Inverse Analysis Estimation Model Generation Apparatus and Inverse Analysis Estimation Model Generation Method) Figure 5 is a diagram illustrating the configuration of the inverse analysis estimation model generation apparatus according to the second embodiment. Figure 6 is a schematic process diagram of the inverse analysis estimation model generation method according to the second embodiment.

[0041] As shown in Figure 5, the inverse analysis estimation model generation device according to the second embodiment includes an input unit including a training data input unit, a processing unit including an inverse analysis estimation model generation unit, an output unit including an inverse analysis estimation model output unit, and a storage unit storing the inverse analysis estimation model generation program according to the second embodiment. The inverse analysis estimation model generation device is composed of, for example, an information processing device. The information processing device is the same as the information processing device according to the first embodiment. The inverse analysis estimation model generation program causes the information processing device to function, thereby executing the inverse analysis estimation model generation method.

[0042] In the method for generating an estimation model for inverse analysis according to the second embodiment, an estimation model for inverse analysis is generated using an estimation model generation device for inverse analysis. Specifically, in the method for generating an estimation model for inverse analysis, first, as shown in FIG. 6, one or more types of process data, which are characteristic quantities representing a crystal transition method in which crystal transition is performed by a crystal transition device, one or more types of raw material data, which are characteristic quantities representing raw materials for the crystal transition, and one or more types of device data, which are characteristic quantities representing the crystal transition device, are selected. A plurality of sets of teacher data, which is a set of condition data including two or more types of data and crystal transition data, which is a characteristic quantity representing the characteristics of the crystal transition, are prepared (teacher data preparation step).

[0043] The crystal transition device, crystal transition, and crystal transition method are the same as those of the crystal transition device, crystal transition, and crystal transition method according to the first embodiment, respectively. The process data, raw material data, and device data are the same as those of the process data, raw material data, and device data according to the first embodiment, respectively. The condition data is the same as the condition data according to the first embodiment, except that it includes two or more types of data selected from one or more types of process data, one or more types of raw material data, and one or more types of device data. The condition data may further include one or more types of environmental data in addition to the above two or more types of data. The environmental data is the same as the environmental data according to the first embodiment. The crystal transition data is the same as the crystal transition data according to the first embodiment. The teacher data is the same as the teacher data according to the first embodiment, except that the condition data includes two or more types of data as described above.

[0044] Next, as shown in FIG. 6, a plurality of sets of teacher data are input into an estimation model generation device via a teacher data input unit (teacher data input step). In this case, the teacher data input unit of the estimation model generation device receives the input of a plurality of sets of teacher data and stores them in a storage unit.

[0045] Next, as shown in FIG. 6, the inverse analysis estimation model generation unit of the inverse analysis estimation model generation device generates, by machine learning, an inverse analysis estimation model that outputs data of the remaining types of the condition data when inputting analysis data including data of some types of the condition data and the crystal transition data, using a plurality of sets of teacher data stored in the storage unit (inverse analysis estimation model generation step). Then, the inverse analysis estimation model generation unit stores the inverse analysis estimation model in the storage unit.

[0046] The data of some types of the condition data input to the estimation model is not particularly limited as long as it is data of some types among two or more types of data included in the condition data, and it may be one type of data or two or more types of data. The data of the remaining types of the condition data output by the estimation model is not particularly limited as long as it is data of the remaining types obtained by excluding the data of some types from two or more types of data included in the condition data, and it may be one type of data or two or more types of data. The combination of the data of some types of the condition data and the crystal transition data and the data of the remaining types of the condition data is not particularly limited as long as there is a certain relationship such as a correlation relationship (hereinafter sometimes referred to as "correlation relationship, etc.") between the data of some types of the condition data and the crystal transition data and the data of the remaining types of the condition data. The machine learning algorithm used for generating the inverse analysis estimation model is not particularly limited as long as it can generate an estimation model capable of estimating the data of the remaining types of the condition data. For example, it is the same as the algorithm according to the first embodiment.

[0047] Next, as shown in FIG. 6, the inverse analysis estimation model output unit of the inverse analysis estimation model generation device outputs the inverse analysis estimation model stored in the storage unit to an inverse analysis estimation device described later (inverse analysis estimation model output step). In the inverse analysis estimation device described later, the inverse analysis estimation model input unit receives the input of the output inverse analysis estimation model and stores it in the storage unit.

[0048] (Inverse Analysis Estimation Device and Inverse Analysis Estimation Method) Figure 7 is a schematic diagram illustrating the configuration of the inverse analysis estimation device according to the second embodiment. Figure 8 is a schematic process diagram of the inverse analysis estimation method according to the second embodiment.

[0049] As shown in Figure 7, the inverse analysis estimation device according to the second embodiment includes an input unit including an inverse analysis estimation model input unit and an analysis data input unit, a processing unit including an inverse analysis estimation unit, an output unit including an estimation result output unit, and a storage unit storing the inverse analysis estimation program and the inverse analysis estimation model according to the second embodiment. The inverse analysis estimation device is composed of, for example, an information processing device. The information processing device is the same as the information processing device described in the above section (Inverse Analysis Estimation Model Generation Device and Inverse Analysis Estimation Model Generation Method). The inverse analysis estimation program causes the inverse analysis estimation method to be executed by activating the inverse analysis estimation device in the information processing device.

[0050] In the inverse analysis estimation method according to the second embodiment, an inverse analysis estimation device is used to estimate condition data that makes the characteristics of the crystal transition a desired characteristic when performing a crystal transition in a crystal transition apparatus, by inverse analysis. Specifically, in the inverse analysis estimation method, first, as shown in Figure 8, data of some types of condition data including two or more types of data selected from one or more types of process data which are feature quantities that represent the crystal transition method in which the crystal transition is performed in a crystal transition apparatus, one or more types of raw material data which are feature quantities that represent the raw materials for the crystal transition, and one or more types of apparatus data which are feature quantities that represent the crystal transition apparatus, and crystal transition data which are feature quantities that represent the characteristics of the crystal transition (analysis data preparation step). At this time, as crystal transition data, crystal transition data (desired crystal transition data) which are feature quantities that represent the desired characteristics of the crystal transition are prepared. The crystal transition apparatus, crystal transition, crystal transition method, condition data, some types of condition data, and crystal transition data are the same as those described in the above section (Inverse Analysis Estimation Model Generation Apparatus and Inverse Analysis Estimation Model Generation Method).

[0051] Next, as shown in Figure 8, some types of the above-mentioned condition data and analysis data including the above-mentioned crystal transition data are input to the analysis estimation device via the analysis data input unit (analysis data input step). In this case, the analysis data input unit of the inverse analysis estimation device receives the analysis data input and stores it in the storage unit.

[0052] Next, as shown in Figure 8, the inverse analysis estimation unit of the inverse analysis estimation device uses the estimation model stored in the memory unit to obtain estimated values ​​for the remaining types of data of the condition data from the analysis data, including some types of condition data and crystal transition data, stored in the memory unit (inverse analysis estimation step). The inverse analysis estimation unit then stores the estimated values ​​for the remaining types of data of the condition data in the memory unit.

[0053] The remaining types of data for the condition data are the same as those described in the section on (Inverse Analysis Estimation Model Generation Device and Inverse Analysis Estimation Model Generation Method) above, and may consist of one type of data or two or more types of data. In the inverse analysis estimation process, the inverse analysis estimation unit may perform further optimization on the estimated values ​​of the remaining types of data obtained as described above, and the optimized estimated values ​​may be obtained as the estimated values ​​of the remaining types of data for the condition data and stored in the memory unit. Examples of optimization methods for multiple objective variables include Bayesian optimization and multi-objective optimization that selects the Pareto optimal solution. Examples of optimization methods for a single objective variable include GA (genetic algorithm), DE (Differential Evolution), PSO (Particle Swarm Optimization), Nelder-Mead method, Hooke-Jeeves pattern search method, BRKGA, etc.

[0054] Next, as shown in Figure 8, the estimation result output unit of the inverse analysis estimation device outputs estimated values ​​of the remaining types of condition data stored in the memory unit (estimation result output step). In this way, the inverse analysis estimation device is used to estimate condition data that will make the characteristics of the crystal transition desired when performing a crystal transition in the crystal transition apparatus, through inverse analysis.

[0055] (Effects) According to the inverse analysis estimation model generation apparatus and inverse analysis estimation apparatus of the second embodiment, condition data can be estimated by inverse analysis to make the characteristics of the crystal transition a desired characteristic when performing a crystal transition in a crystal transition apparatus. Similarly, according to the inverse analysis estimation model generation method and inverse analysis estimation method of the second embodiment, condition data can be estimated by inverse analysis to make the characteristics of the crystal transition a desired characteristic when performing a crystal transition in a crystal transition apparatus. Therefore, for example, in order to achieve a specific crystal transition rate, it becomes possible to control various data of the condition data based on the estimation results of various data of the condition data.

[0056] The following describes, in more detail, the estimation model generation apparatus and estimation apparatus, estimation model generation method and estimation method, and estimation model generation program and estimation program according to the embodiment, with reference to examples.

[0057] [Example] First, an example of the estimation model generation method according to the embodiment was carried out. First, multiple sets of training data were prepared (training data preparation step). For this, a jacketed stirring reaction vessel (made of stainless steel) was prepared as a crystal transition apparatus for performing a crystal transition by solid-phase transition method, which had a tank for initiating the crystal transition reaction (tank volume: 30 L), a stirring device including stirring blades (anchor blades) for stirring the inside of the tank, a jacket for heating the inside of the tank, and a chiller for adjusting the temperature inside the tank. Then, after putting 2200 g of commercially available Q10 Form I crystals (powder) into the tank of the reaction vessel, the inside of the tank was depressurized with a vacuum pump, the inside of the tank was heated with the jacket, and the temperature inside the tank was adjusted with the chiller to raise the temperature inside the tank to 45°C to 50°C, and the process of stirring the raw material inside the tank with the stirring blades at a stirring speed of 200 rpm to 300 rpm was continued for about 40 to 50 hours. In this way, in order to induce a crystal transition by solid-phase transition, a product (powder) containing Form II crystals was obtained by performing a combined heating and shearing treatment on a raw material containing Form I crystals.

[0058] In this crystal transition method, the temperature inside the tank, the temperature of the raw material, the jacket inlet temperature, the jacket outlet temperature, the jacket temperature difference, the chiller temperature, and the power consumption of the impeller were measured every minute. Furthermore, the product (raw material) inside the tank was sampled every four hours, and the crystal transition rate from Form I crystals to Form II crystals in the sample was measured. Then, from the measurements of the tank temperature, raw material temperature, jacket inlet temperature, jacket outlet temperature, jacket temperature difference, and chiller temperature taken every minute, the average of these values ​​for the past two hours (processing time) was calculated every two hours. From the measurements of the power consumption of the impeller taken every minute, the average power consumption of the impeller over the past two hours, the sum and variance of the average power consumption of the impeller per five minutes (unit time) over the past two hours, and the average and sum of the variance of the power consumption of the impeller per five minutes over the past two hours were calculated every two hours. Furthermore, by calculating the simple moving average of the crystal transition rate before and after each 4-hour interval, the crystal transition rate (measured value) [%] every 2 hours was obtained, and the increase in the crystal transition rate (measured value) [%] over the previous 2 hours (processing time) was calculated every 2 hours. As a result, one set of training data was created every 2 hours, corresponding to the duration of the crystal transition. This data consisted of condition data consisting of the average temperature inside the tank (reactor), the average temperature of the raw material (powder), the average jacket inlet temperature, the average jacket outlet temperature, the average jacket temperature difference (average ΔT), and the average chiller temperature over 2 hours, as well as the average power consumption of the stirring blade over 2 hours, the sum and variance of the average power consumption per 5 minutes (unit time) of the stirring blade over 2 hours, and the average and sum of the variance of the power consumption per 5 minutes of the stirring blade over 2 hours, and crystal transition data, which is the increase in the crystal transition rate over 2 hours. The number of sets corresponding to the duration of the crystal transition was created. By repeating the above operations of crystal transition, measurement, and creation of training data twice, multiple sets of training data corresponding to the duration of the crystal transition over a total of 89 hours were obtained.

[0059] Next, using multiple sets of training data, we generated an estimation model that outputs crystal transition data when conditional data is input, using machine learning (estimation model generation process). In this process, we used the data analysis platform provided by Dataku Inc. and performed two types of machine learning algorithms, using random forest and decision tree, respectively, to generate two estimation models, one using random forest and the other using decision tree. Hyperparameters were set appropriately in the random forest machine learning and the other in the decision tree machine learning. As described above, we have implemented one example of an estimation model generation method.

[0060] Next, an example of the estimation method according to the embodiment was performed using the two estimation models described above: random forest and decision tree. In this case, regardless of which estimation model was used, the estimation model was used to obtain estimated values ​​of multiple crystal transition data (increase in crystal transition rate over 2 hours (processing time)) from multiple condition data included in each of the multiple sets of training data. In this way, the increase in crystal transition rate over 2 hours was estimated for each of the multiple condition data.

[0061] A graph was obtained plotting the estimated increase in crystal transition rate [%] over two hours using a random forest estimation model against the measured value, and the feature importance of the conditional data in the random forest estimation model was also obtained. Figure 9 is a graph showing the feature importance of the conditional data in the random forest estimation model. In addition, a graph was obtained plotting the estimated increase in crystal transition rate [%] over two hours using a decision tree estimation model against the measured value, and the feature importance of the conditional data in the decision tree estimation model was also obtained. Figure 10 is a graph showing the feature importance of the conditional data in the decision tree estimation model. From the estimation results of the random forest and decision tree estimation models, RMSE (root mean square error) and R ２ The result of calculating the coefficient of determination showed that the estimated model of the random forest had an RMSE of 2.101, R ２The result was 0.822, and in the decision tree estimation model, the RMSE was 2.326, R ２ The result was 0.782.

[0062] The embodiments of the estimation model generation apparatus and estimation apparatus, estimation model generation method and estimation method, and estimation model generation program and estimation program according to the present invention have been described in detail above. However, the present invention is not limited to the embodiments described above, and various design modifications can be made without departing from the spirit of the invention as described in the claims. All publications, patents and patent applications referenced herein are incorporated herein by direct reference.

Claims

1. An estimation model generation device characterized by comprising: a training data input unit that accepts training data which is a pair of condition data including one or more data selected from one or more types of process data which are feature quantities that represent a crystal transition method in which a crystal transition is performed in a crystal transition apparatus, one or more types of raw material data which are feature quantities that represent the raw materials for the crystal transition, and one or more types of apparatus data which are feature quantities that represent the crystal transition apparatus, and crystal transition data which are feature quantities that represent the characteristics of the crystal transition; and an estimation model generation unit that generates an estimation model by machine learning using the training data which outputs the crystal transition data when the condition data is input.

2. An estimation device comprising: a condition data input unit that accepts input of condition data including one or more types of data selected from one or more types of process data which are feature quantities representing a crystal transition method that performs a crystal transition in a crystal transition apparatus, one or more types of raw material data which are feature quantities representing the raw materials for the crystal transition, and one or more types of apparatus data which are feature quantities representing the crystal transition apparatus; and an estimation unit that obtains estimated values ​​of crystal transition data which are feature quantities representing the characteristics of the crystal transition from the condition data by using an estimation model, wherein the estimation model is an estimation model generated by machine learning using training data which is a pair of the condition data and the crystal transition data, so as to output the crystal transition data when the condition data is input.

3. A method for generating an estimation model, comprising the steps of: preparing training data which is a pair of condition data including one or more data selected from one or more types of process data which are feature quantities that represent a crystal transition method in which a crystal transition is performed in a crystal transition apparatus, one or more types of raw material data which are feature quantities that represent the raw materials for the crystal transition, and one or more types of apparatus data which are feature quantities that represent the crystal transition apparatus, and crystal transition data which are feature quantities that represent the characteristics of the crystal transition; and generating an estimation model by machine learning using the training data which outputs the crystal transition data when the condition data is input.

4. An estimation method comprising the steps of: preparing condition data including one or more types of data selected from one or more types of process data which are feature quantities representing a crystal transition method in which a crystal transition is performed using a crystal transition apparatus, one or more types of raw material data which are feature quantities representing the raw materials for the crystal transition, and one or more types of apparatus data which are feature quantities representing the crystal transition apparatus; and obtaining an estimated value of crystal transition data which are feature quantities representing the characteristics of the crystal transition from the condition data by using an estimation model, wherein the estimation model is an estimation model generated by machine learning using training data which is a pair of the condition data and the crystal transition data, so as to output the crystal transition data when the condition data is input.

5. An estimation model generation program characterized by causing an information processing device to execute the estimation model generation method described in claim 3.

6. An estimation program characterized by causing an information processing device to execute the estimation method described in claim 4.

7. A method for generating an estimation model, comprising: a step of preparing training data which is a pair of condition data, which includes two or more types of data selected from one or more types of process data which are feature quantities that represent a crystal transition method in which a crystal transition is performed in a crystal transition apparatus, one or more types of raw material data which are feature quantities that represent the raw materials for the crystal transition, and one or more types of apparatus data which are feature quantities that represent the crystal transition apparatus, and crystal transition data which are feature quantities that represent the characteristics of the crystal transition; and a step of generating an estimation model by machine learning using the training data which outputs the remaining types of data of the condition data when analysis data including some types of data of the condition data and the crystal transition data is input.

8. An estimation method comprising: a step of preparing some types of condition data, which include two or more types of data selected from one or more types of process data, which are feature quantities representing a crystal transition method in which a crystal transition is performed in a crystal transition apparatus; one or more types of raw material data, which are feature quantities representing the raw materials for the crystal transition; and one or more types of apparatus data, which are feature quantities representing the crystal transition apparatus; and crystal transition data, which are feature quantities representing the characteristics of the crystal transition; and a step of obtaining estimated values ​​of the remaining types of condition data from analysis data including some types of condition data and the crystal transition data by using an estimation model, wherein the estimation model is an estimation model generated by machine learning using training data which are a pair of condition data and crystal transition data, so as to output the remaining types of condition data when analysis data including some types of condition data and the crystal transition data is input.

Images