Prediction device, learning device, methods thereof, and program

Abstract

To predict physical properties of a substance whose structure is unknown.SOLUTION: A prediction device includes an acquisition unit that acquires analysis data including a mass spectrum of a target substance at a plurality of time points, a feature extraction unit that extracts feature data related to a peak of an ion count from the analysis data, and a physical property prediction unit that predicts a physical property of the target substance by inputting the feature data extracted by the feature extraction unit to a prediction model in which a relationship between a predetermined physical property and the feature data is learned.SELECTED DRAWING: Figure 3

Description

[Technical Field]

[0001] The present disclosure relates to a prediction device, a learning device, and methods and programs thereof. [Background technology]

[0002] Techniques for predicting the physical properties of a substance based on machine learning are known. For example, Patent Document 1 discloses a method for estimating a rubber composition, which involves obtaining analytical data obtained by qualitatively analyzing a target rubber composition using an analyzer, inputting input data created based on the analytical data into a trained machine learning model, and deriving an output from the trained machine learning model that corresponds to at least one of the abundance ratio of compounds in the target rubber composition and the physical properties of the target rubber composition. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Publication No. 2023-94142 Summary of the Invention [Problem to be solved by the invention]

[0004] However, in conventional techniques, physical properties cannot be predicted unless the structure of a substance is known. For example, in Patent Document 1, peaks derived from pyrolysis products are identified from gas chromatogram data of a rubber composition, so qualitative analysis is not possible unless the compounds that make up the rubber composition are known.

[0005] One aspect of the present disclosure aims to predict the physical properties of a substance whose structure is unknown. [Means for solving the problem]

[0006] The present disclosure has the following configuration.

[0007] <1> an acquisition unit configured to acquire analytical data including mass spectra of the target substance at multiple time points; a feature extraction unit configured to extract feature data relating to an ion count peak from the analysis data; a physical property prediction unit configured to predict the physical properties of the target substance by inputting the feature data extracted by the feature extraction unit into a prediction model that has learned the relationship between predetermined physical properties and the feature data; A prediction device comprising:

[0008] <2> The analytical data includes analytical results obtained by pyrolysis gas chromatography. the above <1> The prediction device according to claim 1.

[0009] <3> the characteristic data includes time information relating to one or more time points at which the sum of the ion counts reaches a peak, and mass information relating to one or more masses at which the ion counts reach a peak at each of the one or more time points; the above <1> or <2> The prediction device according to claim 1.

[0010] <4> the time information includes information indicating a time at which the sum of ion counts reaches a peak and a relative intensity of the sum of ion counts at the peak time; the above <3> The prediction device according to claim 1.

[0011] <5> the information indicating the time includes a retention index normalized to a retention time in an analytical device that outputs the analytical data; the above <4> The prediction device according to claim 1.

[0012] <6> The mass information includes information indicating a mass at which the ion count peaks and a relative intensity of the ion count at that mass. the above <3> from <5> The prediction device according to any one of the preceding claims.

[0013] <7> the characteristic data is sorted in descending order of the sum of the ion counts; the above <3> from <6> The prediction device according to any one of the preceding claims.

[0014] <8> an acquisition unit configured to acquire analytical data including mass spectra of a target substance at multiple time points and physical property values ​​obtained by measuring predetermined physical properties using the target substance; a feature extraction unit configured to extract feature data relating to an ion count peak from the analysis data; a model learning unit configured to learn a prediction model in which the feature data is an explanatory variable and the physical property is a target variable based on a data set including the feature data and the physical property value; A learning device comprising:

[0015] <9> The computer acquiring analytical data including mass spectra of the target substance at multiple time points; extracting characteristic data relating to ion count peaks from the analysis data; a step of predicting the physical properties of the target substance by inputting the feature data extracted in the extraction step into a prediction model that has learned the relationship between predetermined physical properties and the feature data; A forecasting method that performs.

[0016] <10> The computer a step of acquiring analytical data including mass spectra of a target substance at multiple time points and physical property values ​​obtained by measuring predetermined physical properties using the target substance; extracting characteristic data relating to ion count peaks from the analysis data; a step of learning a prediction model based on a data set including the feature data and the physical property values, the prediction model having the feature data as explanatory variables and the physical property as objective variables; Learning how to do it.

[0017] <11> On the computer, acquiring analytical data including mass spectra of the target substance at multiple time points; extracting feature data from the analysis data; a step of predicting the physical properties of the target substance by inputting the feature data extracted in the extraction step into a prediction model that has learned the relationship between predetermined physical properties and the feature data; A program to execute.

[0018] <12> On the computer, a step of acquiring analytical data including mass spectra of a target substance at multiple time points and physical property values ​​obtained by measuring predetermined physical properties using the target substance; extracting characteristic data relating to ion count peaks from the analysis data; a step of learning a prediction model based on a data set including the feature data and the physical property values, the prediction model having the feature data as explanatory variables and the physical property as objective variables; A program to execute. [Effects of the Invention]

[0019] According to one aspect of the present disclosure, the physical properties of a substance whose structure is unknown can be predicted. [Brief explanation of the drawings]

[0020] [Figure 1] FIG. 1 is a block diagram showing an example of the overall configuration of a physical property prediction system. [Figure 2] FIG. 1 is a block diagram illustrating an example of a computer. [Figure 3] FIG. 2 is a block diagram showing an example of a functional configuration of a physical property prediction system. [Figure 4] FIG. 10 is a diagram illustrating an example of analysis data. [Figure 5] FIG. 10 is a diagram illustrating an example of feature data. [Figure 6]10 is a flowchart illustrating an example of a learning process. [Figure 7] 10 is a flowchart illustrating an example of a feature extraction process. [Figure 8] 10 is a flowchart illustrating an example of a prediction process. [Figure 9] FIG. 10 is a diagram illustrating an example of prediction accuracy based on unaligned feature data. [Figure 10] FIG. 10 is a diagram illustrating an example of prediction accuracy based on aligned feature data. [Figure 11] FIG. 10 is a diagram illustrating an example of prediction accuracy for each number of detected peaks. DETAILED DESCRIPTION OF THE INVENTION

[0021] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. In this specification and drawings, components having substantially the same functional configurations are designated by the same reference numerals, and redundant description will be omitted.

[0022] [Embodiment] An embodiment of the present disclosure is an information processing system that predicts the physical properties of a substance. Hereinafter, the information processing system according to this embodiment will be referred to as a "physical property prediction system." Furthermore, a substance to be predicted will be referred to as a "prediction target substance."

[0023] In this embodiment, the prediction target substance may be a substance whose structure is unknown. The prediction target substance may be any substance that can be thermally decomposed. For example, the prediction target substance may be a polymer material. For example, the polymer material may be a resin such as elastomer, polyester, or polyamide.

[0024] Conventionally, techniques for predicting the physical properties of substances using methods such as machine learning have been known. Many conventional property prediction techniques use chemical structure as an explanatory variable. Conversely, conventional property prediction techniques cannot predict the physical properties of a substance if the chemical structure of the substance is unknown. Although publicly available databases related to substances or materials can be used as datasets for property prediction, only a small number of substances have known chemical structures among those registered in the databases. Therefore, conventional techniques are limited in the substances whose physical properties can be predicted.

[0025] The present embodiment aims to predict the physical properties of a substance whose structure is unknown. To this end, the present embodiment extracts feature data relating to ion count peaks from analytical data including mass spectra of the substance at multiple time points, and predicts the physical properties of the substance by inputting the extracted feature data into a prediction model that has learned the relationship between predetermined physical properties and the feature data.

[0026] In one aspect, this embodiment predicts physical properties based on mass spectra of a substance at multiple time points, making it possible to predict the physical properties of a substance whose structure is unknown. In another aspect, this embodiment makes it possible to construct a dataset without specifying the structure of the substance, making it possible to efficiently train a highly accurate prediction model.

[0027] <Overall structure> The overall configuration of a property prediction system according to this embodiment will be described with reference to Fig. 1. Fig. 1 is a block diagram showing an example of the overall configuration of a property prediction system.

[0028] 1, the property prediction system 1000 includes an analysis device 10, a learning device 20, a prediction device 30, and a terminal device 40. The analysis device 10, the learning device 20, the prediction device 30, and the terminal device 40 are connected to each other so as to be able to communicate data with each other via a communication network N. The communication network N may be, for example, a local area network (LAN), a virtual private network (VPN), the Internet, or the like.

[0029] The analytical device 10 is an example of a device, instrument, or tool for analyzing a substance. The analytical device 10 may be an analytical system including multiple devices, instruments, or tools. The analytical device 10 uses a sample of the substance to be analyzed to measure physical quantities that indicate the characteristics of the substance, and outputs analytical data based on the measurement results. The analytical data is electronic data that indicates the results of the analysis by the analytical device 10.

[0030] In this embodiment, the analysis device 10 may be, for example, a pyrolysis gas chromatograph. The analysis data output by the pyrolysis gas chromatograph includes the analysis results obtained by pyrolysis gas chromatography (i.e., a pyrolysis gas chromatogram). The pyrolysis gas chromatogram includes a total ion count current chromatogram (TICC) and mass spectra (MS) at multiple time points.

[0031] A total ion count current chromatogram is a chromatogram with time on the horizontal axis and total ion count on the vertical axis. The total ion count is the sum of the ion counts detected at a certain time. A mass spectrum is a spectrum with mass (m / z) on the horizontal axis and ion count on the vertical axis. The total ion count or ion count may be expressed as relative intensity.

[0032] The learning device 20 is an example of an information processing device such as a personal computer, workstation, or server that learns a prediction model. The prediction model is an example of a machine learning model that predicts a predetermined physical property. The predetermined physical property may include, for example, a relative permittivity (Dk) or a dielectric loss tangent (Df). The predetermined physical property is not limited to these, and any physical property may be the target of prediction.

[0033] The prediction model may be a machine learning model that learns the relationship between feature data extracted from analytical data on a substance to be learned and physical property values ​​obtained by measuring predetermined physical properties using the substance. The prediction model may be a machine learning model that uses the feature data as explanatory variables and the predetermined physical properties as target variables.

[0034] Hereinafter, the substance to be learned will also be referred to as the "learning substance." The learning substance may include a substance whose structure is unknown. The learning substance may include a substance whose structure is known. The learning substance may be any substance that can be thermally decomposed, similar to the prediction substance.

[0035] The learning device 20 may construct a dataset for training a prediction model based on feature data extracted from analytical data related to the training target substance and measurement data including physical property values ​​obtained by measuring predetermined physical properties using the training target substance. The learning device 20 may generate a prediction model by learning the relationship between the feature data and the physical property values ​​based on the dataset. The learning device 20 may output the trained prediction model to the prediction device 30.

[0036] The prediction device 30 is an example of an information processing device such as a personal computer, workstation, or server that predicts the physical properties of a prediction target substance. The prediction device 30 may predict the physical properties of the prediction target substance based on a trained prediction model. The trained prediction model may be generated by the learning device 20. The prediction device 30 may predict the physical properties of the prediction target substance by inputting feature data extracted from analytical data related to the prediction target substance into the trained prediction model.

[0037] The terminal device 40 is an example of an information processing terminal such as a personal computer, smartphone, or tablet terminal operated by a user of the property prediction system 1000. The terminal device 40 may acquire analytical data related to the learning target substance from the analysis device 10 and transmit it to the learning device 20. The terminal device 40 may transmit measurement data related to the learning target substance to the learning device 20. The terminal device 40 may acquire analytical data related to the prediction target substance from the analysis device 10 and transmit it to the prediction device 30. The terminal device 40 may receive prediction results from the prediction device 30 and present them to the user.

[0038] The overall configuration of the property prediction system 1000 shown in FIG. 1 is an example, and various system configurations are possible depending on the application and purpose. For example, the property prediction system 1000 may include multiple units of one or more of the analysis device 10, learning device 20, prediction device 30, and terminal device 40. For example, the learning device 20 or the prediction device 30 may be implemented by multiple computers, or may be implemented as a cloud computing service. For example, the property prediction system 1000 may be implemented by a standalone computer. The device divisions shown in FIG. 1, such as the analysis device 10, learning device 20, prediction device 30, and terminal device 40, are an example.

[0039] <Hardware configuration> The hardware configuration of the property prediction system 1000 will be described with reference to Fig. 2. The learning device 20, the prediction device 30, and the terminal device 40 are realized by, for example, a computer. Fig. 2 is a block diagram showing an example of the hardware configuration of a computer.

[0040] 2, the computer 500 includes a CPU (Central Processing Unit) 501, a ROM (Read Only Memory) 502, a RAM (Random Access Memory) 503, a HDD (Hard Disk Drive) 504, an input device 505, a display device 506, a communication I / F (Interface) 507, and an external I / F 508. The CPU 501, the ROM 502, and the RAM 503 form a so-called computer. The hardware components of the computer 500 are connected to each other via a bus line 509. The input device 505 and the display device 506 may be connected to the external I / F 508 for use.

[0041] The CPU 501 is a computing device that reads programs and data from a storage device such as the ROM 502 or the HDD 504 onto the RAM 503 and executes the processes, thereby realizing the overall control and functions of the computer 500. The computer 500 may have a GPU (Graphics Processing Unit) in addition to or instead of the CPU 501.

[0042] The ROM 502 is an example of a non-volatile semiconductor memory (storage device) that can retain programs and data even when the power is turned off. The ROM 502 functions as a main storage device that stores various programs, data, etc. required for the CPU 501 to execute various programs installed in the HDD 504. Specifically, the ROM 502 stores boot programs such as a Basic Input Output System (BIOS) and an Extensible Firmware Interface (EFI) that are executed when the computer 500 starts up, as well as data such as OS (Operating System) settings and network settings.

[0043] The RAM 503 is an example of a volatile semiconductor memory (storage device) in which programs and data are erased when the power is turned off. The RAM 503 is, for example, a dynamic random access memory (DRAM) or a static random access memory (SRAM). The RAM 503 provides a working area in which various programs installed in the HDD 504 are expanded when executed by the CPU 501.

[0044] The HDD 504 is an example of a non-volatile storage device that stores programs and data. The programs and data stored in the HDD 504 include an OS, which is basic software that controls the entire computer 500, and applications that provide various functions on the OS. Note that the computer 500 may use a storage device that uses flash memory as a storage medium (e.g., an SSD (Solid State Drive)) instead of the HDD 504.

[0045] The input device 505 includes a touch panel, operation keys and buttons, a keyboard and mouse, a microphone for inputting sound data such as voice, and the like, which are used by the user to input various signals.

[0046] The display device 506 is configured with a display such as a liquid crystal display or organic EL (Electro-Luminescence) display for displaying a screen, a speaker for outputting sound data such as voice, and the like.

[0047] The communication I / F 507 is an interface that connects to a communication network and enables the computer 500 to perform data communication.

[0048] The external I / F 508 is an interface with external devices, such as a drive device 510.

[0049] The drive device 510 is a device for loading a recording medium 511. The recording medium 511 here includes media that record information optically, electrically, or magnetically, such as a CD-ROM, a flexible disk, or a magneto-optical disk. The recording medium 511 may also include semiconductor memories that record information electrically, such as ROMs and flash memories. This allows the computer 500 to read from and / or write to the recording medium 511 via the external I / F 508.

[0050] The various programs to be installed in the HDD 504 are installed, for example, by setting the distributed recording medium 511 in a drive device 510 connected to the external I / F 508 and reading out the various programs recorded on the recording medium 511 by the drive device 510. Alternatively, the various programs to be installed in the HDD 504 may be installed by being downloaded via the communication I / F 507 from the communication network or another network different from the communication network.

[0051] <Functional configuration> The functional configuration of the property prediction system 1000 will be described with reference to Fig. 3. Fig. 3 is a block diagram showing an example of the functional configuration of the property prediction system.

[0052] Learning Device 3, the learning device 20 includes a data storage unit 201, an acquisition unit 210, a feature extraction unit 220, a dataset construction unit 230, and a model learning unit 240. The learning device 20 functions as the data storage unit 201, the acquisition unit 210, the feature extraction unit 220, the dataset construction unit 230, and the model learning unit 240 by executing a learning program installed in advance.

[0053] For example, the data storage unit 201 is realized by the HDD 504 shown in Fig. 2. For example, the acquisition unit 210, the feature extraction unit 220, the dataset construction unit 230, and the model learning unit 240 are realized by processing that is executed by the CPU 501 in accordance with a program loaded from the HDD 504 onto the RAM 503 shown in Fig. 2.

[0054] A dataset is stored in the data storage unit 201. A dataset may be stored in advance in the data storage unit 201. The dataset stored in the data storage unit 201 may be constructed by the dataset construction unit 230.

[0055] A dataset is a collection of training data for multiple training substances. The training data includes feature data extracted from analytical data for the training substances and physical property values ​​obtained by measuring predetermined physical properties using the training substances. The feature data is used as explanatory variables for the prediction model. The physical property values ​​are used as response variables for the prediction model.

[0056] The acquisition unit 210 acquires analytical data and measurement data. The acquisition unit 210 may acquire analytical data and measurement data for each of a plurality of learning target substances. The analytical data includes mass spectra of the learning target substance at a plurality of time points. In this embodiment, the analytical data includes analysis results obtained by pyrolysis gas chromatography. The measurement data includes physical property values ​​obtained by measuring predetermined physical properties using the learning target substance. In this embodiment, the physical properties include the relative permittivity or the dielectric loss tangent.

[0057] The acquiring unit 210 may receive the analysis data and measurement data transmitted by the terminal device 40. The acquiring unit 210 may acquire the analysis data output from the analysis device 10. The acquiring unit 210 may receive the analysis data transmitted by the analysis device 10.

[0058] The feature extraction unit 220 extracts feature data from the analytical data. The feature extraction unit 220 may extract feature data from the analytical data acquired by the acquisition unit 210. The feature extraction unit 220 may extract feature data from each piece of analytical data related to each of the study target substances.

[0059] The feature data includes information about the peak of the ion count. The feature data may include time information and mass information. The time information may be information about one or more time points at which the total ion count peaks. The mass information may be information about one or more masses at which the ion count peaks, at each of one or more time points at which the total ion count peaks. The feature data may include the same number of pieces of mass information as the number of time points included in the time information. The feature data may be a one-dimensional vector linking the time information and the mass information.

[0060] The time information may include information indicating the time at which the total ion count peaks and information indicating the total ion count at that time. The information indicating the time may be, for example, a retention index normalized to the retention time in the analyzer 10. The information indicating the total ion count may be, for example, a relative intensity of the total ion count.

[0061] The time information may include information about a predetermined number of peaks. For example, the number of peaks included in the time information may be 20.

[0062] The mass information may include information indicating the mass at which the ion count peaks and information indicating the ion count at that mass. The information indicating the mass may be, for example, an m / z value. The information indicating the ion count may be, for example, a relative intensity of the ion count.

[0063] The mass information may include information about a predetermined number of peaks. For example, the number of peaks included in the mass information may be 20. The number of peaks included in the mass information may be the same as or different from the number of peaks included in the time information.

[0064] The feature data may be sorted in descending order of the total ion count. When sorting in descending order of the total ion count, the time information included in the feature data is sorted by a set of information indicating the time and information indicating the total ion count in descending order of the total ion count. In this case, the mass information included in the feature data is sorted in the order of the information indicating the time in the time information.

[0065] The dataset construction unit 230 constructs a dataset. The dataset construction unit 230 may construct a dataset based on the measurement data acquired by the acquisition unit 210 and the feature data extracted by the feature extraction unit 220. The dataset construction unit 230 may generate training data for each training target substance, including the feature data and physical property values ​​indicated in the measurement data, and construct a dataset by integrating the training data for each training target substance. The dataset construction unit 230 may store the constructed dataset in the data storage unit 201.

[0066] The model learning unit 240 learns a prediction model. The model learning unit 240 may learn the prediction model based on a dataset constructed by the dataset construction unit 230. The model learning unit 240 may learn the prediction model based on a dataset read from the data storage unit 201. The model learning unit 240 may generate a trained prediction model by learning the relationship between feature data and physical property values ​​included in the dataset.

[0067] The model learning unit 240 outputs the trained prediction model. The model learning unit 240 may transmit the trained prediction model to the prediction device 30. The trained prediction model may be stored in the model storage unit 301 of the prediction device 30.

[0068] <Prediction Device> 3, the prediction device 30 includes a model storage unit 301, an acquisition unit 310, a feature extraction unit 320, and a physical property prediction unit 330. The prediction device 30 functions as the model storage unit 301, the acquisition unit 310, the feature extraction unit 320, and the physical property prediction unit 330 by executing a prediction program installed in advance.

[0069] For example, the model storage unit 301 is realized by the HDD 504 shown in Fig. 2. For example, the acquisition unit 310, the feature extraction unit 320, and the property prediction unit 330 are realized by processing that is executed by the CPU 501 in accordance with a program loaded from the HDD 504 onto the RAM 503 shown in Fig. 2.

[0070] A trained prediction model is stored in the model storage unit 301. A trained prediction model may be stored in advance in the model storage unit 301. The prediction model stored in the model storage unit 301 may be generated by the learning device 20.

[0071] The acquisition unit 310 acquires analytical data. The acquisition unit 310 may acquire analytical data related to the predicted target substance. The analytical data includes mass spectra of the predicted target substance at multiple time points. In this embodiment, the analytical data includes analysis results obtained by pyrolysis gas chromatography.

[0072] The acquiring unit 310 may receive the analysis data transmitted by the terminal device 40. The acquiring unit 310 may acquire the analysis data output from the analysis device 10. The acquiring unit 310 may receive the analysis data transmitted by the analysis device 10.

[0073] The feature extraction unit 320 extracts feature data from the analysis data. The feature extraction unit 320 may extract feature data from the analysis data acquired by the acquisition unit 310. The feature extraction unit 320 may extract feature data in the same manner as the feature extraction unit 220 of the learning device 20.

[0074] The physical property prediction unit 330 predicts physical properties. The physical property prediction unit 330 may predict the physical properties of the prediction target substance. The physical property prediction unit 330 may predict the physical properties of the prediction target substance based on a prediction model read out from the model storage unit 301. The physical property prediction unit 330 may predict the physical properties of the prediction target substance based on feature data extracted by the feature extraction unit 320. The physical property prediction unit 330 may predict the physical properties of the prediction target substance by inputting the feature data into the prediction model.

[0075] The physical property prediction unit 330 may output the predicted physical property results. The physical property prediction unit 330 may transmit the predicted physical property results to the terminal device 40. The physical property prediction unit 330 may display the predicted physical property results on the display device 506 of the prediction device 30. The physical property prediction unit 330 may output information used in the prediction together with the predicted physical property results. The information used in the prediction may include, for example, at least one of analytical data related to the prediction target substance and feature data extracted from the analytical data.

[0076] <Feature data> The feature data extracted from the analysis data by the feature extraction unit (the feature extraction unit 220 of the learning device 20 and the feature extraction unit 320 of the prediction device 30) according to this embodiment will be described in more detail with reference to FIGS.

[0077] Fig. 4 is a diagram showing an example of analytical data. Fig. 4 shows a pyrolysis gas chromatogram, which is an example of analytical data. As shown in Fig. 4, the analytical data includes a total ion count current chromatogram 400 and multiple mass spectra 410 (410-1, 410-2).

[0078] The feature extraction unit detects peaks of the total ion count in the total ion count current chromatogram 400. The detected peaks are a predetermined number of peaks in ascending order of total ion count. After detecting the predetermined number of peaks, the feature extraction unit identifies the time corresponding to each peak. FIG. 4 shows, as an example, that total ion count peaks are detected at times t1 to t3. While FIG. 4 shows an example in which three peaks are detected from the total ion count current chromatogram for the sake of visibility, the number of detected peaks may be 20.

[0079] The feature extraction unit acquires the mass spectrum at the specified time. Mass spectrum 410-1 shown in Figure 4 is the mass spectrum at time t1, and mass spectrum 410-2 is the mass spectrum at time t2.

[0080] The feature extraction unit detects ion count peaks in each of the acquired mass spectra 410-1 and 410-2. The peaks to be detected are a predetermined number of peaks, starting with the peak with the largest ion count. Once the predetermined number of peaks have been detected, the feature extraction unit identifies the masses corresponding to each peak. As an example, FIG. 4 shows that peaks at masses m1, m2, and m3 have been detected in mass spectrum 410-1 at time t1, and peaks at masses m4, m5, and m6 have been detected in mass spectrum 410-2 at time t2. While FIG. 4 shows an example in which three peaks are detected in each mass spectrum for ease of visibility, the number of peaks to be detected may be 20.

[0081] FIG. 5 is a diagram showing an example of feature data. The feature extraction unit generates feature data based on peaks detected from the analysis data. Specifically, the feature extraction unit generates time information 420 based on the peaks of the total ion count detected from the total ion count current chromatogram 400. The time information 420 includes the times at which the total ion count peaks (times 1, 2, . . .) and the relative intensities of the total ion count at each of the times 1, 2, . . . (intensity 1, 2, . . .). The feature extraction unit may sort the time information 420 in descending order of intensity 1, 2, . . .

[0082] Times 1, 2, ... included in the time information 420 are retention indices obtained by normalizing the retention times in the analysis device 10. The retention time is the time a sample is retained in the analysis device 10 after pyrolysis. Since retention times vary depending on the analysis device 10, normalizing the retention times makes it possible to generate feature data with reduced differences between the analysis devices 10.

[0083] It is known that the retention index is proportional to the logarithm of the retention time. For example, the retention time can be converted to the retention index using equation (1).

[0084]

number

[0085] However, RI is a retention index, and t A is the retention time of the sample, and t m is the retention time of a linear alkane with carbon number m. By measuring the retention times of multiple linear alkanes with different carbon numbers and creating a calibration curve, the coefficients a and b in equation (1) can be determined.

[0086] The feature extraction unit generates mass information 430 (430-1 to 430-3) based on the peaks of the ion count detected from each mass spectrum 410. The mass information 430 includes the mass at which the ion count peaks and the relative intensity of the ion count at that mass.

[0087] The feature extraction unit sequentially concatenates the time information 420 and mass information 430-1, 430-2, .... When the time information 420 is sorted in descending order of intensity 1, 2, ..., the mass information 430-1, 430-2, ... is sorted in the order of time 1, 2, .... In this way, the feature extraction unit generates feature data 440 expressed as a one-dimensional vector.

[0088] Note that the functional configuration of the property prediction system 1000 shown in Fig. 3 is one example, and it goes without saying that there are various examples of functional configurations depending on the application and purpose. The division of storage units such as the data storage unit 201 and model storage unit 301 shown in Fig. 3 is one example. Furthermore, the division of processing units such as the acquisition unit 210, feature extraction unit 220, dataset construction unit 230, model learning unit 240, acquisition unit 310, feature extraction unit 320, and property prediction unit 330 shown in Fig. 3 is one example.

[0089] For example, at least two of the acquisition unit 210, feature extraction unit 220, dataset construction unit 230, model learning unit 240, acquisition unit 310, feature extraction unit 320, and physical property prediction unit 330 may be integrated into one processing unit. For example, at least one of the acquisition unit 210, feature extraction unit 220, dataset construction unit 230, model learning unit 240, acquisition unit 310, feature extraction unit 320, and physical property prediction unit 330 may be divided into multiple processing units.

[0090] For example, the data storage unit 201 and the model storage unit 301 may be integrated into a single storage unit. For example, the data storage unit 201 or the model storage unit 301 may be divided into multiple storage units. For example, the data storage unit 201 or the model storage unit 301 may not be included in the learning device 20 or the prediction device 30, but may be included in an external information processing device or storage device. Note that "external" means not included in the property prediction system 1000.

[0091] <Processing Procedure> The property prediction method executed by the property prediction system 1000 will be described with reference to Fig. 6 to Fig. 8. The property prediction method includes a learning process executed by the learning device 20 (see Fig. 6) and a prediction process executed by the prediction device 30 (see Fig. 8).

[0092] <Learning process> 6 is a flowchart showing an example of the learning process. The learning process is a process for learning a prediction model. The learning process is executed by the learning device 20.

[0093] In step S1, a user of the property prediction system 1000 acquires analytical data related to a target substance from the analytical device 10. The user may acquire multiple analytical data to be used in training the prediction model. The number of analytical data to be used in training the prediction model may be determined arbitrarily, but may be approximately 70, for example. The user inputs the acquired multiple analytical data into the terminal device 40. The terminal device 40 transmits the input multiple analytical data to the training device 20.

[0094] Learning device 20 receives multiple pieces of analytical data from terminal device 40. Acquisition unit 210 of learning device 20 acquires the multiple pieces of analytical data received by learning device 20. Acquisition unit 210 sends the acquired multiple pieces of analytical data to feature extraction unit 220 and dataset construction unit 230.

[0095] In step S2, the feature extraction unit 220 of the learning device 20 receives analytical data related to the target substance from the acquisition unit 210. The feature extraction unit 220 extracts feature data related to ion count peaks from the received analytical data. The feature extraction unit 220 may extract feature data from the analytical data related to the target substance by executing a feature extraction process (see FIG. 7 ). The feature extraction unit 220 sends the extracted feature data to the dataset construction unit 230.

[0096] Step S2 is repeatedly executed for each of the analytical data relating to the learning target substance acquired in step S1, whereby a plurality of pieces of feature data extracted from each of the analytical data are sent to the data set constructing unit 230.

[0097] In step S3, a user of the property prediction system 1000 acquires measurement data related to the learning target substance. The user may acquire multiple pieces of measurement data indicating physical property values ​​obtained by measuring predetermined physical properties using each of the learning target substances for which analytical data has been acquired. The user inputs the acquired multiple pieces of measurement data into the terminal device 40. The terminal device 40 transmits the input multiple pieces of measurement data to the learning device 20.

[0098] The learning device 20 receives a plurality of pieces of measurement data from the terminal device 40. The acquisition unit 210 of the learning device 20 acquires the plurality of pieces of measurement data received by the learning device 20. The acquisition unit 210 sends the acquired plurality of pieces of measurement data to the dataset creation unit 230.

[0099] In step S4, the dataset construction unit 230 of the learning device 20 receives feature data related to the learning target substance from the feature extraction unit 220. The dataset construction unit 230 also receives measurement data related to the learning target substance from the acquisition unit 210.

[0100] The dataset construction unit 230 acquires the physical property values ​​indicated in the measurement data. The dataset construction unit 230 generates training data including feature data and physical property values. The dataset construction unit 230 generates training data for each training target substance for which analysis data was acquired in step S1. The dataset construction unit 230 constructs a dataset by integrating the training data for each training target substance.

[0101] In step S5, the dataset construction unit 230 of the learning device 20 deletes, from among the data strings included in the dataset, data strings whose values ​​match in all of the feature data. If there is no data string whose values ​​match in all of the feature data, the dataset construction unit 230 does not need to delete the data string. The dataset construction unit 230 stores the dataset after the deletion of the data string in the data storage unit 201.

[0102] In step S6, the model learning unit 240 of the learning device 20 reads a dataset from the data storage unit 201. The model learning unit 240 may read only a portion of the dataset stored in the data storage unit 201, and not read the rest for verifying the prediction model. The model learning unit 240 generates a trained prediction model by learning the relationship between feature data and physical property values ​​based on the read dataset.

[0103] In step S7, the model learning unit 240 of the learning device 20 outputs the learned prediction model generated in step S6. The learning device 20 transmits the learned prediction model output from the model learning unit 240 to the prediction device 30.

[0104] The prediction device 30 receives the trained prediction model from the learning device 20. The prediction device 30 stores the received trained prediction model in the model storage unit 301.

[0105] Feature extraction processing 7 is a flowchart showing an example of the feature extraction process. The feature extraction process corresponds to step S2 of the learning process (see FIG. 6). The feature extraction process also corresponds to step S12 of the prediction process (see FIG. 8).

[0106] In step S2-1, the feature extraction unit 220 acquires a total ion count current chromatogram from the analysis data relating to the learning target substance, and detects a peak of the total ion count from the total ion count current chromatogram.

[0107] In step S2-2, the feature extraction unit 220 extracts a predetermined number of peaks from the total ion count peaks detected in step S2-1, starting with the peaks with the largest total ion count. The feature extraction unit 220 identifies the time corresponding to each of the extracted total ion count peaks.

[0108] In step S2-3, the feature extraction unit 220 converts the time (retention time) extracted in step S2-2 into a retention index. The feature extraction unit 220 generates time information to be included in the feature data based on the retention index corresponding to the peak of the total ion count and the relative intensity of the total ion count.

[0109] In step S2-4, the feature extraction unit 220 acquires a mass spectrum at the time extracted in step S2-2 from the analytical data related to the learning target substance. The feature extraction unit 220 detects an ion count peak from the acquired mass spectrum.

[0110] In step S2-5, the feature extraction unit 220 extracts a predetermined number of peaks, starting with the largest ion count, from the ion count peaks detected in step S2-4. The feature extraction unit 220 identifies the mass corresponding to each of the extracted ion count peaks. The feature extraction unit 220 generates a table containing data strings corresponding to the identified masses, with all intensities being zero.

[0111] In step S2-6, the feature extraction unit 220 acquires the relative intensities of the ion counts corresponding to the masses extracted in step S2-5, and then substitutes the acquired relative intensities into the table generated in step S2-5.

[0112] Step S2-6 is repeatedly performed for each mass extracted in step S2-5. As a result, feature extraction unit 220 generates a table containing masses corresponding to peaks in the mass spectrum and the relative intensities of the ion counts corresponding to those masses. Feature extraction unit 220 generates mass information based on the table containing masses and relative intensities.

[0113] The processes from step S2-4 to step S2-6 are repeatedly executed for each time extracted in step S2-2, whereby the feature extraction unit 220 generates multiple pieces of mass information for each time.

[0114] In step S2-7, the feature extraction unit 220 sequentially concatenates the time information generated in step S2-3 with the multiple mass information generated in step S2-6. This generates one-dimensional feature data including the time information and the multiple mass information. The feature extraction unit 220 outputs the generated feature data. The feature data output from the feature extraction unit 220 is sent to the dataset construction unit 230.

[0115] <Prediction processing> 8 is a flowchart showing an example of the prediction process. The prediction process is a process of predicting physical properties based on a trained prediction model. The prediction process is executed by the prediction device 30.

[0116] In step S11, a user of the property prediction system 1000 acquires analytical data related to the prediction target substance from the analytical device 10. The user inputs the acquired analytical data into the terminal device 40. The terminal device 40 transmits the input analytical data to the prediction device 30.

[0117] Prediction device 30 receives analytical data from terminal device 40. Acquisition unit 310 of prediction device 30 acquires the analytical data received by prediction device 30. Acquisition unit 310 sends the acquired analytical data to feature extraction unit 320.

[0118] In step S12, the feature extraction unit 320 of the prediction device 30 receives analytical data related to the predicted target substance from the acquisition unit 310. The feature extraction unit 320 extracts feature data related to ion count peaks from the received analytical data. The feature extraction unit 320 may extract feature data from the analytical data related to the predicted target substance by executing a feature extraction process (see FIG. 7).

[0119] In step S13, the feature extraction unit 320 of the prediction device 30 deletes data strings that are not included in the dataset from among the data strings included in the feature data. Data strings that are not included in the dataset are data strings that were deleted from the dataset in step S5 of the learning process (see FIG. 6). In other words, the feature extraction unit 320 deletes data strings that are not present in the explanatory variables of the prediction model from among the data strings included in the feature data. If no data strings were deleted in step S5 of the learning process, the feature extraction unit 320 does not need to delete the data strings. The feature extraction unit 320 sends the feature data after the deletion of the data strings to the property prediction unit 330.

[0120] In step S14, the physical property prediction unit 330 of the prediction device 30 receives feature data from the feature extraction unit 320. The physical property prediction unit 330 reads out a trained prediction model from the model storage unit 301. The physical property prediction unit 330 inputs the received feature data into the read prediction model. The prediction model predicts a predetermined physical property based on the input feature data and outputs a predicted value of the predetermined physical property. The physical property prediction unit 330 acquires the predicted value output from the prediction model.

[0121] In step S15, the physical property prediction unit 330 of the prediction device 30 outputs the predicted physical property results. The predicted physical property values ​​obtained in step S14 are included in the predicted physical property results. The predicted physical property results may also include information used in the prediction of the physical property. For example, the information used in the prediction may include at least one of analytical data related to the prediction target substance and feature data extracted from the analytical data.

[0122] The prediction device 30 transmits the prediction results of the physical properties to the terminal device 40. The terminal device 40 receives the prediction results of the physical properties from the prediction device 30. The terminal device 40 presents the received prediction results of the physical properties to the user. The terminal device 40 may display the predicted values ​​included in the prediction results on the display device 506. The terminal device 40 may display information used for the prediction together with the predicted values.

[0123] A user of the property prediction system 1000 may refer to the property prediction results displayed on the display device 506 of the terminal device 40 and use them in material development. For example, the user may use the predicted values ​​of the property of each candidate substance to select, from among multiple candidate substances, a substance whose property meets a target as a raw material for a specific product.

[0124] <Prediction accuracy> The prediction accuracy of the physical property prediction by the physical property prediction system 1000 will be described with reference to FIGS.

[0125] 9 is a diagram showing an example of prediction accuracy based on unsorted feature data. Unsorted feature data means that the data is not sorted in descending order of total ion count.

[0126] Figure 9 is a scatter plot plotting the relationship between the predicted values ​​(vertical axis) of a prediction model that uses feature data extracted from a pyrolysis gas chromatogram as explanatory variables and the predicted values ​​(horizontal axis) of a prediction model that uses chemical structure as explanatory variables. The predicted values ​​of a prediction model that uses chemical structure as explanatory variables include some of the actual measured values. In Figure 9, the coefficient of determination (R 2 ) was evaluated. As shown in Figure 9, the prediction accuracy based on the unaligned feature data was shown to be good, with a coefficient of determination of 0.414.

[0127] 10 is a diagram showing an example of prediction accuracy based on sorted feature data. Sorted feature data means that the data is sorted in descending order of total ion count.

[0128] Similar to FIG. 9, FIG. 10 is a scatter plot plotting the relationship between the predicted values ​​(vertical axis) of a prediction model that uses feature data extracted from a pyrolysis gas chromatogram as explanatory variables and the predicted values ​​(horizontal axis) of a prediction model that uses chemical structure as explanatory variables. The predicted values ​​of a prediction model that uses chemical structure as explanatory variables include some of the actual measured values. Also, similar to FIG. 9, FIG. 10 shows the coefficient of determination (R 2 As shown in Figure 10, the prediction accuracy based on the aligned feature data was 0.472, demonstrating even better prediction accuracy than that based on unaligned feature data.

[0129] 11 is a diagram showing an example of prediction accuracy for each number of detected peaks. The peaks detected in this embodiment include total ion count peaks in a total ion count current chromatogram and ion count peaks in a mass spectrum. In this embodiment, an example in which 20 peaks of each type are detected has been described, but these numbers may be set arbitrarily.

[0130] FIG. 11 shows the results of evaluating the prediction accuracy for each combination of the total ion count detection number and the ion count detection number, with the combination being variable. The vertical axis (Using TICCs) in FIG. 11 represents the total ion count detection number. The horizontal axis (Using m / z) in FIG. 11 represents the ion count detection number. The prediction accuracy is calculated as the coefficient of determination R between the predicted value of the prediction model using chemical structure as an explanatory variable and the predicted value of the prediction model according to this embodiment. 2 was used.

[0131] As shown in Figure 11, prediction accuracy tends to be high when the number of detected ion counts is around 20. The number of detected total ion counts did not change significantly from 5 to 25, indicating that any number of peaks could be detected.

[0132] <Effects of the embodiment> The prediction device 30 according to this embodiment acquires analytical data including mass spectra of a target substance at multiple time points, and predicts the physical properties of the target substance by inputting feature data relating to ion count peaks extracted from the analytical data into a prediction model that has learned the relationship between predetermined physical properties and the feature data. In one aspect, this embodiment predicts physical properties based on mass spectra of the target substance at multiple time points, making it possible to predict the physical properties of substances whose structures are unknown.

[0133] The analytical data may include analytical results obtained by pyrolysis gas chromatography. According to one aspect, the present embodiment makes it possible to predict the physical properties of a substance based on analytical results obtained by pyrolysis gas chromatography.

[0134] The feature data may include time information relating to one or more time points at which the total ion count reaches a peak, and mass information relating to one or more masses at which the ion count reaches a peak at each of the one or more time points. In one aspect, according to this embodiment, the physical properties of a substance can be predicted with high accuracy based on feature data relating to two types of peaks.

[0135] The time information may include information indicating the time at which the sum of ion counts reaches a peak and the relative intensity of the sum of ion counts at that time. The information indicating the time may include a retention index normalized to the retention time in the analyzer that output the analysis data. The mass information may include information indicating the mass at which the ion count reaches a peak and the relative intensity of the ion count at that mass. In one aspect, according to this embodiment, the peak time and mass can be handled on the same scale, thereby improving the efficiency of predictive model training and property prediction.

[0136] The feature data may be sorted in descending order of the total ion count. According to one aspect, the present embodiment allows the order of explanatory variables to have meaning, thereby making it possible to predict physical properties with higher accuracy.

[0137] The learning device 20 according to this embodiment acquires analytical data including mass spectra of a target substance at multiple time points and physical property values ​​obtained by measuring predetermined physical properties using the target substance, and learns a predictive model using the feature data as explanatory variables and the physical properties as target variables based on a dataset including the feature data and physical property values ​​extracted from the analytical data. In one aspect, this embodiment allows for the efficient learning of a highly accurate predictive model, since the dataset can be constructed without specifying the structure of the substance.

[0138] [supplement] Each function of the above-described embodiments can be realized by one or more processing circuits. Here, the term "processing circuit" in this specification includes a processor programmed to execute each function by software, such as a central processing unit (CPU) or a graphics processing unit (GPU) implemented by an electronic circuit, as well as devices such as an application specific integrated circuit (ASIC), a digital signal processor (DSP), a field programmable gate array (FPGA), and conventional circuit modules designed to execute each of the above-described functions.

[0139] Although the embodiments of the present disclosure have been described in detail above, the embodiments disclosed herein are illustrative in all respects and are not limiting. The embodiments can be modified and improved in various ways without departing from the scope and spirit of the appended claims. The matters described in the above embodiments can be configured in other ways as long as they are not inconsistent, and can be combined as long as they are not inconsistent. [Explanation of symbols]

[0140] 10: Analyzer 20: Learning device 30: Prediction device 40: Terminal device 201: Data storage unit 210: Acquisition Department 220: Feature extraction unit 230: Dataset Construction Department 240: Model learning section 301: Model storage unit 310: Acquisition Department 320: Feature extraction unit 330: Physical Properties Prediction Department 1000: Physical property prediction system

Claims

1. an acquisition unit configured to acquire analytical data including mass spectra of the target substance at multiple time points; a feature extraction unit configured to extract feature data relating to an ion count peak from the analysis data; a physical property prediction unit configured to predict the physical properties of the target substance by inputting the feature data extracted by the feature extraction unit into a prediction model that has learned the relationship between predetermined physical properties and the feature data; A prediction device comprising:

2. The analytical data includes analytical results obtained by pyrolysis gas chromatography. The prediction device according to claim 1 .

3. the characteristic data includes time information relating to one or more time points at which the sum of the ion counts reaches a peak, and mass information relating to one or more masses at which the ion counts reach a peak at each of the one or more time points; The prediction device according to claim 1 .

4. the time information includes information indicating a time at which the sum of ion counts reaches a peak and a relative intensity of the sum of ion counts at the peak time; The prediction device according to claim 3 .

5. the information indicating the time includes a retention index normalized to a retention time in an analytical device that outputs the analytical data; The prediction device according to claim 4 .

6. The mass information includes information indicating a mass at which the ion count peaks and a relative intensity of the ion count at that mass. The prediction device according to claim 3 .

7. the characteristic data is sorted in descending order of the sum of the ion counts; The prediction device according to any one of claims 3 to 6.

8. an acquisition unit configured to acquire analytical data including mass spectra of a target substance at multiple time points and physical property values ​​obtained by measuring predetermined physical properties using the target substance; a feature extraction unit configured to extract feature data relating to an ion count peak from the analysis data; a model learning unit configured to learn a prediction model in which the feature data is an explanatory variable and the physical property is a target variable based on a data set including the feature data and the physical property value; A learning device comprising:

9. The computer acquiring analytical data including mass spectra of the target substance at multiple time points; extracting characteristic data relating to ion count peaks from the analysis data; a step of predicting the physical properties of the target substance by inputting the feature data extracted in the extraction step into a prediction model that has learned the relationship between predetermined physical properties and the feature data; A forecasting method that performs.

10. The computer a step of acquiring analytical data including mass spectra of a target substance at multiple time points and physical property values ​​obtained by measuring predetermined physical properties using the target substance; extracting characteristic data relating to ion count peaks from the analysis data; a step of learning a prediction model based on a data set including the feature data and the physical property values, the prediction model having the feature data as explanatory variables and the physical property as objective variables; Learning how to do it.

11. On the computer, acquiring analytical data including mass spectra of the target substance at multiple time points; extracting feature data from the analysis data; a step of predicting the physical properties of the target substance by inputting the feature data extracted in the extraction step into a prediction model that has learned the relationship between predetermined physical properties and the feature data; A program to execute.

12. On the computer, a step of acquiring analytical data including mass spectra of a target substance at multiple time points and physical property values ​​obtained by measuring predetermined physical properties using the target substance; extracting characteristic data relating to ion count peaks from the analysis data; a step of learning a prediction model based on a data set including the feature data and the physical property values, the prediction model having the feature data as explanatory variables and the physical property as objective variables; A program to execute.

Images