Data parsing device, data parsing method, method for generating learned model, system and program

Abstract

This invention provides a data analysis apparatus, a data analysis method, a method for generating a learned model, a system, and a program capable of performing high-precision and easy FT-IR spectrum analysis. The data analysis apparatus (500) includes an acquisition unit (520) for acquiring the FT-IR spectrum to be analyzed, a learned model (M2), and an analysis unit (511) for inputting the analyzed object into the learned model. The learned model (M2) is obtained through machine learning in such a way that when an FT-IR spectrum is input, it outputs information indicating whether the input FT-IR spectrum contains peaks originating from a learned atomic group. The learned atomic group includes atomic groups with three or more atoms.

Description

Technical Field

[0001] This disclosure relates to data parsing apparatus, data parsing methods, methods for generating learned models, systems, and programs. Background Technology

[0002] A Fourier transform infrared spectrometer (hereinafter also referred to as a "FT-IR spectrometer") illuminates a sample with infrared interference light and detects the reflected or transmitted light. The FT-IR spectrometer then performs a Fourier transform on a graph (interferogram) recording its detection signal to obtain a spectrum (hereinafter also referred to as a "FT-IR spectrum") with wavelength or wavenumber on the horizontal axis and intensity (e.g., absorbance or transmittance) on the vertical axis. Since the FT-IR spectrum presents a pattern corresponding to the molecular structure of the sample, the user can use the FT-IR spectrum for qualitative analysis. Furthermore, since the intensity represented by the vertical axis of the FT-IR spectrum is approximately proportional to the concentration or thickness of the sample, the user can perform quantitative analysis based on the height or area of ​​the peaks in the FT-IR spectrum. An example of an FT-IR spectrometer is disclosed in International Patent Publication No. 2018 / 193499 (Patent Document 1).

[0003] Existing technical documents

[0004] Patent documents

[0005] Patent Document 1: International Publication No. 2018 / 193499 Summary of the Invention

[0006] The problem the invention aims to solve

[0007] Users can perform the aforementioned qualitative and / or quantitative analyses by resolving FT-IR spectra. Generally, FT-IR spectrum resolution utilizes a library (e.g., a database showing the spectral patterns of each compound or functional group). Users visually observe the FT-IR spectra and compare them to the spectral patterns shown in the library. However, FT-IR spectrum resolution can be challenging for users unfamiliar with this method. Furthermore, because compounds with complex structures exhibit complex spectral patterns, even experienced users may find accurate resolution difficult.

[0008] This disclosure was made to solve the above-mentioned problems. The purpose of this disclosure is to provide a data analysis apparatus, a data analysis method, a method for generating a learned model, a system, and a program that can perform FT-IR spectrum analysis with high accuracy and ease.

[0009] Solution for solving the problem

[0010] The data analysis apparatus according to the first aspect of this disclosure includes an acquisition unit for acquiring an FT-IR spectrum as the object of analysis, a learned model, and an analysis unit for inputting the object of analysis into the learned model. The learned model is obtained through machine learning in such a way that, when an FT-IR spectrum is input, it outputs information indicating whether the input FT-IR spectrum contains peaks originating from learned atomic groups. The learned atomic groups include atomic groups with three or more atoms.

[0011] The data parsing method involved in the second aspect of this disclosure includes the acquisition steps and input steps described below.

[0012] In the acquisition step, the FT-IR spectrum of a compound with an unknown molecular structure is acquired. In the input step, the FT-IR spectrum acquired in the acquisition step and specified information for identifying a group of atoms consisting of three or more atoms are input to the learned model. The learned model is a mathematical model obtained through machine learning in the following manner: when given the input FT-IR spectrum and specified information, it outputs information indicating whether the input FT-IR spectrum contains peaks originating from a group of atoms consisting of three or more atoms specified in the input specified information.

[0013] The method for generating a learned model according to the third approach disclosed herein includes the analysis steps, the saving steps, and the learning steps described below.

[0014] In the analysis step, Fourier transform infrared spectroscopy (FT-IR) is used to analyze compounds with known molecular structures to obtain FT-IR spectra. In the storage step, the FT-IR spectra obtained in the analysis step are associated with and stored along with the group information of the compounds analyzed in the analysis step. In the learning step, the FT-IR spectra and group information stored in the storage step are used as training data for machine learning to generate a learned model capable of resolving whether peaks originating from specified group members are present in the FT-IR spectra. Specified group members include groups of three or more atoms. Group information for compounds containing specified group members indicates the specified group members present in the compound. Group information for compounds not containing specified group members indicates that the compound does not contain the specified group members.

[0015] The fourth method disclosed herein involves a system that has one or more computers capable of performing the above-described data parsing method or the method for generating a learned model.

[0016] The fifth method of this disclosure involves a program for causing a computer to perform the aforementioned data parsing method or the method for generating a learned model.

[0017] The effects of the invention

[0018] The inventors of this application have successfully generated a learned model capable of highly accurate analysis of whether peaks originating from learned atomic groups (groups of three or more atoms) are present in FT-IR spectra (the analysis object input into the learned model). This model utilizes machine learning on atomic groups of three or more atoms. FT-IR spectra of compounds with atomic groups of three or more atoms tend to exhibit complex spectral patterns. However, in the aforementioned data analysis apparatus and method, the use of a learned model as described above allows for highly accurate and easy analysis of such FT-IR spectra. Furthermore, the method for generating the learned model described above allows for the appropriate generation of learned models as described. Attached Figure Description

[0019] Figure 1 This is a diagram illustrating the structure of a Fourier transform infrared spectrometer according to an embodiment of the present disclosure.

[0020] Figure 2 This is a diagram showing an example of an FT-IR spectrum.

[0021] Figure 3 This is a diagram illustrating an example of how users utilize a rule base to analyze FT-IR spectra.

[0022] Figure 4 It is shown Figure 1 A diagram showing the detailed structure of the control device.

[0023] Figure 5 This is a diagram illustrating an example of learning data in a method for generating a learned model according to an embodiment of this disclosure.

[0024] Figure 6 It is shown Figure 4 The diagram shows a detailed view of the server's structure.

[0025] Figure 7 This is a diagram illustrating the learned model mounted in the data parsing apparatus according to embodiments of this disclosure.

[0026] Figure 8 It is shown Figure 7 The graph shows the average AUC of each learned cluster in the learned model.

[0027] Figure 9 This is a figure showing an example of the evaluation results of a learned model when the specified atomic group is set to a benzamide structure.

[0028] Figure 10 This is a diagram showing an example of the evaluation results of a learned model when the specified atomic group is set to nitro.

[0029] Figure 11 This is a diagram showing an example of the evaluation results of a learned model when the specified atomic group is set to bromine.

[0030] Figure 12 This is a diagram illustrating an example of the changes in the value of the loss function and the classification accuracy in the method for generating a learned model according to an embodiment of this disclosure.

[0031] Figure 13 It is shown Figure 7 A diagram showing the detailed structure of the data parsing device.

[0032] Figure 14 This is a flowchart illustrating a data parsing method involved in an embodiment of the present disclosure.

[0033] Figure 15 It is shown in Figure 14 The image shown is an example of a screen displaying the input of the storage location of the FT-IR spectrum during the processing.

[0034] Figure 16 It is shown in Figure 14 The image shown is an example of a screen that accepts input from atomic groups during the processing.

[0035] Figure 17 It is shown in Figure 14 The image shown is an example of a screen that notifies the user of the parsing results during the processing.

[0036] Figure 18 It is shown Figure 16 The image shown is a variation of the scene.

[0037] Figure 19 It is shown Figure 14 A diagram showing a variation of the processing. Detailed Implementation

[0038] The embodiments of this disclosure will now be described in detail with reference to the accompanying drawings. Furthermore, the same or equivalent parts in the drawings will be labeled with the same reference numerals, and their descriptions will not be repeated.

[0039] Figure 1 This is a diagram showing the structure of the FT-IR spectrometer (Fourier transform infrared spectrometer) according to this embodiment.

[0040] Reference Figure 1The FT-IR spectrometer according to this embodiment includes an analysis device 100 and a control device 200 for controlling the analysis device 100. The analysis device 100 includes an interference wave generation unit 110, an irradiation unit 120, and a detector 130. The interference wave generation unit 110 generates infrared interference light. The irradiation unit 120 irradiates a sample M with the infrared interference light generated by the interference wave generation unit 110. The infrared interference light is reflected by the sample M. Then, the infrared interference light reflected by the sample M is emitted from the irradiation unit 120 to the detector 130. The detector 130 detects the infrared interference light incident from the irradiation unit 120.

[0041] The interference wave generating unit 110 includes a light source 111, a condenser lens 112, a collimating lens 113, a beam splitter 114, a fixed mirror 115, a movable mirror 116, and a mirror actuator 117. The illumination unit 120 includes a condenser lens 121, a prism 122, a condenser lens 123, and a pressing mechanism 124. The various actuators included in the interference wave generating unit 110 and the illumination unit 120 are controlled by the control device 200.

[0042] The light source 111 is configured to emit infrared light. At least one of, for example, a ceramic light source and a tungsten filament lamp can be used as the light source 111. Figure 1 In the image, the optical path of infrared light is shown using a double-dotted line.

[0043] Infrared light emitted from light source 111 is reflected by condenser lens 112 and collimating lens 113 to become parallel light, and then enters beam splitter 114. In beam splitter 114, a portion of the infrared light is reflected towards fixed lens 115. The remaining portion of the infrared light passes through beam splitter 114 and goes to movable lens 116. The infrared light reflected by fixed lens 115 and movable lens 116 respectively enters beam splitter 114 and is combined.

[0044] The movable mirror 116 is configured to move along the optical axis. The movable mirror 116 is moved along the optical axis by a mirror actuator 117. The control device 200 controls the reflecting surface of the movable mirror 116 by moving the movable mirror 116 via the mirror actuator 117. The collimating mirror 113, beam splitter 114, fixed mirror 115, movable mirror 116, and mirror actuator 117 constitute a Michelson interferometer. This Michelson interferometer is configured to generate infrared interference light using infrared light emitted from the light source 111. The infrared interference light generated by the Michelson interferometer is emitted from the beam splitter 114 towards the irradiation unit 120.

[0045] The FT-IR spectrometer described in this embodiment measures the FT-IR spectrum of sample M using the ATR (Attenuated Total Reflection) method. When sample M is brought into contact with prism 122 and infrared light is passed through prism 122 and irradiated onto sample M, the infrared light transmitted from the interior of prism 122 undergoes total internal reflection at the interface between sample M and prism 122. During this total internal reflection, the infrared light slightly penetrates towards sample M, and therefore, by detecting the total internal reflection light, the FT-IR spectrum of the surface of sample M can be obtained. For example, a diamond prism can be used as prism 122. However, it is not limited to this; a Ge prism or a ZnSe prism can also be used instead of a diamond prism. The pressing mechanism 124 is configured to fix sample M to the surface of prism 122 while pressing it down.

[0046] Infrared interference light incident from the interference wave generation unit 110 onto the illumination unit 120 is reflected by the condenser lens 121 and converged onto the prism 122. The infrared interference light incident onto the prism 122 undergoes total internal reflection at the interface between the sample M and the prism 122 after passing through the interior of the prism 122, and is then reflected by the condenser lens 123 and incident onto the detector 130.

[0047] Detector 130 is an infrared detector configured to output a detection signal corresponding to the infrared interference light incident from irradiation unit 120. Detector 130 can be at least one of, for example, a DLATGS (deuterated L-alanine doped triglycine sulfate) detector, an MCT (HgCdTe) detector, and an InGaAs detector. Detector 130 may also include a temperature control mechanism. Detector 130 outputs the aforementioned detection signal to control device 200.

[0048] In this embodiment, a computer comprising a processor 201, RAM (Random Access Memory) 202, a storage device 203, and a communication device 204 is used as the control device 200. The communication method of the communication device 204 is arbitrary and can be either wireless or wired communication. The processor 201 can be, for example, a CPU (Central Processing Unit). The number of processors in the control device 200 is arbitrary; it can be one or more. The RAM 202 functions as working memory for temporarily storing data processed by the processor 201. The storage device 203 is configured to store stored information. In addition to storing programs, the storage device 203 also stores information used in the programs (e.g., graphs, formulas, and various parameters).

[0049] The control device 200 performs an A / D (analog-to-digital) conversion on the detection signal input from the detector 130 and records the converted digital signal in the storage device 203. Thus, an interferogram is recorded in the storage device 203 of the control device 200. Furthermore, the control device 200 uses a Fourier transform to separate the composite waveform spectrum of the interference waves shown in the interferogram into the light intensity of each wavenumber component, thereby obtaining an FT-IR spectrum. In this embodiment, an FT-IR spectrum with wavenumber on the horizontal axis and absorbance on the vertical axis is used. Details of the function of the control device 200 will be described later (see [reference]). Figure 4 ).

[0050] As described above, the FT-IR spectrometer involved in this embodiment can analyze sample M to obtain the FT-IR spectrum of sample M. Sample M can also be a compound with an unknown molecular structure. Figure 2 This is a diagram showing an example of an FT-IR spectrum. Figure 2 The FT-IR spectrum shown is that of methanol, containing peaks originating from the hydroxyl group (-OH). (See reference...) Figure 2 In FT-IR spectra, peaks originating from hydroxyl groups appear in the wavenumber region of 3200 cm⁻¹ to 3600 cm⁻¹. The infrared absorption represented by the peaks in this region is characteristic of hydroxyl groups.

[0051] The position, intensity, and width of peaks appearing in the FT-IR spectrum of a compound vary depending on the atomic groups (i.e., substructures) contained in the compound. Users can determine the atomic groups contained in a compound by resolving its FT-IR spectrum. Libraries (e.g., databases showing spectral patterns for each compound or each atomic group) can be used in the resolution of FT-IR spectra. Users can resolve FT-IR spectra by visually observing the spectra and comparing them with spectral patterns shown in the library.

[0052] Figure 3 This is a diagram illustrating an example of how a user utilizes a rule base to analyze FT-IR spectra. (See reference...) Figure 3If the FT-IR spectrum contains a strong peak in the wavenumber region of 1650 cm⁻¹ to 1780 cm⁻¹ (hereinafter also referred to as the "C=O peak"), a carbonyl group is inferred to be present. If the FT-IR spectrum contains a broad and weak peak near the wavenumber region of 2500 cm⁻¹ to 3300 cm⁻¹ in addition to the C=O peak, the carbonyl group is inferred to be the C=O of a carboxylic acid. Furthermore, if the FT-IR spectrum contains strong peaks in the wavenumber regions of 1000 cm⁻¹ to 1150 cm⁻¹ and 1200 cm⁻¹ to 1300 cm⁻¹ in addition to the C=O peak, the carbonyl group is inferred to be the C=O of an ester. Furthermore, if the FT-IR spectrum contains a strong peak in the wavenumber region of 1025 cm⁻¹ to 1250 cm⁻¹ in addition to the C=O peak, the carbonyl group is inferred to be the C=O of an aliphatic ketone. Furthermore, if the FT-IR spectrum contains a strong peak in the wavenumber range of 1215 cm⁻¹ to 1325 cm⁻¹ in addition to the C=O peak, it is inferred that the carbonyl group is the C=O of an aromatic ketone. Conversely, if the FT-IR spectrum contains a weak peak near wavenumber 2720 cm⁻¹ in addition to the C=O peak, it is inferred that the carbonyl group is the C=O of an aldehyde.

[0053] FT-IR spectra can also be resolved using the methods described above. However, resolving FT-IR spectra using these methods may not be easy for users unfamiliar with such methods. Furthermore, because the FT-IR spectra of compounds with complex structures exhibit complex spectral patterns, even experienced users may find it difficult to perform accurate resolution.

[0054] Therefore, in the data analysis method of this embodiment, by using a data analysis device equipped with a learned model to analyze FT-IR spectra, high-precision and easy analysis of FT-IR spectra can be achieved. The learned model described above is a mathematical model obtained through machine learning in the following manner: when an FT-IR spectrum is input, the output indicates whether the input FT-IR spectrum contains peaks originating from learned atomic groups. The inventors of this application have successfully generated a learned model capable of highly accurate analysis of whether an FT-IR spectrum (the object of analysis input to the learned model) contains peaks originating from learned atomic groups (atomic groups with three or more atoms) by performing machine learning on atomic groups with three or more atoms. The method for generating the learned model described above will be explained below.

[0055] Figure 4 This is a diagram showing the detailed structure of the control device 200. (Refer to...) Figure 1 and Figure 4The control device 200 includes an analysis and control unit 210, a data processing unit 220, and a data creation unit 230. In this embodiment, the data is generated by the processor 201 ( Figure 1 The analysis and control unit 210, the data processing unit 220, and the data production unit 230 are implemented by executing the program stored in the storage device 203. However, this is not the only option; these units can also be implemented using dedicated hardware (electronic circuits).

[0056] The analysis control unit 210 is configured to control the analysis device 100. The data processing unit 220 generates signals indicating the state of the analysis device 100 based on the outputs of various sensors (including the detector 130) installed in the analysis device 100, and outputs the generated signals to the analysis control unit 210. The analysis control unit 210 controls the analysis device 100 to analyze the sample M based on the signals received from the data processing unit 220. As a result, the detector 130 detects infrared interference light representing the characteristics of the sample M.

[0057] The data processing unit 220 performs an A / D conversion on the detection signal input from the detector 130 and records the converted digital signal in the storage device 203. As a result, an interferogram is recorded in the storage device 203. When the recording of the interferogram is complete, the data processing unit 220 conveys this information to the data production unit 230. The data production unit 230 produces an FT-IR spectrum by performing a Fourier transform on the interferogram recorded by the data processing unit 220.

[0058] In acquiring FT-IR spectra for qualitative analysis of compounds with unknown molecular structures (hereinafter also referred to as "unknown compounds"), the user places the unknown compound as sample M in the analysis device 100. The placed unknown compound is analyzed by the analysis device 100. The interferogram is recorded in the storage device 203 through this analysis. The data production unit 230 uses the interferogram to produce the FT-IR spectrum of the unknown compound and saves it in the storage device 203. This FT-IR spectrum is saved separately from the learning data described later in the storage device 203.

[0059] The FT-IR spectra of the unknown compounds obtained as described above are analyzed by the data analysis device 500 (see reference 500) described later. Figure 13The data is analyzed. Details will be described later. The user can transfer data from storage device 203 to data analysis device 500 using a portable storage medium 600 (e.g., a memory card, memory stick, or memory disk). Control device 200 is configured to load and unload storage medium 600 and to communicate with the installed storage medium 600. The user can copy or move FT-IR spectral data from storage device 203 to storage medium 600 by operating control device 200 using an input device (not shown).

[0060] On the other hand, when the analysis device 100 analyzes compounds with known molecular structures (hereinafter also referred to as "known compounds") to obtain data for machine learning (hereinafter also referred to as "learning data"), the user places the known compound as a sample M in the analysis device 100. The placed known compound is analyzed by the analysis device 100. The interferogram is recorded in the storage device 203 by the analysis. The data production unit 230 uses the interferogram to produce the FT-IR spectrum of the known compound. Then, the data production unit 230 uses the obtained FT-IR spectrum to produce learning data and saves it in the storage device 203. The learning data involved in this embodiment will be described in detail below.

[0061] In this embodiment, server 300 performs machine learning (e.g., supervised learning). Server 300 is configured to communicate with control device 200 and obtain learning data from control device 200 via communication. However, it is not limited to this; users can also use the aforementioned storage medium 600 to transfer learning data from control device 200 to server 300.

[0062] Server 300 uses the learning data acquired from control device 200 as training data to perform machine learning to generate a learned model capable of analyzing whether peaks in FT-IR spectra originate from specified atomic groups (hereinafter also referred to as "target atomic groups"). In this embodiment, labeled (correct answer) data is used as the learning data. The number of atomic groups included in the target atomic group is arbitrary, and the target atomic group includes at least one atomic group with three or more atoms. In this embodiment, atomic groups No. 1 to No. 19 (see below) are used. Figure 8 Set as the object atom group. Before learning, for example, the user sets the object atom group in the data production department 230.

[0063] Before the analysis is performed by the analysis device 100, the data production unit 230 acquires sample information (i.e., information related to sample M). For example, sample information input by a user via an input device (not shown) is input to the data production unit 230. Alternatively, sample information read from a label mounted on the container of sample M by a reader (not shown) can also be input to the data production unit 230. The data production unit 230 uses the interferogram recorded in the storage device 203 through the analysis described above to produce the FT-IR spectrum of the known compound and saves it in the storage device 203. At this time, the data production unit 230 saves the FT-IR spectrum in association with the group information of sample M. In this embodiment, the step of producing the FT-IR spectrum of the known compound and the step of saving the produced FT-IR spectrum in association with the group information of sample M in the storage device 203 are respectively equivalent to an example of the "analysis step" and "saving step" in the method for generating a learned model according to this disclosure.

[0064] If a sample M (a known compound) has at least one of the target groups, the group information of sample M indicates which target group sample M possesses. Group information can also be indicated by a group number. For example, if sample M has group No. 3 (…), the target group can be identified by a group number. Figure 8 In the case of sample M having group No. 7 and group No. 16, the content of the group information is "3". Figure 8 In the case of sample M, the group information is "7, 16". On the other hand, if sample M does not have the target group, the group information of sample M indicates that sample M does not have the target group. The group information can also be represented by a number other than the aforementioned group number (e.g., "0") to indicate that sample M does not have the target group.

[0065] In this embodiment, the sample information includes the compound name of sample M. The data creation unit 230 obtains the group information of sample M based on the compound name of sample M. The data creation unit 230 is configured to communicate with the compound database 400. The compound database 400 stores information on various compounds in a searchable manner. By accessing the compound database 400 and performing a search using the compound name of sample M, the data creation unit 230 can obtain the group information of sample M. The data creation unit 230 is configured to transform the compound name represented by the sample information into a name that follows the nomenclature adopted by the compound database 400. Therefore, even if the sample information displays the compound name using a nomenclature different from that of the compound database 400, the data creation unit 230 can transform the compound name to match the compound database 400 to perform the aforementioned search. Furthermore, the method by which the data creation unit 230 obtains group information is not limited to the method described above. For example, group information may also be included in the sample information input to the data creation unit 230.

[0066] The learning data used is, for example, the FT-IR spectrum obtained by measuring substances containing known atomic groups as input data, and a pair of data in which the presence or absence of the contained atomic group is the correct answer data. This time, the existing compound DB was used. Figure 5 This is a graph illustrating an example of data used in learning. (See reference...) Figure 5 The learning data includes FT-IR spectra and atomic group information. For example, atomic group information indicating the presence or absence of atomic groups of each object is obtained from sample information. In this embodiment, data that has been numerically quantified as described above is used as atomic group information. FT-IR spectra are obtained as the analysis results of the analysis apparatus 100. In this embodiment, the FT-IR spectra are also numerically quantified. The FT-IR spectra included in the learning data can also be data in which the absorbance is numerically expressed for each wavenumber. For example, the absorbance can also be shown at predetermined intervals (e.g., every 4 cm⁻¹) within a predetermined wavenumber range, such as "0.15, 0.18, 0.25, …". As described above, in this embodiment, the numerically quantified data of the FT-IR spectra and atomic group information are set as the learning data.

[0067] Figure 6 This is a diagram showing the detailed structure of server 300. (Refer to...) Figure 4 and Figure 6 In this embodiment, a computer equipped with a processor 301, RAM 302, storage device 303, and communication device 304 is used. Figure 4The server 300 includes a data acquisition unit 310 and a learning execution unit 320. In this embodiment, the data acquisition unit 310 and the learning execution unit 320 are specifically implemented by the processor 301 executing a program stored in the storage device 303. However, this is not a limitation, and these units may also be specifically implemented by dedicated hardware (electronic circuits).

[0068] The data acquisition unit 310 acquires learning data from the control device 200. Furthermore, the data acquisition unit 310 is configured to communicate with both the FT-IR spectrum database 410 and the open library 420. The FT-IR spectrum database 410 and the open library 420 each store the FT-IR spectrum of each compound. The data acquisition unit 310 transforms the data acquired from the FT-IR spectrum database 410 and the open library 420 into the learning data described above (see reference). Figure 5 The same data format.

[0069] Before learning, the unlearned mathematical model M1 (e.g., a neural network) is stored in storage device 303. The neural network is a mathematical model that simulates the neural circuit mechanism of the brain. The learning execution unit 320 uses the learning data acquired by the data acquisition unit 310 to perform machine learning on the mathematical model M1. This generates, for example... Figure 7 The learned model M2 is shown. The object atom cluster is learned through the aforementioned machine learning. That is, the object atom cluster is the learned atom cluster in the learned model M2. In this embodiment, the step of performing machine learning on the mathematical model M1 is equivalent to an example of the "learning step" in the method for generating a learned model disclosed herein.

[0070] Figure 7 This is a diagram used to illustrate the completed learning of model M2. (See reference...) Figure 7 When the FT-IR spectrum and one atomic group (e.g., ) are input to the learned model M2, the learned atomic group is specified. Figure 8When given specified information (e.g., the name of the atomic group) from one of the atomic groups No. 1 to No. 19, the learned model M2 outputs information indicating whether the input FT-IR spectrum (hereinafter referred to as the "input spectrum") contains peaks originating from the atomic group specified by the input specified information (hereinafter referred to as the "specified atomic group"). That is, the learned model M2 functions as a classifier. The information output from the learned model M2 will be referred to as the "model analysis result" below. Furthermore, the case where the input spectrum contains peaks originating from the specified atomic group is referred to as "having an atomic group," and the case where the input spectrum does not contain peaks originating from the specified atomic group is referred to as "not having an atomic group." The learned model M2 can also output "1" as the model analysis result when an atomic group is present and "0" as the model analysis result when no atomic group is present.

[0071] The inventors of this application actually generated a fully trained model M2 and evaluated the classification accuracy of the generated fully trained model M2. As the untrained mathematical model M1, a CNN (Convolutional Neural Network) with convolutional layers, pooling layers, and dense layers was used. Specifically, a CNN was obtained by alternating between three convolutional layers and one pooling layer twice, followed by a two-layer dense connection. The ReLU function was used as the activation function, and a softmax function was used in the output layer (final activation). The fully trained model M2 was generated by performing backpropagation-based machine learning (more specifically, deep training) on ​​the mathematical model M1 using a dataset containing 5276 training data points. The number of epochs, representing the number of training iterations, was set to 30.

[0072] The target atom group (after learning about atom groups) is Figure 8The atomic groups shown are No. 1 to No. 19. The atomic groups (those already studied) include atomic groups with fewer than 3 atoms (hereinafter also called "small atomic groups") and atomic groups with more than 3 atoms (hereinafter also called "large atomic groups"). Specifically, atomic groups No. 1 to No. 7 are equivalent to small atomic groups, and atomic groups No. 8 to No. 19 are equivalent to large atomic groups. Atomic groups No. 1 to No. 7 are fluorine, chlorine, bromine, iodo, hydroxyl, amino, and cyano, respectively. Atomic groups No. 8 to No. 19 are nitro, ester, acrylic acid structure, toluene structure, thiophene structure, pyrimidine structure, toluidine structure (more specifically, o-toluidine structure), benzoic acid structure (a type of aromatic carboxylic acid structure), benzamide structure, salicylic acid structure, benzimidazole structure, and benzothiazole structure, respectively. Groups No. 11 and Nos. 14 through No. 19 are each containing carbon rings, while groups Nos. 12, 13, 18, and 19 are each containing heterocyclic rings.

[0073] The inventors prepared approximately 500 validation data points and used these data to evaluate the learned model M2 generated as described above. While changing the specified atomic groups for the learned model M2, the inventors evaluated the accuracy of the model in correctly classifying the presence of atomic groups (positive) and the absence of atomic groups (negative). The inventors obtained the ROC (Receiver Operating Characteristic) curve and calculated the AUC (Area Under the Curve) based on it. The ROC curve is a line obtained by plotting the results of changing the threshold (cutoff point) while dividing the true positive rate (TPR) on the vertical axis and the false positive rate (FPR) on the horizontal axis, and connecting the plotted data. The AUC is equivalent to the area under the ROC curve and takes a value from 0 to 1. A higher AUC indicates a higher classification accuracy.

[0074] Calculate the AUC for each learned atomic group. Figure 8 This is a graph showing the average AUC (i.e., the average of all evaluation results) for each learned cluster of the learned model M2. (See also...) Figure 8 For large atomic groups (atomic groups No. 8 to No. 19), an AUC of over 0.900 was obtained. The AUC for small atomic groups (atomic groups No. 1 to No. 7) was lower than that for large atomic groups.

[0075] Figure 9This is a diagram showing an example of the evaluation results of the learned model M2 when the specified group is set to group No. 16 (benzamide structure). Figure 10 This is a diagram showing an example of the evaluation results of the learned model M2 when the specified group is set to group No.8 (nitro). Figure 11 This is a diagram illustrating an example of the evaluation results of the learned model M2 when the specified radical is set to radical No. 3 (bromo). Figure 9 and 10 As shown, for the benzamide structure corresponding to the large atomic group and the nitro group, ROC curves were plotted showing the TPR rising to near 1 in the low FPR stage, and a high AUC was obtained. On the other hand, for the bromine group corresponding to the small atomic group, such as... Figure 11 As shown, the ROC curve rises gently, and the AUC is low.

[0076] Figure 12 This is a graph showing the changes in the value of the loss function and the classification accuracy in machine learning for mathematical model M1. Figure 12 In the diagram, line L11 represents the classification accuracy on the training data, and line L12 represents the classification accuracy on the validation data. Line L21 represents the value of the loss function on the training data (train loss), and line L22 represents the value of the loss function on the validation data (validation loss). (See reference...) Figure 12 As shown in the chart, learning was performed during the generation of the learned model M2 until the value of the loss function roughly converged.

[0077] Refer again Figure 7 In this embodiment, the learned model M2 is installed in the data parsing device 500. In this embodiment, a computer equipped with a processor 501, RAM 502, storage device 503, and communication device 504 is used as the data parsing device 500. The learned model M2 is stored in the storage device 503.

[0078] Figure 13 This is a diagram showing the detailed structure of the data parsing device 500. (Refer to...) Figure 7 and Figure 13 The data parsing device 500 includes an analysis application 511, parsing software 512, and a data acquisition unit 520. The analysis application 511 and parsing software 512 are stored together with the learned model M2 in a storage device 503. Figure 7 In this embodiment, by means of processor 501 ( Figure 7The data acquisition unit 520 is specifically implemented by executing the program stored in the storage device 503. Furthermore, the analysis application 511 is an application program that functions as the "request unit," "parsing unit," and "notification unit" involved in this disclosure. This is achieved by the processor 501 (… Figure 7 These components are implemented by executing the analysis application program 511 stored in the storage device 503. Furthermore, the functions implemented in software in this embodiment can also be implemented using dedicated hardware (electronic circuitry).

[0079] The data parsing device 500 is configured to exchange information with both the input device 710 and the notification device 720. The input device 710 is configured to receive input from a user. The input device 710 outputs a signal corresponding to the user's input to the data parsing device 500. Examples of the input device 710 include various pointing devices (mouse, touchpad, etc.), keyboards, and touch panels. Alternatively, the input device 710 may include a smart speaker for receiving voice input. The notification device 720 is configured to notify the user. The data parsing device 500 can notify the user of information via the notification device 720. Examples of the notification device 720 include various displays. The notification device 720 may also have a speaker function. In this embodiment, a touch panel display combining the functions of both is used as the input device 710 and the notification device 720. The input device 710 and the notification device 720 can also be mounted on portable devices such as tablet terminals, smartphones, or wearable devices (i.e., electronic devices that can be carried by the user).

[0080] The data analysis device 500 is configured to be able to load and unload the storage medium 600 and to communicate with the installed storage medium 600. The data analysis device 500 can read data from the installed storage medium 600. The data analysis device 500 may also include a reader, driver, or port for reading data from the storage medium 600. The user acquires the target data for FT-IR spectroscopy according to, for example, the following process, so that the data analysis device 500 can utilize the acquired target data.

[0081] User utilization Figure 1 and Figure 4 The FT-IR spectrometer shown is used to analyze unknown compounds to obtain FT-IR spectra. The obtained FT-IR spectra are stored in storage device 203. Figure 4In this embodiment, after the user copies the FT-IR spectral data from storage device 203 to storage medium 600, the storage medium 600 is installed on data parsing device 500. Thus, data parsing device 500 becomes capable of reading the FT-IR spectral data from storage medium 600. In this embodiment, the FT-IR spectrum within storage medium 600 is considered the object of parsing.

[0082] The analysis application 511 is configured to request the user to input the storage location (e.g., data path) of the analytical object (FT-IR spectrum). When the user inputs the storage location of the analytical object in the storage medium 600 via the input device 710 in response to the request, the data acquisition unit 520 acquires the analytical object from the storage location input by the user.

[0083] Analysis application 511 is configured to request the user to select from all learned atomic groups (in this embodiment, ...) Figure 8 Users can select one or more atomic groups from the options shown (atomic groups No. 1 to No. 19). Users can select any atomic group from the options via the input device 710. The atomic groups selected by the user are input into the analysis application 511.

[0084] The analysis application 511 is configured to: for each atomic group selected by the user in response to the above request, obtain an analysis result indicating whether the analysis object contains a peak originating from that atomic group. Specifically, the analysis application 511 obtains at least one of the model analysis result and the rule base analysis result described below.

[0085] If the user-selected atomic groups include large atomic groups, the analysis application 511 obtains the analytical results related to the selected large atomic group by learning the model M2. The analysis application 511 obtains the model analytical results by inputting the FT-IR spectrum (analysis object) specified by the user and the specified information for specifying the large atomic group selected by the user (specified atomic group) into the learned model M2.

[0086] If the user-selected atomic groups include small atomic groups, the analysis application 511 obtains the analysis results related to the selected small atomic group through the analysis software 512. The analysis software 512 is configured to, when an FT-IR spectrum is provided, use a rule base to analyze whether the provided FT-IR spectrum contains peaks originating from a specified atomic group (hereinafter also referred to as "registered atomic groups"). In this embodiment, registered atomic groups include... Figure 8 The atomic groups shown are No. 1 to No. 7. The analysis application 511 will use the FT-IR spectra (analyzable objects) specified by the user and the small atomic groups (more specifically) selected by the user. Figure 8Information about one of the atomic groups No. 1 to No. 7 (as shown) is input into the analysis software 512. The analysis software 512 then performs rule-based analysis of the FT-IR spectrum. As the analysis result obtained by the analysis software 512, information indicating whether the analyzed object contains peaks originating from the aforementioned small atomic group (i.e., the registered atomic group represented by the input information) is obtained (hereinafter also referred to as the "rule-based analysis result"). For example... Figure 3 The process shown is used to perform rule-based analysis of FT-IR spectra.

[0087] The analysis application 511 controls the notification device 720 to send notifications to the user. Upon obtaining the aforementioned parsing results, the analysis application 511 notifies the user of the parsing results (i.e., the model parsing results and / or the rule base parsing results) via the notification device 720.

[0088] Figure 14 This is a flowchart illustrating the data parsing process performed by the data parsing apparatus 500 according to this embodiment. For example, the process shown in this flowchart begins when the analysis application 511 is started. The user can operate the input device 710 to start the analysis application 511.

[0089] Reference Figure 13 and Figure 14 In step (hereinafter also abbreviated as "S") 11, the analysis application 511 requests the user to input the storage location of the analytical object (FT-IR spectrum) and the atomic group. First, the analysis application 511 causes the notification device 720 to display the first input screen as described below.

[0090] Figure 15 This is a diagram illustrating an example of an input screen (i.e., a first input screen) for determining the storage location of an object to be parsed by a user. See reference. Figure 15 The first input screen displays message M110, a text box M121, a button M122 for displaying a file selection dialog box, and an OK button M130. Since the notification device 720 is a touch panel display, it can sense the location of the touched screen when a user's finger or pen touches it. The user can perform screen operations by touching the screen.

[0091] Message M110 requests the user to input the storage location of the FT-IR spectrum (analysis object). The user can directly input the storage location of the FT-IR spectrum (e.g., data path) into the text box M121 using an on-screen keyboard (virtual keyboard) not shown. Additionally, when the user presses button M122, a file selection dialog box is displayed. The user can also use the file selection dialog box to select the FT-IR spectrum data (file). When the FT-IR spectrum data (file) is selected using the file selection dialog box, the storage location of the FT-IR spectrum is returned to the text box M121. When the user presses the OK button M130 while the storage location of the FT-IR spectrum has been input into the text box M121, the data acquisition unit 520 acquires the FT-IR spectrum from the storage location input by the user. In this embodiment, the step of the data acquisition unit 520 acquiring the FT-IR spectrum is equivalent to an example of the "acquisition step" in the data analysis method disclosed herein. Afterwards, the analysis application 511 causes the notification device 720 to display the second input screen described below.

[0092] Figure 16 This is an example of a screen (i.e., a second input screen) showing input of atomic groups performed by a user. See reference. Figure 16 The second input screen displays message M210, checkbox CB1 for selecting large atomic groups from the options, description bar M221 for each checkbox CB1, checkbox CB2 for selecting small atomic groups from the options, description bar M222 for each checkbox CB2, select all button M230, and OK button M240.

[0093] Message M210 is used to request the user to select a radical. The user can select the large radical shown in description bar M221 using checkbox CB1. Additionally, the user can select the small radical shown in description bar M222 using checkbox CB2. The user can select all options by pressing the select all button M230. When selecting from the options shown in description bars M221 and M222 (in this embodiment...) Figure 8 When the user presses the OK button M240 while one or more atomic groups (No. 1 to No. 19) are selected, the process enters... Figure 14 S12. In Figure 16 In the example shown, hydroxyl, nitro, and toluene structures were selected.

[0094] Refer again Figure 13 and Figure 14In S12, the analysis application 511 determines whether the groupings selected by the user in S11 include large groupings. Then, if the groupings selected by the user include large groupings ("yes" in S12), in S13, the analysis application 511 inputs the FT-IR spectrum (analysis object) specified by the user in S11 and the specified information for specifying a particular grouping among the large groupings selected by the user in S11 (specified grouping) to the learned model M2, thereby obtaining the model analysis result. S13 in this embodiment corresponds to an example of the "input step" in the data analysis method disclosed herein. Afterwards, in S14, the analysis application 511 determines whether the analysis of the FT-IR spectra for all large groupings selected by the user has been completed.

[0095] If two or more large atomic groups are selected in S11, FT-IR spectra are analyzed sequentially for each large atomic group. If there are large atomic groups that have been selected by the user but have not yet been analyzed, the result is "No" in S14, and the process returns to S13. In S13, the analysis application 511 changes the specified atomic group to the unanalyzed large atomic group and performs FT-IR spectrum analysis (analysis object) again after learning model M2. The process in S13 is executed the same number of times as the number of large atomic groups selected by the user in S11. Then, when the FT-IR spectrum analysis for all large atomic groups selected by the user is completed ("Yes" in S14), the process proceeds to S15. On the other hand, if no large atomic groups are selected by the user ("No" in S12), the processes in S13 and S14 are not executed, and the process proceeds to S15.

[0096] In S15, the analysis application 511 determines whether the group of atoms selected by the user in S11 includes small groups. Then, if the group of atoms selected by the user includes small groups ("yes" in S15), the analysis application 511 inputs the FT-IR spectrum (analysis object) specified by the user in S11 and information representing a small group among the small groups (registered groups) selected by the user in S11 into the analysis software 512 in S16, thereby obtaining the rule base analysis result. Afterwards, the analysis application 511 determines in S17 whether the analysis of the FT-IR spectrum has ended for all small groups selected by the user.

[0097] If two or more small atomic groups are selected in S11, FT-IR spectra are analyzed sequentially for each small atomic group. If there are small atomic groups that have been selected by the user but have not yet been analyzed, a "No" condition is met in S17, and the process returns to S16. In S16, the analysis application 511 inputs the unanalyzed small atomic groups (registered atomic groups) into the analysis software 512, and the analysis software 512 performs FT-IR spectra analysis (analysis targets) again. The process in S16 is executed the same number of times as the number of small atomic groups selected by the user in S11. Then, when the FT-IR spectra analysis for all small atomic groups selected by the user is completed ("Yes" in S17), the process proceeds to S18. On the other hand, if the atomic groups selected by the user do not include small atomic groups ("No" in S15), the processes in S16 and S17 are not executed, and the process proceeds to S18.

[0098] In S18, the analysis application 511 notifies the user of the parsing results (i.e., model parsing results and / or rule base parsing results) obtained in at least one of S13 and S16 by controlling the notification device 720. The analysis application 511 causes the notification device 720 to display a notification screen as described below, for example.

[0099] Figure 17 This is an example of a screen used to notify the user of the parsing results (i.e., a notification screen). See reference. Figure 17 The notification screen displays the display unit M310, messages M320 and M340, parsing results M330, an append button M350, and an end button M360.

[0100] Display unit M310 shows the object of analysis as an FT-IR spectrum. Message M320 shows explanations related to the analysis result M330. The analysis result M330 is for... Figure 14 In S11, the atomic groups selected by the user indicate whether the parsed object contains peaks originating from that atomic group. Figure 17 The example shown illustrates the analysis results when the user selected the hydroxyl, nitro, and toluene structures. Message M340 provides information related to the add button M350 and the end button M360. When the user presses either the add button M350 or the end button M360, processing begins. Figure 14 S19.

[0101] Reference Figure 17 and Figure 14 In S19, the analysis application 511 determines whether the user wishes to add atomic groups. This is done in the aforementioned notification screen ( Figure 17If the user presses the add button M350 in the notification screen, it is determined that the user wants to add an atomic group ("Yes" in S19), and the process returns to S11. Figure 17 If the end button M360 is pressed, it is determined that the user does not wish to add more atomic groups (in S19, this is "No"). Figure 14 The series of processes shown has ended.

[0102] [Variation Example]

[0103] In the above implementation, the options that the user can select (see...) Figure 16 The options include all learned groups, but may also include only a portion of learned groups.

[0104] Atoms that become objects of machine learning (object atoms) are not limited to Figure 8 The shown atomic groups can be appropriately modified. For example, the o-toluidine structure can be replaced, or its isomers (m-toluidine structure, p-toluidine structure) can be used instead of the o-toluidine structure. Additionally, other aromatic carboxylic acid structures (phthalic acid structure, isophthalic acid structure, terephthalic acid structure) can be used instead of the benzoic acid structure. In the above embodiments, in order to compare the evaluation results of small atomic groups and large atomic groups related to the learned model M2, the target atomic groups (learned atomic groups) include small atomic groups, but the target atomic groups (learned atomic groups) can also include only large atomic groups.

[0105] In the above implementation, the options that the user can select (see...) Figure 16 ) includes small atomic groups included in the registered atomic groups ( Figure 8 The atomic groups shown are No. 1 to No. 7, and the macroatomic groups included in the atomic groups have been studied. Figure 8 The analysis application 511, which analyzes atomic groups No. 8 to No. 19, is configured to: obtain analysis results for selected small atomic groups using analysis software 512, and obtain analysis results for selected large atomic groups using the learned model M2. However, the data analysis device 500 does not necessarily need to have analysis software 512. For example, in Figure 14 In S16, the user can also perform data analysis visually without using the analysis software 512. Alternatively, when the analysis application 511 requests the user to select a group of atoms, the user can only select large groups of atoms. Furthermore, the analysis application 511 can also be configured to obtain analysis results by learning model M2 for the selected large groups of atoms.

[0106] Figure 18 It is shown Figure 16 The image shown is a distorted example. (Refer to...) Figure 18 The screen is identical to the previous one, except that the checkbox CB2 for selecting small atomic groups and the description bar M222 for each checkbox CB2 are omitted. Figure 16 The images shown are the same.

[0107] Figure 19 It is shown Figure 14 A diagram showing a modified example of the processing. In Figure 19 The processing shown omits... Figure 14 S12 and S15 to S17 are shown. Additionally, in Figure 19 In the process shown, in S11, instead of Figure 16 The image shown uses Figure 18 The screen shown. Therefore, small atomic groups are not included in the atomic groups that the user can select in S11. Large atomic groups are always selected in S11 (e.g., Figure 8 (Atomic groups No. 8 to No. 19 are shown). Figure 19 S13, S14, S18, and S19 are the same as those mentioned above. Figure 14 S13, S14, S18, and S19 are the same. Figure 19 In the processing shown, no parsing based on parsing software 512 is performed. Therefore, the data parsing device 500 may not need to have parsing software 512.

[0108] The machine learning method is not limited to the method used in the above implementation, but can be any method. The learned model can also consist of multiple classifiers.

[0109] The structure of a Fourier transform infrared spectrometer for acquiring FT-IR spectra is not limited to... Figure 1 The structure shown can be modified appropriately. For example, in Figure 1 The structure shown detects infrared interference light reflected by sample M, but the structure of the Fourier transform infrared spectrometer can be modified to detect infrared interference light transmitted through sample M.

[0110] The server 300 and data analysis device 500 can also be integrated into the Fourier transform infrared spectrometer. Alternatively, the control device 200 can also function as both the server 300 and the data analysis device 500.

[0111] The above examples illustrate FT-IR spectra with wavenumber on the horizontal axis and absorbance on the vertical axis, but the horizontal axis of an FT-IR spectrum can also be wavelength. Additionally, the vertical axis of an FT-IR spectrum can also be transmittance.

[0112] [Way]

[0113] Those skilled in the art will understand that the above-described exemplary embodiments and their variations are specific examples of the following approaches.

[0114] (First item) A data analysis apparatus according to one method includes an acquisition unit for acquiring an FT-IR spectrum as the object of analysis, a learned model, and an analysis unit for inputting the object of analysis into the learned model. The learned model is obtained by machine learning in such a way that when an FT-IR spectrum is input, it outputs information indicating whether the input FT-IR spectrum contains peaks originating from learned atomic groups. The learned atomic groups include atomic groups with three or more atoms.

[0115] The data analysis apparatus described in the first item is capable of performing high-precision and easy FT-IR spectrum analysis on atomic groups of three or more atoms using the aforementioned learned model. The inventors of this application have successfully generated a learned model (see reference) by performing machine learning on atomic groups of three or more atoms, which can accurately analyze whether an FT-IR spectrum (the analysis object input into the learned model) contains peaks originating from the learned atomic group (atomic group of three or more atoms). Figure 8 ).

[0116] (Second item) In the data parsing apparatus described in the first item, the learned group may also include multiple groups. The learned model may also be configured such that, when an FT-IR spectrum and information for specifying a group included in the learned model are input, the output indicates whether the input FT-IR spectrum contains peaks originating from the group specified by the input information.

[0117] According to the data analysis device described in the second item, it is possible to perform FT-IR spectrum analysis on multiple atomic groups with high precision and ease.

[0118] (Third item) The data parsing apparatus described in the first or second item may further include: a request unit that requests a user to select one or more atomic groups from an option including at least a portion of the atomic groups that have been learned; and a notification unit that causes a notification device to notify the user. The parsing unit may also be configured to: for each of the one or more atomic groups selected in response to the above request, obtain a parsing result indicating whether a peak originating from that atomic group is contained in the parsing object. The notification unit may also be configured to notify the user of the parsing result via a notification device.

[0119] According to the data analysis device described in the third item, it is able to analyze the FT-IR spectra of each atomic group selected by the user and notify the user of the analysis results.

[0120] (Fourth) The data analysis apparatus described in the third item may also include analysis software, which, when an FT-IR spectrum is provided, uses a rule base to analyze whether the provided FT-IR spectrum contains peaks originating from a specified atomic group (registered atomic group). The above options may also include learning atomic groups containing three or more atoms and registered atomic groups containing fewer than three atoms. The analysis unit may also be configured to: when one or more atomic groups selected for the above request contain three or more atoms, obtain analysis results related to the selected atomic groups containing three or more atoms by learning the model. The analysis unit may also be configured to: when one or more atomic groups selected for the above request contain fewer than three atoms, obtain analysis results related to the selected atomic groups containing fewer than three atoms using the analysis software.

[0121] The data analysis device described in the fourth item performs FT-IR spectrum analysis using a learned model and analysis software. The learned model only performs FT-IR spectrum analysis on atomic groups with three or more atoms. Since the learned model only needs to learn atomic groups with three or more atoms, the learning burden can be reduced.

[0122] (Fifth item) In any of the data parsing apparatuses described in the first to fourth items, the learned atomic group may also include at least one of an atomic group containing a carbon ring and an atomic group containing a heterocyclic ring.

[0123] Based on the above structure, it is easy to obtain a fully learned model that can accurately resolve FT-IR spectra through machine learning.

[0124] (Sixth item) In any of the data parsing apparatuses described in the first to fourth items, the learned atomic group may also include one or more atomic groups selected from the group consisting of nitro, ester, acrylic acid structure, toluene structure, thiophene structure, pyrimidine structure, toluidine structure, aromatic carboxylic acid structure, benzamide structure, salicylic acid structure, benzimidazole structure and benzothiazole structure.

[0125] Based on the above structure, it is easy to obtain a fully learned model that can accurately resolve FT-IR spectra through machine learning.

[0126] (Seventh) The data parsing method involved in one approach includes the acquisition steps and input steps described below.

[0127] In the acquisition step, the FT-IR spectrum of a compound with an unknown molecular structure is acquired. In the input step, the FT-IR spectrum acquired in the acquisition step and specified information for identifying a group of atoms consisting of three or more atoms are input to the learned model. The learned model is a mathematical model obtained through machine learning in the following manner: when given the input FT-IR spectrum and specified information, it outputs information indicating whether the input FT-IR spectrum contains peaks originating from a group of atoms consisting of three or more atoms specified in the input specified information.

[0128] The data analysis method described in item seven can perform FT-IR spectrum analysis of atomic groups with more than three atoms with high accuracy and ease by utilizing the learned model described above.

[0129] (Item 8) The method for generating a learned model in one approach includes the analysis steps, storage steps and learning steps described below.

[0130] In the analysis step, Fourier transform infrared spectroscopy (FT-IR) is used to analyze compounds with known molecular structures to obtain FT-IR spectra. In the saving step, the FT-IR spectra obtained in the analysis step are saved in association with the group information of the compounds analyzed in the analysis step. In the learning step, the FT-IR spectra and group information saved in the saving step are used as training data for machine learning to generate a learned model capable of resolving whether peaks originating from a specified group (target group) are present in the FT-IR spectrum. The target group includes groups of three or more atoms. Group information for compounds containing the target group indicates the target group present in the compound. Group information for compounds not containing the target group indicates that the compound does not contain the target group.

[0131] According to the method for generating the fully learned model described in item 8, learning data for machine learning can be easily obtained. Furthermore, by using the obtained learning data for machine learning, a fully learned model capable of resolving FT-IR spectra with high accuracy can be generated.

[0132] (Item 9) A system involved in one method is a system having one or more computers capable of performing the data parsing method described in Item 7 or the method for generating a learned model described in Item 8.

[0133] (Item 10) One method involves a program for causing a computer to perform the data parsing method described in Item 7 or the method for generating a learned model described in Item 8. The program may also be stored on a non-transitory computer-readable medium.

[0134] The embodiments disclosed herein should be considered illustrative rather than restrictive in all respects. The scope of the invention is not shown by the description of the above embodiments, but by the claims, which are intended to include all modifications within the meaning and scope of the claims.

[0135] Explanation of reference numerals in the attached figures

[0136] 100: Analysis device; 110: Interference wave generation unit; 111: Light source; 112: Condenser lens; 113: Collimating lens; 114: Beam splitter; 115: Fixed mirror; 116: Movable mirror; 117: Mirror actuator; 120: Irradiation unit; 121, 123: Condenser lenses; 122: Prism; 124: Pressing mechanism; 130: Detector; 200: Control device; 201, 301, 501: Processor; 202, 302, 502: RAM; 203, 303, 503: Storage device; 204, 304 504: Communication device; 210: Analysis and control unit; 220: Data processing unit; 230: Data production unit; 300: Server; 310: Data acquisition unit; 320: Learning execution unit; 400: Compound database; 410: FT-IR spectrum database; 420: Open library; 500: Data analysis device; 511: Analysis application program; 512: Analysis software; 520: Data acquisition unit; 600: Storage medium; 710: Input device; 720: Notification device; M1: Mathematical model; M2: Model after learning.

Claims

1. A data parsing device, comprising: The acquisition unit acquires data that is analyzed using Fourier transform infrared spectra. The learned model is obtained by machine learning in a manner that outputs information indicating whether the input Fourier transform infrared spectrum contains peaks originating from the atomic group specified by the input specified information when the input Fourier transform infrared spectrum is given as input and specified information is used to specify one atomic group included in the learned atomic group. The request section requests the user to select one or more atomic groups from the options; and The parsing unit inputs the parsing object into the learned model. in, The learned atomic groups include atomic groups with more than 3 atoms. The options include atomic groups with three or more atoms. The analytical unit is configured as follows: Determine whether the group of atoms selected in response to the request from the requesting unit includes a group of atoms with three or more atoms. If the selected group of atoms includes a group of atoms with three or more atoms, the parsing object obtained by the acquisition unit and the specified information for specifying the group of atoms with three or more atoms selected by the user are input into the learned model, thereby performing parsing of the parsing object. If the selected group of atoms does not include a group of atoms with three or more atoms, then the parsing of the parsing object based on the learned model will not be performed.

2. The data parsing apparatus according to claim 1, wherein, It also includes: a notification unit, which enables the notification device to send notifications to the user. The analysis unit is configured to: for one or more atomic groups selected in response to the request, obtain analysis results indicating whether the analysis object contains a peak originating from that atomic group. The notification unit is configured to notify the user of the parsing result via the notification device.

3. The data parsing apparatus according to claim 1, wherein, It also includes analysis software that, when provided with a Fourier transform infrared spectrum, uses a rule base to analyze whether the provided Fourier transform infrared spectrum contains peaks originating from specified atomic groups. The analysis unit is configured to: for one or more atomic groups selected in response to the request, obtain analysis results indicating whether the analysis object contains a peak originating from that atomic group. The options include atomic groups containing three or more atoms in the learned atomic group and atomic groups containing fewer than three atoms in the specified atomic group. The analytical unit is configured as follows: If, in response to the request, one or more atomic groups include atomic groups with three or more atoms, the analysis results related to the selected atomic groups with three or more atoms are obtained through the learned model. If one or more atomic groups selected in response to the request include an atomic group with fewer than three atoms, the analysis results related to the selected atomic group with fewer than three atoms are obtained by the analysis software.

4. The data parsing apparatus according to any one of claims 1 to 3, wherein, The learned atomic groups include at least one of atomic groups containing carbon rings and atomic groups containing heterocycles.

5. The data parsing apparatus according to any one of claims 1 to 3, wherein, The learned atomic groups include one or more atomic groups selected from the group consisting of nitro, ester, acrylic acid, toluene, thiophene, pyrimidine, toluidine, aromatic carboxylic acid, benzamide, salicylic acid, benzimidazole, and benzothiazole.

6. A data parsing method, comprising the following steps: The acquisition steps involve obtaining the Fourier transform infrared spectrum of compounds with unknown molecular structures. The request step asks the user to select one or more atomic groups from an option that includes atomic groups with three or more atoms. The determination step involves determining whether the selected group of one or more atoms includes a group of three or more atoms, and... In the input step, if the selected group of atoms includes a group of atoms with three or more atoms, the acquired Fourier transform infrared spectrum and specified information for specifying the group of atoms with three or more atoms selected by the user are input into the learned model. The learned model is a mathematical model obtained by machine learning in the following manner: when the Fourier transform infrared spectrum and the specified information for specifying the atomic group of the three or more atoms are input, the output indicates whether the input Fourier transform infrared spectrum contains peaks originating from the atomic group of the three or more atoms specified by the input specified information.

7. A system comprising one or more computers performing the method according to claim 6.

8. A computer program product comprising a program for causing a computer to perform the method according to claim 6.

9. A method for generating a learned model, comprising the following steps: The analytical steps, including the analysis device and control device, involve using a Fourier transform infrared spectrometer to analyze compounds with known molecular structures to obtain Fourier transform infrared spectra. In the storage step, the control device associates the acquired Fourier transform infrared spectrum with the analyzed atomic group information of the compound and stores it. as well as The learning process involves a computer capable of communicating with the control device using the stored Fourier transform infrared spectrum and atomic group information as training data to perform machine learning to generate a learned model capable of analyzing whether the Fourier transform infrared spectrum contains peaks originating from specified atomic groups. The defined atomic groups do not include atomic groups with fewer than 3 atoms, and do not include atomic groups with more than 3 atoms. The atomic group information of a compound having the specified atomic group indicates the specified atomic group possessed by the compound, and the atomic group information of a compound not having the specified atomic group indicates that the compound does not have the specified atomic group.

Images