Search method for colorant material, information processing apparatus, and non-transitory computer-readable recording medium

Abstract

The present invention improves a search technique for a pigment material. A search method for a pigment material, executed by an information processing device (10), includes the steps of training a VAE encoder (3) and a VAE decoder (4) respectively, the VAE encoder (3) outputs a latent variable on a latent space (5) corresponding to pigment material information as input with pigment material information expressed by a prescribed notation, the VAE decoder (4) outputs pigment material information expressed by a prescribed notation as input with an arbitrary latent variable on the latent space (5), and determining a desired pigment material satisfying all of a plurality of physical properties among the plurality of physical properties based on data related to the VAE encoder (3), the VAE decoder (4), and the plurality of physical properties of the pigment material.

Description

Technical Field

[0001] This disclosure relates to a method for searching pigment materials, an information processing apparatus, and a non-transitory computer-readable recording medium. This application claims priority based on Japanese Patent Application No. 2022-050780, filed on March 25, 2022, the contents of which are incorporated herein by reference. Background Technology

[0002] Techniques for searching for new materials using machine learning are known in the past. For example, techniques for searching for new agents using Bayesian optimization have been proposed in Non-Patent Literature 1 and Non-Patent Literature 2.

[0003] Existing technical documents

[0004] Patent documents

[0005] Non-patent literature 1: Rafael Gomez-Bombarelli et al. "Automatic Chemical Design Using a Data-Driven Continuous Representation of Molecules" ACS Cent. Sci. 2018, 4, 2, 268-276.

[0006] Non-patent document 2: Ryan-Rhys Griffiths et.al. "Constrained Bayesian Optimization for Automatic Chemical Design using Variational Autoencoders" Chem. Sci. 2020, 11, 577-586. Summary of the Invention

[0007] The problem the invention aims to solve

[0008] The substance search techniques described in Non-Patent Literature 1 and Non-Patent Literature 2 are for searching pharmaceuticals, and the prior art has not yet considered searching for new pigment materials. In other words, there is room for improvement in the pigment material search techniques.

[0009] The purpose of this disclosure, which was made in view of such circumstances, is to improve the search technology for pigment materials.

[0010] Solution for solving the problem

[0011] One embodiment of this disclosure relates to a pigment material search method executed by an information processing device. The search method includes the following steps: a training step, in which a VAE encoder and a VAE decoder are trained respectively. The VAE encoder takes pigment material information represented by a predetermined labeling method as input and outputs latent variables in a latent space corresponding to the pigment material information. The VAE decoder takes any latent variable in the latent space as input and outputs pigment material information represented by the predetermined labeling method. A determination step, based on the VAE encoder, the VAE decoder, and data related to various physical properties of the pigment material, determines a pigment material that satisfies the desired properties of all of the various physical properties.

[0012] In addition, one embodiment of the pigment material search method disclosed herein further includes the following steps: training a property prediction model that takes arbitrary variables in the latent space as input and outputs predicted values ​​of the various properties, wherein data related to the various properties of the pigment material are used in the training of the property prediction model, and in the determination step, searching for latent variables corresponding to the desired pigment material by optimization processing based on the property prediction model.

[0013] In addition, in one embodiment of the pigment material search method of this disclosure, the desired pigment material is determined by inputting a portion of the data relating to various physical properties of the pigment material into the VAE encoder and the VAE decoder.

[0014] In addition, in one embodiment of the pigment material search method of this disclosure, the desired pigment material is determined by inputting a portion of the data related to various physical properties of the pigment material into the physical property prediction model.

[0015] In addition, in one embodiment of the pigment material search method, the desired pigment material is determined by inputting data related to various physical properties of the pigment material into the VAE encoder and the VAE decoder.

[0016] Furthermore, in one embodiment of the pigment material search method disclosed herein, the various physical properties include information related to hue.

[0017] Furthermore, in one embodiment of the pigment material search method disclosed herein, the various physical properties include information related to stability.

[0018] In addition, in one embodiment of the pigment material search method, the training step uses actual condition data of the composition used as a pigment material and actual condition data of the composition used in applications other than pigment materials to train the VAE encoder and the VAE decoder.

[0019] In addition, in one embodiment of the pigment material search method disclosed herein, the optimization process is a Bayesian optimization process.

[0020] Furthermore, in one embodiment of the pigment material search method disclosed herein, at least a portion of the data related to the various physical properties is continuous value information, discrete value information, or classification information.

[0021] In addition, in one embodiment of the pigment material search method disclosed herein, the pigment material is a dichroic pigment material, and the various physical properties include the maximum absorption wavelength and the dichroic ratio.

[0022] Furthermore, one embodiment of this disclosure relates to an information processing apparatus for searching pigment materials, which includes a control unit. The control unit performs the following processing: training a VAE encoder and a VAE decoder, wherein the VAE encoder takes pigment material information represented by a predetermined labeling method as input and outputs latent variables in a latent space corresponding to the pigment material information; the VAE decoder takes any latent variable in the latent space as input and outputs pigment material information represented by the predetermined labeling method; and based on data related to the VAE encoder, the VAE decoder, and various physical properties of the pigment material, determines a pigment material that satisfies the desired properties of all of the various physical properties.

[0023] Furthermore, one embodiment of this disclosure involves a non-transitory computer-readable recording medium that stores commands for searching pigment materials. When executed by a processor, these commands cause the processor to perform the following processes: training a VAE encoder and a VAE decoder, wherein the VAE encoder takes pigment material information represented by a prescribed notation as input and outputs latent variables in a latent space corresponding to the pigment material information; and the VAE decoder takes any latent variable in the latent space as input and outputs pigment material information represented by the prescribed notation; and determining, based on data related to the VAE encoder, the VAE decoder, and various physical properties of the pigment material, the desired pigment material satisfying all of the various physical properties.

[0024] The effects of the invention

[0025] According to one embodiment of the present disclosure, the pigment material search method, information processing apparatus, and non-transitory computer-readable recording medium can improve the pigment material search technology. Attached Figure Description

[0026] Figure 1 This is a diagram illustrating an outline of a pigment material search technique according to one embodiment of this disclosure.

[0027] Figure 2 This is a flowchart illustrating an outline of a method for searching pigment materials according to one embodiment of this disclosure.

[0028] Figure 3 This is a block diagram illustrating the general structure of an information processing apparatus for performing a method for searching pigment materials according to an embodiment of the present disclosure.

[0029] Figure 4 This is a diagram illustrating a first specific example of a method for searching pigment materials according to an embodiment of this disclosure.

[0030] Figure 5 This is a flowchart illustrating a first specific example of a method for searching pigment materials according to an embodiment of this disclosure.

[0031] Figure 6 This is a diagram illustrating an example of the penalty involved in the maximum absorption wavelength.

[0032] Figure 7 This is a diagram showing variations and comparative examples of the penalty involved in the maximum absorption wavelength.

[0033] Figure 8 This is a graph showing the differences in output results corresponding to each penalty involved with the maximum absorption wavelength.

[0034] Figure 9 This is a diagram illustrating an example of the penalty involved in dichroism ratio.

[0035] Figure 10 This is a diagram illustrating one example of the penalties involved in lightfastness.

[0036] Figure 11 This is a diagram illustrating a second specific example of a method for searching pigment materials according to an embodiment of this disclosure.

[0037] Figure 12 This is a flowchart illustrating a second specific example of a method for searching pigment materials according to an embodiment of this disclosure.

[0038] Figure 13This is a diagram illustrating a variation of a second specific example of a method for searching pigment materials according to an embodiment of this disclosure.

[0039] Figure 14 This is a figure illustrating a third specific example of a method for searching pigment materials according to an embodiment of this disclosure.

[0040] Figure 15 This is a flowchart illustrating a third specific example of a method for searching pigment materials according to an embodiment of this disclosure. Detailed Implementation

[0041] Hereinafter, the pigment material search technique involved in the embodiments of this disclosure will be described with reference to the accompanying drawings.

[0042] In each figure, the same or equivalent parts are labeled with the same reference numerals. In the description of this embodiment, the description of the same or equivalent parts is appropriately omitted or simplified.

[0043] First, refer to Figure 1 and Figure 2 This section provides an overview of the present embodiment. The pigment material search technique described in this embodiment utilizes... Figure 1 The experimental data 1 and the public database 2 (public DB 2) are shown. The pigment material searched using this search technique is, for example, a dichroic dye, but is not limited to this. Experimental data 1 and public DB 2 contain information related to the pigment material (hereinafter also referred to as pigment material information) and actual condition data related to the pigment material. Furthermore, an example is shown where both experimental data 1 and public DB 2 are used in the pigment material search technique according to this embodiment, but it is not limited to this. Alternatively, only one of experimental data 1 and public DB 2 may be used in the pigment material search technique according to this embodiment.

[0044] Pigment material information includes information related to the molecular structure of the pigment material. This pigment material information is represented using a prescribed notation method. In this embodiment, the prescribed notation method is described as SMILES (Simplified molecular-input line-entry system), but it is not limited to this.

[0045] The actual condition data related to the pigment material refers to the data related to the various physical properties of the pigment material (hereinafter also referred to as physical property data), including data obtained through experiments, etc. Furthermore, the pigment material search technology involved in this embodiment is a method executed by the information processing device 10, which includes a VAE encoder 3 and a VAE decoder 4. In other words, the pigment material search technology involved in this embodiment uses a Variational Autoencoder (VAE).

[0046] VAE Encoder 3 is a learning model that takes pigment material information as input and outputs latent variables that correspond to the pigment material information. The latent variables are variables in the latent space 5. Figure 1 The latent space 5 is represented by two-dimensional coordinates, but the dimension of the latent space 5 is not limited to 2. The dimension of the latent space can also be 3 or higher. The VAE decoder 4 is a learning model that takes any latent variable in the latent space 5 as input and outputs pigment material information.

[0047] like Figure 2 As shown, the VAE encoder 3 and VAE decoder 4 are trained based on pigment material information (step S10). Then, in the pigment material search technique according to this embodiment, the pigment material that satisfies the expectation of all physical properties among multiple physical properties is determined based on the trained VAE encoder 3 and VAE decoder 4, as well as the physical property data (step S20). Such determination processing can be performed using various methods. For example, in step S20, a prediction model trained using physical property data can also be used. Such a prediction model is a model that takes a latent variable in the latent space 5 as input and outputs a predicted value of the physical property corresponding to that latent variable. When using a prediction model, the pigment material that satisfies the expectation of all physical properties among multiple physical properties is determined through optimization processing based on the prediction model. Furthermore, as described later, the pigment material that satisfies the expectation of all physical properties among multiple physical properties can also be determined by a method that does not use the prediction model.

[0048] Thus, according to the pigment material search technology involved in this embodiment, pigment materials that satisfy any physical property can be determined using the VAE encoder 3, the VAE decoder 4, and physical property data. Therefore, the pigment material search technology is improved in terms of being able to search for pigment materials that satisfy the expectation of all physical properties among a variety of physical properties.

[0049] (Structure of an information processing device)

[0050] Next, the various structures of the information processing apparatus 10 will be described in detail. The information processing apparatus 10 is any device used by a user. As the information processing apparatus 10, it can be, for example, a personal computer, a server computer, a general-purpose electronic device, or a special-purpose electronic device.

[0051] like Figure 3 As shown, the information processing device 10 includes a control unit 11, a storage unit 12, an input unit 13, and an output unit 14.

[0052] The control unit 11 includes at least one processor, at least one dedicated circuit, or a combination thereof. The processor is a general-purpose processor such as a CPU (central processing unit) or GPU (graphics processing unit), or a dedicated processor for specific processing. The dedicated circuit is, for example, an FPGA (field-programmable gate array) or an ASIC (application-specific integrated circuit). The control unit 11 controls the various parts of the information processing device 10 while performing processing related to the operation of the information processing device 10.

[0053] The storage unit 12 includes at least one semiconductor memory, at least one magnetic memory, at least one optical memory, or a combination of at least two of these. The semiconductor memory is, for example, RAM (random access memory) or ROM (read-only memory). RAM is, for example, SRAM (static random access memory) or DRAM (dynamic random access memory). ROM is, for example, EEPROM (electrically erasable programmable read-only memory). The storage unit 12 functions as, for example, a main storage device, an auxiliary storage device, or a cache. The storage unit 12 stores data for the operation of the information processing device 10 and data obtained through the operation of the information processing device 10.

[0054] The input unit 13 includes at least one input interface. The input interface may be, for example, a physical key, a capacitive key, an indicator device, or a touchscreen integrated with the display. Alternatively, the input interface may be, for example, a microphone for receiving voice input, or a camera for receiving gesture input. The input unit 13 accepts data input for the operation of the information processing device 10. The input unit 13 may also be connected to the information processing device 10 as an external input device, replacing the information processing device 10 having its own input unit 13. As a connection method, for example, any method such as USB (Universal Serial Bus), HDMI (High-Definition Multimedia Interface), or Bluetooth (registered trademark) can be used.

[0055] The output unit 14 includes at least one output interface. The output interface may be, for example, a display that outputs image information. The display may be, for example, an LCD (liquid crystal display) or an organic EL (electroluminescence) display. The output unit 14 displays and outputs data obtained through the operation of the information processing device 10. The output unit 14 may also be connected to the information processing device 10 as an external output device, replacing the information processing device 10 in having an output unit 14. As a connection method, any method such as USB, HDMI (registered trademark), or Bluetooth (registered trademark) can be used.

[0056] The functions of the information processing device 10 are implemented by a processor, which is equivalent to the information processing device 10, executing the program involved in this embodiment. That is, the functions of the information processing device 10 are implemented by software. The program enables the computer to function as the information processing device 10 by causing the computer to perform the actions of the information processing device 10. In other words, the computer functions as the information processing device 10 by executing the actions of the information processing device 10 according to the program.

[0057] In this embodiment, the program can be recorded on a computer-readable recording medium. Computer-readable recording media include non-transitory computer-readable media, such as magnetic recording devices, optical discs, optical-magnetic recording media, or semiconductor memory. For example, the program can be circulated by selling, transferring, or renting portable recording media such as DVDs (digital versatile discs) or CD-ROMs (compact disc read-only memory) containing the program. Alternatively, the program can be circulated by storing it on a storage device on an external server and sending the program from the external server to other computers. Furthermore, the program can also be provided as a program product.

[0058] Some or all of the functions of the information processing device 10 can also be implemented by a dedicated circuit equivalent to the control unit 11. That is, some or all of the functions of the information processing device 10 can also be implemented by hardware.

[0059] In this embodiment, the storage unit 12 stores experimental data 1, public database 2, VAE encoder 3, and VAE decoder 4.

[0060] As described above, Experimental Data 1 and Public DB 2 contain information about the pigment material and its physical properties. For example, the physical properties data may also include information related to hue. Additionally, the physical properties data may also include information related to stability. Furthermore, in the case where the pigment material is a dichroic pigment, the hue-related information may include, for example, the maximum absorption wavelength and the dichroic ratio. Additionally, in the case where the pigment material is a dichroic pigment, the stability-related information may include, for example, lightfastness. Hereinafter, in this embodiment, the case where the pigment material is a dichroic pigment will be described as an example. Furthermore, the case where the aforementioned various physical properties include the maximum absorption wavelength, dichroic ratio, and lightfastness will be explained.

[0061] Furthermore, experimental data 1, publicly available DB 2, VAE encoder 3, and VAE decoder 4 can also be stored on an external device different from the information processing device 10. In this case, the information processing device 10 can also have an external communication interface. The communication interface can be either a wired or wireless communication interface. In the case of wired communication, the communication interface is, for example, a LAN interface or a USB. In the case of wireless communication, the communication interface is, for example, an interface supporting mobile communication standards such as LTE, 4G, or 5G, or an interface supporting short-range wireless communication such as Bluetooth (registered trademark). The communication interface can receive data for the operation of the information processing device 10, and can also transmit data obtained through the operation of the information processing device 10.

[0062] (First specific example)

[0063] Reference Figure 4 and Figure 5 This will be used to illustrate a first specific example and operation of a method for searching pigment materials according to an embodiment of the present disclosure.

[0064] like Figure 4 As shown, the first specific example includes a VAE encoder 103, a VAE decoder 104, and a property prediction model 106. That is, in this case, the storage unit 12 stores the VAE encoder 103, the VAE decoder 104, and the property prediction model 106.

[0065] The VAE encoder 103 is a learning model that takes pigment material information as input and outputs latent variables equivalent to that information. As described above, the pigment material information is data labeled with SMILES. The latent variables output by the VAE encoder 103 are variables in the latent space 105. Figure 4 The latent space 105 is represented by two-dimensional coordinates, but the dimension of the latent space 105 is not limited to 2. The dimension of the latent space can also be 3 or higher. The VAE encoder 103 is trained based on pigment material information.

[0066] VAE decoder 104 is a learning model that takes any latent variable in the latent space 105 as input and outputs pigment material information. VAE decoder 104 is trained based on pigment material information.

[0067] The property prediction model 106 is a learning model that takes latent variables in the latent space 105 as input and outputs predicted property values ​​corresponding to those latent variables. The property prediction model 106 is trained based on property data. This training is performed by correcting the error between the output of the property prediction model 106 for a given input and the property data (learning data). Here, the property prediction model 106 consists of multiple prediction models corresponding to each property. That is, in this embodiment, the property prediction model 106 consists of prediction models related to the maximum absorption wavelength, prediction models related to the dichroic ratio, and prediction models related to lightfastness. The prediction models related to the maximum absorption wavelength, the dichroic ratio, and lightfastness are trained based on property data. In other words, the prediction models related to the maximum absorption wavelength, the dichroic ratio, and lightfastness are trained respectively based on actual data regarding the maximum absorption wavelength, dichroic ratio, and lightfastness of the pigment material.

[0068] Figure 5 This is a flowchart illustrating a first specific example of a method for searching pigment materials according to an embodiment of this disclosure.

[0069] Step S110: The control unit 11 of the information processing device 10 trains the VAE encoder 103 and the VAE decoder 104 based on the pigment material information.

[0070] Step S120: The control unit 11 trains the property prediction model 106 based on the property data. That is, the control unit 11 trains prediction models related to the maximum absorption wavelength, dichroism ratio, and lightfastness of the pigment material, respectively, based on actual data regarding the maximum absorption wavelength, dichroism ratio, and lightfastness. Furthermore, the training of the property prediction model 106 can be performed in parallel with the training of the VAE encoder 103 and VAE decoder 104 in step S110.

[0071] Step S130: The control unit 11 determines the latent variables corresponding to the desired pigment material through optimization processing based on the physical property prediction model 106. Such optimization processing includes Bayesian optimization processing. That is, for example, the control unit 11 determines the latent variables corresponding to the desired pigment material through Bayesian optimization processing.

[0072] Here, in the Bayesian optimization process, the control unit 11 searches for latent variables that minimize the comprehensive index value determined by the sum of penalties corresponding to the predicted physical properties. When the various physical properties are maximum absorption wavelength, dichroism ratio, and lightfastness, this total penalty is determined by the sum of the penalties corresponding to the predicted maximum absorption wavelength, the predicted dichroism ratio, and the predicted lightfastness. Each penalty is explained below.

[0073] Figure 6 This is a diagram illustrating an example of the penalty corresponding to the predicted value of the maximum absorption wavelength. The penalty corresponding to the predicted value of the maximum absorption wavelength is determined by a function that increases accordingly with the increase in deviation from the first target range. For example, in the case of cyan pigment, the first target range is 600 nm or more and 700 nm or less. Additionally, in the case of magenta pigment, the first target range is 500 nm or more and 600 nm or less. Furthermore, in the case of yellow pigment, the first target range is 400 nm or more and 500 nm or less. Figure 6 The penalty shown represents the penalty associated with the maximum absorption wavelength of the cyan pigment. Figure 6 The penalty is zero when the maximum absorption wavelength is at the center of the first target range (650 nm in this case), and increases as a function of a fixed gradient when the maximum absorption wavelength deviates from the center. This gradient is set to penalize a first threshold value when the maximum absorption wavelength deviates from the first target range by a specified value. Figure 6 In this context, the specified value is 20nm. Additionally, the first threshold is 50. That is, the gradient is penalized with 50 at 580nm, which is 20nm lower than the lower limit of the first target range. Furthermore, the gradient is penalized with 50 at 720nm, which is 20nm higher than the upper limit of the first target range. Moreover, in... Figure 6 An example with a fixed gradient (an example with a penalty as a linear function) is shown, but this is not a limitation. The penalty can also be a higher-order function. Additionally, an example with a first threshold of 50 is shown, but this is not a limitation. The first threshold could also be, for example, 20. The first threshold should be appropriately determined based on the penalty involved in the dichroism ratio.

[0074] exist Figure 7 The diagram shows variations and comparative examples of the penalty corresponding to the predicted maximum absorption wavelength. It is possible to employ... Figure 7 The functions f0 to f5 shown, with f1 to f5 serving as penalties corresponding to the predicted maximum absorption wavelength. f1 and Figure 6 The penalties shown are the same. f2 to f5 are variations of f1. Similar to f1, f2 to f5 are also functions determined by the increase in penalty corresponding to the increase in deviation relative to the first target range. On the other hand, f0 is a comparative example of the penalty related to the maximum absorption wavelength. With respect to f0, the penalty remains constant even as the deviation relative to the first target range increases.

[0075] Figure 8 This is a graph showing the differences in the output results corresponding to each penalty and the predicted value of the maximum absorption wavelength. For example... Figure 8 As shown, when the penalty corresponding to the predicted value of the maximum absorption wavelength is determined by f0, a relatively large number of candidates that do not meet the maximum absorption wavelength are found among the pigment materials searched as a result of Bayesian optimization. Therefore, it is preferable that the penalty corresponding to the predicted value of the maximum absorption wavelength is determined by a function, such as f1 to f5, that increases the penalty in accordance with the increase in deviation from the first target range.

[0076] Figure 9 This is a diagram illustrating an example of the penalty corresponding to the predicted value of the dichroism ratio. The penalty corresponding to the predicted value of the dichroism ratio is determined by a function that decreases accordingly with an increase in the dichroism ratio. Furthermore, the penalty corresponding to the predicted value of the dichroism ratio is determined to be a minimum value equal to a second threshold. The second threshold is, for example, 20. Figure 9 In this context, the penalty corresponding to the predicted dichroism ratio is determined by a linear function. Furthermore, the gradient of this penalty is, for example, 1. Additionally, in... Figure 9 The example shown is one with a fixed gradient (an example where the penalty is a linear function), but it is not limited to this. The penalty can also be a higher-order function.

[0077] Figure 10 This is a graph illustrating an example of the penalty corresponding to the predicted value of lightfastness. The penalty corresponding to the predicted value of lightfastness is determined by a function such as the sigmoid function. In this embodiment, this function includes the sigmoid function, the step function, and functions with properties similar to the sigmoid function (cumulative normal distribution function, Gompertz function, Gudermannian function, etc.). Figure 10 The penalty shown is a step function, which is 0 when the lightfastness index is within the target range (lightfastness index (1000-hour OK probability) is 50%–100%). Furthermore, when the lightfastness index is outside the target range (0%–50%), the penalty is determined to be a third threshold (20 in this case). The third threshold is appropriately determined based on the penalty related to the dichroism ratio. The third threshold can be the same as or different from the first threshold.

[0078] The control unit 11 searches for a latent variable that minimizes the comprehensive index value based on the total penalty determined as described above (hereinafter also referred to as the comprehensive index value). The pigment material corresponding to this latent variable is equivalent to the desired pigment material. Furthermore, in this embodiment, the case where the comprehensive index value is small is defined as the desired case, but it is not limited to this. For example, the case where the comprehensive index value is large can also be defined as the desired case based on the penalty determination method. In this case, the control unit 11 searches for a latent variable that maximizes the comprehensive index value.

[0079] Step S140: The control unit 11 outputs the desired pigment material corresponding to the latent variables determined in step S130 via the VAE decoder 104. Specifically, the control unit 11 confirms whether the pigment material information output by the VAE decoder 104 is suitable for the SMILES syntax rules. If the output pigment material information is suitable for the SMILES syntax rules, the control unit 11 outputs the pigment material information as the desired pigment material.

[0080] Thus, according to this embodiment, the VAE encoder 103 and VAE decoder 104 trained based on pigment material information, and the physical property prediction model 106 trained based on physical property data, can be used to search for pigment materials that satisfy the expectations of all physical properties among a variety of physical properties.

[0081] (Second specific example)

[0082] Reference Figure 11 and Figure 12 This will be used to illustrate a second specific example and operation of a method for searching pigment materials according to an embodiment of the present disclosure.

[0083] like Figure 11 As shown, the second specific example includes a VAE encoder 203, a VAE decoder 204, and a property prediction model 206. That is, in this case, the storage unit 12 stores the VAE encoder 203, the VAE decoder 204, and the property prediction model 206.

[0084] The VAE encoder 203 is a learning model that takes a portion of physical property data and pigment material information as input and outputs a latent variable equivalent to the pigment material information. The portion of physical property data refers to data relating to any of a variety of physical properties. Such data may be continuous value information or non-continuous value information. In other words, such data may be any of continuous value information, discrete value information, or categorical information. In this embodiment, it is described that such a portion of physical property data is data relating to lightfastness and is categorical information (hereinafter also referred to as categorical property).

[0085] As described above, the pigment material information is data tagged with SMILES. The latent variables output by the VAE encoder 203 are variables in the latent space 205. Figure 4 The latent space 205 is represented by two-dimensional coordinates, but the dimension of the latent space 205 is not limited to 2. The dimension of the latent space can also be 3 or higher. The VAE encoder 203 is trained based on classification properties and pigment material information. That is, in this embodiment, the VAE encoder 203 is trained based on data related to lightfastness and pigment material information.

[0086] VAE decoder 204 is a learning model that takes classification properties and arbitrary latent variables in the latent space 205 as inputs and outputs pigment material information. VAE decoder 204 is trained based on classification properties and pigment material information.

[0087] The property prediction model 206 is a learning model that takes the categorical properties and latent variables in the latent space 205 as inputs and outputs the predicted property values ​​corresponding to those latent variables. The property prediction model 206 is trained based on the categorical properties and property data. Here, the property prediction model 206 consists of multiple prediction models corresponding to each property. That is, in this embodiment, the property prediction model 206 consists of a prediction model related to the maximum absorption wavelength and a prediction model related to the dichroic ratio. The prediction model related to the maximum absorption wavelength and the prediction model related to the dichroic ratio are trained based on the categorical properties and corresponding property data. In other words, the prediction model related to the maximum absorption wavelength is trained based on the actual data of the maximum absorption wavelength of the pigment material and the actual data of its lightfastness. Furthermore, the prediction model related to the dichroic ratio is trained based on the actual data of the dichroic ratio and the actual data of its lightfastness.

[0088] Figure 12 This is a flowchart illustrating a second specific example of a method for searching pigment materials according to an embodiment of this disclosure.

[0089] Step S210: The control unit 11 of the information processing device 10 trains the VAE encoder 203 and VAE decoder 204 based on pigment material information and classification properties.

[0090] Step S220: The control unit 11 trains the property prediction model 206 based on the property data and classified properties. That is, the control unit 11 trains the prediction models related to the maximum absorption wavelength and the dichroism ratio, which constitute the property prediction model 206, based on the actual data of the maximum absorption wavelength and lightfastness of the pigment material, and the actual data of the dichroism ratio and lightfastness. Furthermore, the training of the property prediction model 206 can be performed in parallel with the training of the VAE encoder 203 and VAE decoder 204 in step S210.

[0091] Step S230: The control unit 11 determines the latent variables corresponding to the desired pigment material through optimization processing based on the physical property prediction model 206. This optimization processing includes Bayesian optimization. That is, for example, the control unit 11 determines the latent variables corresponding to the desired pigment material through Bayesian optimization. In the Bayesian optimization, the control unit 11 searches for latent variables that minimize the comprehensive index value determined by the sum of penalties corresponding to the physical property prediction values. The penalty setting method is the same as that in the first specific example.

[0092] Step S240: The control unit 11 outputs the desired pigment material corresponding to the latent variables determined in step S230 via the VAE decoder 204. Specifically, the control unit 11 confirms whether the pigment material information output by the VAE decoder 204 is suitable for the SMILES syntax rules. If the output pigment material information is suitable for the SMILES syntax rules, the control unit 11 outputs the pigment material information as the desired pigment material.

[0093] Thus, according to this embodiment, a VAE encoder 203 and VAE decoder 204 trained based on pigment material information and classification properties, and a property prediction model 206 trained based on property data and classification properties, can be used to explore pigment materials that satisfy the expectations of all properties among multiple properties. In particular, according to the method of the second specific example, even when the actual condition data includes classification properties, it is possible to explore pigment materials that satisfy the expectations of all properties among multiple properties with high precision. For example, regarding actual condition data related to lightfastness, experimental conditions or evaluation criteria may sometimes vary depending on each experiment or each database. According to this embodiment, even in such cases, it is possible to search for pigment materials that satisfy the expectations of all properties among multiple properties with high precision by processing the actual condition data as classification information or discrete value information.

[0094] Furthermore, although the property prediction model 206 has been described as a learning model that takes categorical properties and latent variables in the latent space 205 as inputs and outputs predicted property values ​​corresponding to those latent variables, it is not limited to this. The property prediction model 206 can also be a learning model that takes latent variables in the latent space 205 as inputs and outputs predicted property values ​​corresponding to those latent variables. In other words, the inputs to the property prediction model 206 may not include categorical properties. For example, when categorical properties are independent of the properties of the predicted object, the properties of the predicted object can be predicted with high accuracy even without inputting categorical properties into the property prediction model 206.

[0095] (A variation of the second specific example)

[0096] Reference Figure 13 This is a variation of a second specific example of the method for searching pigment materials according to an embodiment of the present disclosure.

[0097] like Figure 13 As shown, a modified example of the second specific example includes a VAE encoder 203, a VAE decoder 204, a property prediction model 206, and a learned prediction model 207. That is, in this case, the storage unit 12 stores the VAE encoder 203, the VAE decoder 204, the property prediction model 206, and the learned prediction model 207.

[0098] The VAE encoder 203, VAE decoder 204, and property prediction model 206 are the same as those in the second specific example. The learned prediction model 207 is a learning model that takes pigment material information as input and outputs predicted values ​​of the properties corresponding to that pigment material information. These predicted values ​​correspond to the classified properties input to the VAE encoder 203. In this embodiment, the predicted values ​​output by the learned prediction model 207 are assumed to be prediction values ​​related to lightfastness.

[0099] In a variation of the second specific example, the predicted values ​​of the physical properties (in this case, the predicted values ​​related to lightfastness) predicted by the fully learned prediction model 207 are used as inputs to the VAE encoder 203, the VAE decoder 204, and the physical property prediction model 206. By doing so, even when there is insufficient actual condition data related to the classified physical properties, it is possible to search for pigment materials that satisfy the expectations of all physical properties across multiple properties. For example, there may be cases where actual condition data related to lightfastness has not been measured as experimental results or has not been documented. Even in such cases, the insufficient actual condition data can be supplemented by using the fully learned prediction model 207. Furthermore, by supplementing the insufficient actual condition data through the fully learned prediction model 207, it is possible to search for pigment materials that satisfy the expectations of all physical properties across multiple properties with the same high accuracy as in the second specific example.

[0100] (Third specific example)

[0101] Reference Figure 14 This will be used to illustrate a third specific example and operation of a method for searching pigment materials according to an embodiment of the present disclosure.

[0102] like Figure 14 As shown, the third specific example includes a VAE encoder 303 and a VAE decoder 304. That is, in this case, the storage unit 12 stores the VAE encoder 303 and the VAE decoder 304.

[0103] The VAE encoder 303 is a learning model that takes physical property data and pigment material information as input and outputs latent variables equivalent to the pigment material information. As described above, the physical property data consists of various physical property values ​​related to the pigment material. Such physical property values ​​can be continuous, discrete, or categorical information. Alternatively, physical property values ​​can be information obtained by converting continuous or discrete information into categorical information. Here, the physical property data is categorical information. As described above, the pigment material information is data labeled with SMILES. The latent variables output by the VAE encoder 303 are variables in the latent space 305. Figure 14The latent space 305 is represented by two-dimensional coordinates, but the dimension of the latent space 305 is not limited to 2. The dimension of the latent space can also be 3 or higher. The VAE encoder 303 is trained based on physical property data and pigment material information.

[0104] VAE decoder 304 is a learning model that takes arbitrary latent variables and physical property data in the latent space 305 as input and outputs pigment material information. VAE decoder 304 is trained based on physical property data and pigment material information.

[0105] Figure 15 This is a flowchart illustrating a third specific example of a method for searching pigment materials according to an embodiment of this disclosure.

[0106] Step S310: The control unit 11 of the information processing device 10 trains the VAE encoder 303 and the VAE decoder 304 based on the pigment material information and physical property data.

[0107] Step S320: The control unit 11 inputs the latent variables randomly selected from the latent space into the VAE decoder 304 and outputs the corresponding pigment material information. Various methods can be used to select the latent variables. For example, the control unit 11 can simply and randomly select latent variables from the latent space. Alternatively, the control unit 11 can randomly select latent variables from the vicinity of latent variables corresponding to known pigment materials for which corresponding experimental data exists. Or, the control unit 11 can randomly select latent variables from the vicinity of latent variables that can satisfy the desired physical properties (in other words, from the vicinity of the target pigment material).

[0108] Step S330: If the pigment material information output by the control unit 11 in step S320 meets a predetermined criterion, the control unit 11 outputs the pigment material information as the desired pigment material. The predetermined criterion refers to, for example, that the output pigment material information conforms to the syntax rules of SMILES. That is, if, for example, the pigment material information output in step S320 conforms to the syntax rules of SMILES, the control unit 11 outputs the pigment material information as the desired pigment material.

[0109] Thus, according to this embodiment, a VAE encoder 303 and a VAE decoder 304 trained based on pigment material information and physical property data can be used to search for pigment materials that satisfy the expectations of all physical properties among multiple physical properties. In particular, according to the method involved in the third specific example, even if all the actual condition data of multiple physical property values ​​are classification information, it is possible to search for pigment materials that satisfy the expectations of all physical properties among multiple physical properties.

[0110] Furthermore, while experimental data 1 and disclosed DB 2 may contain information related to the pigment material and actual condition data related to the pigment material, they are not limited to this. Experimental data 1 and disclosed DB 2 may also contain actual condition data of the composition used in applications other than pigment materials. Additionally, the VAE encoder 3, VAE encoder 103, VAE encoder 203, VAE encoder 303, VAE decoder 4, VAE decoder 104, VAE decoder 204 and VAE decoder 304, property prediction model 106, property prediction model 206, and learning-completed prediction model 207 described above may all be trained using actual condition data of the composition used in applications other than the pigment material. In other words, the VAE encoder, VAE decoder, property prediction model, and learning-completed prediction model in this embodiment may be trained using both actual condition data of the composition used as a pigment material and actual condition data of the composition used in applications other than pigment materials. In this way, by including a wide range of applicable compositions in the training data used for training the VAE encoder, VAE decoder, property prediction model, and the completed prediction model, it is possible to prevent the reduction in accuracy due to extrapolation.

[0111] While this disclosure has been described with reference to the accompanying drawings and embodiments, it should be noted that various modifications and variations can be easily made based on this disclosure by those skilled in the art. Therefore, it should be understood that such modifications and variations are included within the scope of this disclosure. For example, the functions contained in each unit or step can be reconfigured in a logically consistent manner, and multiple units or steps can be combined into one or divided.

[0112] Explanation of reference numerals in the attached figures

[0113] 1: Experimental data; 2: Public database; 3, 103, 203, 303: VAE encoders; 4, 104, 204, 304: VAE decoders; 5, 105, 205, 305: Latent space; 10: Information processing device; 11: Control unit; 12: Storage unit; 13: Input unit; 14: Output unit; 106, 206: Property prediction models; 207: Prediction model after learning completion.

Claims

1. A method for searching pigment materials, performed by an information processing device, the method comprising the following steps: The training steps involve training the VAE encoder and VAE decoder separately. The VAE encoder takes pigment material information represented by a prescribed labeling method as input and outputs latent variables in a latent space corresponding to the pigment material information. The VAE decoder takes any latent variable in the latent space as input and outputs pigment material information represented by the prescribed labeling method. The determination step involves identifying a pigment material that satisfies the desired properties among the various physical properties of the pigment material, based on data related to the VAE encoder, the VAE decoder, and the multiple physical properties of the pigment material. In the training step, the VAE encoder and the VAE decoder are trained using actual condition data of the composition used as a pigment material and actual condition data of the composition used in applications other than pigment materials.

2. The method for searching pigment materials according to claim 1, wherein, It also includes the following steps: A property prediction model is trained by taking arbitrary variables of the latent space as input and outputting predicted values ​​of the various physical properties. Data related to the various physical properties of the pigment material are used in the training of the property prediction model. In the determination step, latent variables corresponding to the desired pigment material are searched through optimization processing based on the physical property prediction model.

3. The method for searching pigment materials according to claim 2, wherein, The desired pigment material is determined by inputting a portion of the data relating to various physical properties of the pigment material into the VAE encoder and the VAE decoder.

4. The method for searching pigment materials according to claim 3, wherein, The desired pigment material is also determined by inputting a portion of the data related to various physical properties of the pigment material into the property prediction model.

5. The method for searching pigment materials according to claim 1, wherein, The desired pigment material is determined by inputting data related to various physical properties of the pigment material into the VAE encoder and the VAE decoder.

6. The method for searching pigment materials according to any one of claims 1 to 5, wherein, The various physical properties include information related to hue.

7. The method for searching pigment materials according to any one of claims 1 to 5, wherein, The various physical properties include information related to stability.

8. The method for searching pigment materials according to any one of claims 2 to 4, wherein, The optimization process is a Bayesian optimization process.

9. The method for searching pigment materials according to claim 1, wherein, At least a portion of the data involved in the various physical properties is continuous value information, discrete value information, or classification information.

10. The method for searching pigment materials according to claim 1, wherein, The pigment material is a dichroic pigment material, and the various physical properties include the maximum absorption wavelength and the dichroic ratio.

11. An information processing apparatus comprising a control unit, the information processing apparatus being used for searching for pigment materials, wherein, The control unit performs the following processing: The VAE encoder and VAE decoder are trained separately. The VAE encoder takes pigment material information represented by a prescribed labeling method as input and outputs latent variables in a latent space corresponding to the pigment material information. The VAE decoder takes any latent variable in the latent space as input and outputs pigment material information represented by the prescribed labeling method. Based on the data related to the VAE encoder, the VAE decoder, and the various physical properties of the pigment material, a pigment material that satisfies the desired properties among all of the various physical properties is determined. Specifically, when training the VAE encoder and the VAE decoder, actual condition data of the composition used as a pigment material and actual condition data of the composition used in applications other than pigment materials are used to train the VAE encoder and the VAE decoder.

12. A non-transitory computer-readable recording medium storing commands for searching pigment materials, wherein, When the command is executed by the processor, it causes the processor to perform the following processing: The VAE encoder and VAE decoder are trained separately. The VAE encoder takes pigment material information represented by a prescribed labeling method as input and outputs latent variables in a latent space corresponding to the pigment material information. The VAE decoder takes any latent variable in the latent space as input and outputs pigment material information represented by the prescribed labeling method. Based on the data related to the VAE encoder, the VAE decoder, and the various physical properties of the pigment material, a pigment material that satisfies the desired properties among all of the various physical properties is determined. Specifically, when training the VAE encoder and the VAE decoder, actual condition data of the composition used as a pigment material and actual condition data of the composition used in applications other than pigment materials are used to train the VAE encoder and the VAE decoder.

13. A computer program product comprising commands for searching for pigment materials, wherein, When the command is executed by the processor, it causes the processor to perform the following processing: The VAE encoder and VAE decoder are trained separately. The VAE encoder takes pigment material information represented by a prescribed labeling method as input and outputs latent variables in a latent space corresponding to the pigment material information. The VAE decoder takes any latent variable in the latent space as input and outputs pigment material information represented by the prescribed labeling method. Based on the data related to the VAE encoder, the VAE decoder, and the various physical properties of the pigment material, a pigment material that satisfies the desired properties among all of the various physical properties is determined. Specifically, when training the VAE encoder and the VAE decoder, actual condition data of the composition used as a pigment material and actual condition data of the composition used in applications other than pigment materials are used to train the VAE encoder and the VAE decoder.

Images