Analysis device, analysis method, and program

Abstract

Provided is an analysis device, an analysis method, and a program with which analysis processing according to a user's request input in a natural language can be executed on measurement data of a substance measured by a predetermined spectroscopy method.　An analysis device 1 comprises: a measurement data acquisition unit 31 that acquires measurement data including the spectrum of a substance to be analyzed measured by a predetermined spectroscopy method; a user request reception unit 32 that receives, in a natural language, a user's request related to analysis of the measurement data; a module selection unit 33 that performs selection processing for selecting a processing module corresponding to the user's request from among a plurality of types of processing modules that perform processing on the measurement data with respect to a natural language model 5 which has been trained for interpreting the natural language; a processing execution unit 34 that causes the processing module selected by the selection processing to execute processing; and a presentation processing unit 35 that presents an analysis result based on the processing of the processing module to the user.

Description

Analysis Device, Analysis Method, and Program

[0001] The present invention relates to an analysis device, an analysis method, and a program.

[0002] Conventionally, spectroscopy methods such as Raman spectroscopy are powerful analytical techniques that can measure substances with high precision and are utilized in various fields such as materials science, chemical analysis, and pharmaceutical development. However, despite their usefulness, there is a major problem that the interpretation of the obtained spectra is very difficult. There are numerous peaks, and advanced expertise is indispensable for accurately identifying them.

[0003] Raman spectroscopy spectrum analysis has been basically a visual operation by skilled experts. Based on experience and knowledge, experts visually confirm the peaks on the spectrum and use their positions, shapes, relative intensities, etc. as clues to infer the structure and composition of the target substance. In order to automate such manual work of experts, analysis methods utilizing machine learning have been proposed. For example, Patent Document 1 and Patent Document 2 describe this type of technology.

[0004] U.S. Patent No. 10,955,362 Specification, U.S. Patent Application Publication No. 2024 / 0369490 Specification

[0005] In order to apply machine learning to the spectrum analysis of spectroscopy methods such as Raman spectroscopy, a great deal of learning costs are involved, and the training of analysts and the efficiency improvement of analysis work have become major issues. However, in machine learning, the model structure is fixed, and there is room for improvement from the perspective of dynamically selecting an optimal analysis method according to user requirements and data characteristics.

[0006] By the way, in recent years, technologies have also been developed that automatically generate responses to analysis instructions and questions in natural language using generative AI and large language models (LLMs). However, general generative AI and information retrieval models cannot sufficiently understand and utilize the structures peculiar to spectral spectra (such as peak - spectrum shapes), making it difficult to optimize analysis and prevent false detections. There is a risk of hallucination, i.e., unfounded answers, and it is difficult to directly use them for measurement data analysis, and there are also issues of reproducibility.

[0007] This invention has been made in view of the above circumstances, and aims to provide an analysis device, analysis method, and program that can perform analysis processing on measurement data of a substance measured by a predetermined spectroscopic method, in accordance with the user's requests input in natural language.

[0008] To achieve the above objective, one aspect of the present invention is an analysis device comprising: a measurement data acquisition unit that acquires measurement data including the spectrum of a substance to be analyzed measured by a predetermined spectroscopic method; a user request reception unit that receives user requests regarding the analysis of the measurement data in natural language; a module selection unit that performs a selection process to cause a natural language model trained to interpret the natural language to select a processing module corresponding to the user's request from among a plurality of processing modules that process the measurement data; a processing execution unit that causes the processing module selected by the selection process to execute processing; and a presentation processing unit that presents the analysis results based on the processing of the processing module to the user.

[0009] Furthermore, one aspect of the present invention is an analysis method comprising: a measurement data acquisition step of acquiring measurement data including the spectrum of a substance to be analyzed measured by a predetermined spectroscopic method; a user request acceptance step of receiving a user's request regarding the analysis of the measurement data in natural language; a module selection step of causing a natural language model that has been trained to interpret the natural language to select a processing module corresponding to the user's request from among a plurality of processing modules that process the measurement data; a processing execution step of causing the processing module selected by the selection process to execute processing; and a presentation processing step of presenting the analysis results based on the processing of the processing module to the user.

[0010] Furthermore, one aspect of the present invention is a program for causing a computer to execute the following steps: a measurement data acquisition step of acquiring measurement data including the spectrum of a substance to be analyzed measured by a predetermined spectroscopic method; a user request acceptance step of receiving a user's request regarding the analysis of the measurement data in natural language; a module selection step of causing a natural language model that has been trained to interpret the natural language to select a processing module corresponding to the user's request from among a plurality of processing modules that process the measurement data; a processing execution step of causing the processing module selected by the selection process to execute processing; and a presentation processing step of presenting the analysis results based on the processing of the processing module to the user.

[0011] According to the present invention, it is possible to provide an analysis device, an analysis method, and a program that can perform analysis processing on measurement data of a substance measured by a predetermined spectroscopic method, in accordance with the user's requests input in natural language.

[0012] This figure shows an analysis system to which an analysis device according to one embodiment of the present invention is applied. This is a block diagram showing the hardware configuration of the analysis device according to this embodiment. This is a functional block diagram showing an example of the functional configuration of the analysis device of the first embodiment. This is a flowchart showing the processing operation of the analysis device of the first embodiment. This is a functional block diagram showing an example of the functional configuration of the analysis device of the second embodiment. This is a flowchart showing the processing operation of the analysis device of the second embodiment. This is a functional block diagram showing an example of the functional configuration of the analysis device of the third embodiment. This is a flowchart showing the processing operation of the analysis device of the third embodiment.

[0013] One embodiment of the present invention will be described below with reference to the drawings. In the description of the second embodiment and subsequent embodiments, components common to the first embodiment will be denoted by the same reference numerals, and their descriptions may be omitted as appropriate.

[0014] <System Configuration> First, the overall system configuration will be explained. Figure 1 shows an analysis system 100 to which an analysis device 1 according to one embodiment of the present invention is applied. The analysis system 100 automatically interprets the content of the analysis request and measurement conditions described by the user in natural language, along with measurement data of a predetermined spectroscopic method, performs the analysis that is most suitable for the conditions and purpose, and presents the analysis results in a visual and easy-to-understand format.

[0015] The specified spectroscopic methods include, for example, Raman spectroscopy, which utilizes Raman scattering light; infrared spectroscopy, which uses infrared light irradiation; and fluorescence spectroscopy, which utilizes fluorescence generated when a substance is excited. The specified spectroscopic methods are not particularly limited, but Raman spectroscopy will be described below as an example of an embodiment.

[0016] The analysis system 100 is implemented by an analysis device 1 that transmits and receives various information with the user terminal 2 via a communication network such as the Internet. The analysis device 1 is an information processing device that performs analysis processing on measurement data of substances measured by Raman spectroscopy. In this embodiment, the analysis device 1 cooperates with a natural language model 5 via a communication network such as the Internet, selects a processing module suitable for the user's requirements, and performs processing related to the analysis of measurement data using that processing module.

[0017] Natural Language Model 5 is a Large Language Model (LLM) that has been trained to interpret natural language and provide responses. It is preferable that Natural Language Model 5 be fine-tuned to allow the selection of processing modules according to the user's requirements for measurement data obtained by spectroscopy.

[0018] User terminal 2 is an information processing device used by a user who measures a substance using a predetermined spectroscopic method. User terminal 2 may exchange various information with analysis device 1 using a pre-installed program, or it may exchange various information through a web browser.

[0019] <Hardware Configuration> Next, an example of the hardware comprising the analysis device 1 will be described. Figure 2 is a block diagram showing the hardware configuration of the analysis device 1 according to this embodiment. The analysis device 1 includes a CPU (Central Processing Unit) 11 as a processor, a ROM (Read Only Memory) 12, a RAM (Random Access Memory) 13, a bus 14, an input / output interface 15, an output unit 16, an input unit 17, a storage unit 18, a communication unit 19, and a drive 20.

[0020] The CPU 11 executes various processes according to the program recorded in the ROM 12 or the program loaded from the storage unit 18 into the RAM 13. The RAM 13 also stores data necessary for the CPU 11 to execute various processes. The CPU 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output interface 15 is also connected to this bus 14.

[0021] The input / output interface 15 is connected to an output unit 16, an input unit 17, a storage unit 18, a communication unit 19, and a drive 20. The output unit 16 consists of a display, speakers, etc., and outputs various information as images and sounds. The input unit 17 consists of a keyboard, mouse, etc., and inputs various information. The storage unit 18 consists of a hard disk, DRAM (Dynamic Random Access Memory), etc., and stores various data. The communication unit 19 communicates with other devices via a network, including the Internet.

[0022] A removable media 21, such as a magnetic disk, optical disk, magneto-optical disk, or semiconductor memory, is appropriately mounted on the drive 20. Programs read from the removable media 21 by the drive 20 are installed in the storage unit 18 as needed. The removable media 21 can also store various data stored in the storage unit 18, just like the storage unit 18.

[0023] The hardware configuration described here is merely an example. The computer described in this embodiment, including the analysis device 1, may have the same configuration as that in Figure 2, or it may have a different configuration. Furthermore, the computer may consist of two or more computers. The user terminal 2 in Figure 1 is, for example, a personal computer, tablet, or smartphone having a configuration similar to the hardware configuration shown in Figure 2.

[0024] <Functional Configuration of the Analysis Device of the First Embodiment> Next, the functional configuration of the analysis device 1 of the first embodiment will be described. Figure 3 is a functional block diagram showing an example of the functional configuration of the analysis device 1 according to the first embodiment.

[0025] As shown in Figure 3, the analysis device 1 comprises a measurement data acquisition unit 31, a user request reception unit 32, a module selection unit 33, a processing execution unit 34, and a presentation processing unit 35.

[0026] The measurement data acquisition unit 31 performs processing to acquire measurement data, including the spectrum of the substance to be analyzed, measured by a predetermined spectroscopic method.

[0027] The spectrum included in the measurement data is, for example, the intensity distribution of light at each wavelength obtained by irradiating a substance with light and performing spectral analysis. The measurement data may be acquired, for example, by uploading and saving it by the user, or it may be transmitted from the measuring instrument via a communication network such as the Internet.

[0028] Furthermore, the measurement data may include conditional information such as the emission conditions of the light source, the measurement conditions of the measuring instrument, ambient temperature, and date and time. This conditional information may also be obtained, for example, by being uploaded or saved by the user, or transmitted from the measuring instrument via a communication network such as the Internet.

[0029] The user request reception unit 32 executes a process to receive user requests in natural language regarding the processing to be performed for the analysis of measurement data.

[0030] In this embodiment, the user request receiving unit 32 acquires natural language indicating the user's request from the user terminal 2 via a communication network. For example, the user request receiving unit 32 acquires natural language entered into a text box on the request input screen displayed on the user terminal 2. The user request receiving unit 32 may also be in a format where a virtual character accepts the user's request through a dialogue screen such as a chat. Alternatively, the user request receiving unit 32 may receive the user's request through the input unit 17, with the operator inputting the request.

[0031] The module selection unit 33 performs a process to select a processing module for the natural language model 5, which performs a process to select a processing module that corresponds to the user's request from among several types of processing modules for the measurement data.

[0032] A processing module is an analysis tool consisting of blocks or units that perform predetermined processing on measurement data. Multiple candidates for processing modules, each performing different types of processing, are configured. For example, a processing module might perform processing related to displaying the waveform of the spectrum contained in the measurement data. Alternatively, a processing module might execute processing that performs analysis on the measurement data using a predetermined method.

[0033] The processing module may be a single process such as PLS (Partial Least Squares regression), polynomial fitting, or low-pass filtering, or it may be a combination of multiple processes, such as a combination of multivariate analysis and clustering. When combining multiple processes, it may also include specifying the order of processing, such as performing multivariate analysis on the measured data before clustering.

[0034] The module selection unit 33 sends a prompt to the natural language model 5 to select a processing module that corresponds to the user's request, and selects a processing module based on the output of the natural language model 5. In the first embodiment, the module selection unit 33 selects a processing module according to the user's request using a direct selection method in which the natural language model 5 selects a processing module on its own.

[0035] In the direct selection method, a finely tuned natural language model 5 is used. For example, a large number of pairs of input examples such as "I want to distinguish between the spectra of sample A, sample B, and sample C," "I want to group multiple Raman data," and "Tell me the best way to divide samples into clusters" and output "Multivariate analysis → clustering (e.g., PCA → k-means method) + result display" are created, and the natural language model 5, which has been finely tuned and trained based on these pairs, selects a processing module according to the user's request.

[0036] The processing execution unit 34 executes processing to enable the processing by the processing module selected by the processing of the module selection unit 33.

[0037] The processing execution unit 34 may present the selected processing module and its contents to the user, and execute the processing module after obtaining the user's permission to execute it. Alternatively, the processing execution unit 34 may instruct the natural language model 5 to execute the selected processing module after the module selection unit 33 has performed the processing module selection process.

[0038] The presentation processing unit 35 presents the execution result of the processing module to the user as an analysis result. In this embodiment, the presentation processing unit 35 transmits information for displaying the analysis result to the user terminal 2 via a communication network.

[0039] Next, an example of the analysis process flow by the analysis device 1 of the first embodiment will be described. Figure 4 is a flowchart showing the operation of the analysis device 1 of the first embodiment.

[0040] In step S1, the measurement data acquisition unit 31 acquires measurement data from the user through an upload operation on the user terminal 2 or the like. The measurement data is, for example, spectral data including the spectrum of a substance measured by Raman spectroscopy. The measurement data may also include information such as the emission conditions of the light source used for measurement, the measurement conditions of the measuring instrument, ambient temperature, date and time, etc.

[0041] In step S2, the user request reception unit 32 receives, in natural language, the user's request for the processing of measurement data through the user terminal 2. The user's request is, for example, a natural language request such as "Differentiate the spectra of sample A, sample B, and sample C", "Since spectra with different concentrations were taken, draw a calibration curve", "Subtract the background", etc.

[0042] In step S3, the module selection unit 33 executes a process of causing the natural language model 5 to select a processing module corresponding to the acquired user request. In this embodiment, the natural language model 5 fine-tuned for Raman spectroscopy measurement interprets the natural language of the user's request, and a processing module corresponding to the user's request is selected.

[0043] In step S4, the processing execution unit 34 executes the processing by the processing module selected by the natural language model 5 according to the user's request. The execution procedure of the processing module by the processing execution unit 34 is not particularly limited. For example, the processing execution unit 34 may present the selection result of the processing module of the natural language model 5 to the user and execute the processing module when obtaining the user's permission, or may automatically execute the processing module.

[0044] In step S5, the presentation processing unit 35 visually or numerically displays the analysis result to the user. In this display process, visualization and layout formatting are executed. In the presentation to the user by the presentation processing unit 35, display processing such as numerical values and graphs may be performed by a result display processing module selected by the natural language model 5 based on the user request, the analysis result, or both. Also, a brief supplementary explanation by the natural language model 5 may be displayed.

[0045] As described above, the analysis apparatus 1 of the present embodiment includes a measurement data acquisition unit 31 that acquires measurement data including the spectrum of a substance to be analyzed measured by a predetermined spectroscopic method, a user request reception unit 32 that receives a user's request regarding the analysis of the measurement data in natural language, a module selection unit 33 that performs a selection process for selecting a processing module corresponding to the user's request from a plurality of types of processing modules that process the measurement data with respect to a natural language model 5 in which learning for interpreting natural language has been performed, a processing execution unit 34 that causes the processing module selected by the selection process to execute processing, and a presentation processing unit 35 that presents the analysis result based on the processing of the processing module to the user.

[0046] Further, the analysis method of the present embodiment includes a measurement data acquisition step of acquiring measurement data including the spectrum of a substance to be analyzed measured by a predetermined spectroscopic method, a user request reception step of receiving a user's request regarding the analysis of the measurement data in natural language, a module selection step of performing a selection process for selecting a processing module corresponding to the user's request from a plurality of types of processing modules that process the measurement data with respect to a natural language model 5 in which learning for interpreting natural language has been performed, a processing execution step of causing the processing module selected by the selection process to execute processing, and a presentation processing step of presenting the analysis result based on the processing of the processing module to the user.

[0047] Further, the program of the present embodiment causes the analysis apparatus 1, which is a computer, to execute a measurement data acquisition step of acquiring measurement data including the spectrum of a substance to be analyzed measured by a predetermined spectroscopic method, a user request reception step of receiving a user's request regarding the analysis of the measurement data in natural language, a module selection step of performing a selection process for selecting a processing module corresponding to the user's request from a plurality of types of processing modules that process the measurement data with respect to a natural language model 5 in which learning for interpreting natural language has been performed, a processing execution step of causing the processing module selected by the selection process to execute processing, and a presentation processing step of presenting the analysis result based on the processing of the processing module to the user.

[0048] In this way, the analysis device 1, analysis method, or program is configured to automatically construct spectral analysis procedures that would otherwise require specialized knowledge, simply by the user giving instructions in natural language, effectively reducing the user's burden. Furthermore, since the natural language model 5 selects a processing module from among several types of processing modules according to the user's requirements, pre-processing and pattern analysis of spectral analysis can be clearly separated from the interpretation processing of the natural language model 5, ensuring reproducibility. This also prevents hallucination, eliminates unfounded explanations and erroneous inferences by generated AI, and improves the reliability of the analysis results. Moreover, the system can be expanded simply by adding new processing modules.

[0049] Furthermore, the processing module of this embodiment includes at least one of the following: analysis of the spectrum included in the measurement data, and processing related to the display of the spectral waveform.

[0050] This ensures that processing by the processing module is performed based on predefined and established analysis or display methods, thereby further enhancing the effects of preventing hallucination and improving reproducibility. Specifically, for example, if the signal-to-noise ratio of the measurement data is low, the natural language model 5 selects a smoothing processing module based on the user's request and the characteristics of the measurement data, and that module performs numerical processing such as an established Savitzky-Golay filter. In this way, the natural language model 5 is only responsible for selecting the processing module, and the actual numerical processing is performed by a deterministic algorithm, so the same output is always obtained for the same input, ensuring reproducibility. Furthermore, unlike when the natural language model 5 directly interprets spectral data to generate analysis results, the processing module operates based on scientifically established methods, eliminating unfounded inferences and misinterpretations, and improving the reliability of the analysis results.

[0051] <Second Embodiment> In the first embodiment, an example of a direct selection method in which the natural language model 5 alone selects a processing module was described. Next, we will describe an analysis device 1a of the second embodiment that employs a RAG selection method in which a search-enhanced RAG (Retrievable-Augmented Generation) is used to select a processing module. In the RAG selection method of the second embodiment, in addition to the user's requests input by the user, the characteristic information of the measurement data to be analyzed is referenced, and a processing module that meets the conditions is selected based on the search results of the RAG database 6, which stores past analysis cases and a rule base such as predetermined policies. In the second embodiment, as in the first embodiment, a finely tuned natural language model 5 is used, but a general-purpose natural language model 5 that is not finely tuned may also be used.

[0052] Figure 5 is a functional block diagram showing an example of the functional configuration of the analysis device 1a in the second embodiment. In the second embodiment, a RAG database 6 used by the natural language model 5 is constructed. The RAG database 6 may be constructed in the storage unit 18, in a connected external storage device, or in a cloud server.

[0053] The RAG database 6 stores relevant information that allows the natural language model 5 to select a processing module according to the user's requests. In the second embodiment, the relevant information is a rule-based system that corresponds to the user's requests, the features of the measurement data described later, or both. For example, the relevant information consists of judgment criteria (policies) for each case, such as when the number of clusters is specified / not specified, when the S / N ratio is high / low, when the fluorescence intensity is high / low, and the history of past processing module selections.

[0054] Furthermore, if user requests and judgment criteria conflict, the rule base may also include a provision for determining which of the user requests or pre-registered judgment criteria takes precedence. For example, if a request is made for operation when the signal-to-noise ratio (S / N) is low despite the S / N ratio being above a certain level (smoothing, e.g., Savitzky-Golay filter), the system may be configured to select a processing module that does not perform smoothing.

[0055] The analysis device 1a of the second embodiment includes a measurement data acquisition unit 31, a user request reception unit 32, a module selection unit 33, a processing execution unit 34, a presentation processing unit 35, an extraction processing unit 36, and a search condition generation unit 37.

[0056] Referring to Figure 6, the extraction processing unit 36 ​​and the search condition generation unit 37 of the analysis device 1a of the second embodiment will be described. Figure 6 is a flowchart showing the operation of the analysis device 1a of the second embodiment.

[0057] In step S11, the measurement data acquisition unit 31 acquires measurement data from the user through an upload operation on the user terminal 2, similar to the process in step S1 of the first embodiment.

[0058] In step S12, the user request receiving unit 32 receives user requests for processing of measurement data in natural language through the user terminal 2, similar to the processing in step S2 of the first embodiment.

[0059] In step S13, the module selection unit 33 obtains the purpose of measurement from the user's request. For example, if the user's request is "I want to distinguish between the spectra of sample A, sample B, and sample C," the purpose of measurement is interpreted as "distinguishing." The purpose of measurement may be interpreted from the user's request by the natural language model 5, or it may be obtained based on the user's selection from pre-set options.

[0060] In step S14, the extraction processing unit 36 ​​performs processing to extract feature information indicating the waveform characteristics from the measurement data. The feature information extraction processing by the extraction processing unit 36 ​​is performed, for example, by a pre-configured processing module. The feature information is, for example, the signal-to-noise ratio (S / N) and the number of peaks of the measurement data. The S / N calculation in Raman spectroscopy is performed in the region without Raman peaks (e.g., 2100-2400 cm⁻¹). －１ The standard deviation of the data is denoted as N, and the height of the highest Raman peak is denoted as S.

[0061] In step S15, the search condition generation unit 37 executes a process to generate search conditions based on user requests and feature information. In this embodiment, the search condition generation unit 37 generates search conditions using the natural language model 5 based on the purpose of measurement obtained from the user requests and the feature information extracted by the extraction processing unit 36. For example, if the user requests "I want to distinguish between the spectra of sample A, sample B, and sample C," and the S / N ratio of the feature information is low and the fluorescence intensity is high, then search conditions such as "Classification (with specified number of clusters)," "Low S / N," and "High fluorescence intensity" will be generated. The generation of search conditions may use the natural language model 5 or a rule-based algorithm.

[0062] In step S16, the module selection unit 33 refers to the generated search conditions and the rule base registered in the RAG database 6, and selects a processing module that matches the conditions. These conditions may include that the selected processing module satisfies the pre- and post-constraint relationships. Pre- and post-constraint relationships are set for each processing module and are information that restricts the processing performed before and after the processing of the processing module.

[0063] In step S17, the presentation processing unit 35 displays the analysis results to the user visually or numerically, similar to the processing in step S5 of the first embodiment.

[0064] The following describes a specific implementation of the selection of processing modules in step S16 using pre- and post-processing constraints. The module selection unit 33 refers to the input / output data format, preconditions, and metadata of subsequent processes set for each processing module and dynamically constructs a processing dependency graph. For example, if the output data format of the baseline removal module is "corrected spectral data" and the input data format of the smoothing module is "spectral data (corrected possible)", it is determined that data can be passed between these modules. Also, since the PCA module has "precondition: denoised" and the k-means method module has "precondition: dimensionality reduced, parameter: number of clusters k required", the order "smoothing → PCA → k-means method" is automatically determined.

[0065] The following describes the coordinated execution between processing modules by the processing execution unit 34. Each processing module implements a standardized API interface, receives the processing results from the previous stage, executes the processing, and passes the results to the next stage. As a specific example, for measurement data with a low S / N ratio, (1) the baseline removal module removes the fluorescence background by polynomial fitting, (2) the Savitzky-Golay filter module receives the output and performs smoothing processing while automatically adjusting the window width, (3) the PCA module performs dimensionality reduction of the preprocessed data, and (4) finally the k-means method module performs clustering. In this series of processes, data verification between each module, error handling, and recording of processing logs are performed automatically, enabling stable execution of complex analysis flows.

[0066] As described above, the analysis device 1a of the second embodiment further comprises an extraction processing unit 36 ​​that extracts feature information indicating waveform characteristics from measurement data, and a search condition generation unit 37 that generates search conditions based on the feature information. The module selection unit 33 performs processing to cause the natural language model 5 to select a processing module based on the user's request, the search conditions, and a rule base that pre-defines rules according to the characteristics of the measurement data.

[0067] This allows RAG searches to be performed based on search criteria that reflect the characteristics of the measurement data, thereby improving the accuracy of the analysis and the reliability of the explanation. Furthermore, the procedure for generating RAG search criteria can be easily added and updated, ensuring consistency in the analysis while allowing for flexible application to Raman spectroscopy, other spectroscopic methods (infrared, fluorescence, etc.), and various analytical purposes (quantitative analysis, state estimation, etc.).

[0068] In the configuration of the second embodiment, the selected processing module and the processing module approved for use by the user may be registered in the RAG database 6 along with user information and characteristic information, and used in subsequent analyses. This makes it possible to reproduce the analysis results more reliably. In particular, by accumulating the combination of processing modules and their execution order along with the characteristic information of the measurement data and the user's requests, the module selection unit 33 can prioritize the selection of processing flows that have been effective in the past for similar measurement conditions and user requests, thereby further improving the accuracy and reproducibility of the analysis.

[0069] <Third Embodiment> Next, we will describe the analysis device 1b of the third embodiment, which presents the user with supplementary explanations along with the analysis results. In the third embodiment, the user can choose between a "direct display method" that displays the analysis results as they are, and a "display method with explanations" that provides justification by referring to papers.

[0070] Figure 7 is a functional block diagram showing an example of the functional configuration of the analysis device 1b of the third embodiment. In the third embodiment as well, a RAG database 6a used by the natural language model 5 is constructed. The RAG database 6a may be constructed in the storage unit 18, in a connected external storage device, or in a cloud server.

[0071] The RAG database 6a of the third embodiment stores relevant information that is structured to allow the natural language model 5 to select a processing module according to the user's request. The relevant information in the third embodiment includes academic papers and materials related to a predetermined spectroscopic method such as Raman spectroscopy. Tag information for use in the search described later may be added to the relevant information.

[0072] The analysis device 1b of the third embodiment includes a measurement data acquisition unit 31, a user request reception unit 32, a module selection unit 33, a processing execution unit 34, a presentation processing unit 35, an extraction processing unit 36a, a search condition generation unit 37a, and an explanatory information creation unit 38.

[0073] In the third embodiment, the user can choose whether or not to display explanatory information generated by the explanatory information creation unit 38 along with the analysis results. Referring to Figure 8, the processing when the user selects to display explanatory information in the analysis device 1b of the third embodiment will be described. Figure 8 is a flowchart showing the operation of the processing of the analysis device 1b of the third embodiment. If the user chooses not to display explanatory information along with the analysis results, the same processing as in the first embodiment will be performed.

[0074] In step S21, the measurement data acquisition unit 31 acquires measurement data from the user through an upload operation on the user terminal 2, similar to the process in step S1 of the first embodiment.

[0075] In step S22, the user request receiving unit 32 receives the user's request for processing the measurement data in natural language through the user terminal 2, similar to the processing in step S2 of the first embodiment.

[0076] In step S23, the module selection unit 33 performs a process to cause the natural language model 5 to select a processing module that corresponds to the acquired user request, similar to the process in step S3 of the first embodiment.

[0077] In step S24, the processing execution unit 34 executes processing by the processing module selected by the natural language model 5 according to the user's request, similar to the processing in step S4 of the first embodiment.

[0078] In step S25, the extraction processing unit 36a performs a process to extract feature information that indicates the characteristics of the waveform from the measurement data. The feature information extracted by the extraction processing unit 36 ​​is, for example, characteristic peaks or patterns. Peaks are obtained, for example, by second derivative, prominence value, or a combination thereof.

[0079] In step S26, the search condition generation unit 37a generates search conditions based on the extracted feature information. The search condition generation unit 37a sends a prompt to the natural language model 5 instructing it to create search conditions that specify, for example, tolerance range, wavenumber range, peak combination, related categories, etc., based on feature information such as peaks. The search condition generation unit 37a may also reflect supplementary information, such as "Sample A / B is plastic," included in the user's request, in the generation of search conditions. The generation of search conditions by the search condition generation unit 37a may be performed using a predetermined rule-based algorithm.

[0080] In step S27, the explanatory information creation unit 38 searches for papers in the RAG database 6a based on the search conditions and performs processing to obtain supporting information. The explanatory information creation unit 38 sends a prompt to the natural language model 5 that includes an instruction to refer to the paper information registered in the RAG database 6a based on the generated search conditions and to obtain supporting information that matches the search conditions. The supporting information is information such as the content of excerpts from papers that match the search conditions and the source (e.g., author, year, journal name, page name, paper title). The explanatory information creation unit 38 may also set the prompt and algorithm to block transmission to the natural language model 5 if spectral data is included in the search conditions and to perform a check instructing the creation of a search query. Publicly available information on the internet may also be obtained as reference information.

[0081] In step S28, the explanatory information creation unit 38 performs a process to cause the natural language model 5 to create explanatory information that explains the processing results of the processing execution unit 34 based on the evidence information. The explanatory information creation unit 38 obtains explanatory information by, for example, sending a prompt to the natural language model 5 that includes an instruction to generate explanatory information that explains the analysis results within the scope of the evidence information, citing the sources. At the end of each sentence and paragraph of the explanatory information generated by the natural language model 5, the source on which it was based (paper title, chapter / section, figure / table number, etc.) is always described. The explanatory information creation unit 38 may also cause the natural language model 5 to check whether the peak wavenumber is within the acceptable range of the search query, and whether each claim has at least one citation.

[0082] In step S29, the presentation processing unit 35 visually or numerically displays the analysis results, along with explanatory information, to the user. The explanatory information includes source information such as excerpts from the relevant section of the paper, author, year, journal name, page number, and paper title.

[0083] As described above, the analysis apparatus 1b of the third embodiment further comprises an extraction processing unit 36a that extracts feature information indicating waveform characteristics from measurement data, a search condition generation unit 37a that generates search conditions based on the feature information, and an explanatory information creation unit 38 that obtains justification information based on the search conditions from a related information registration unit in which a plurality of related pieces of information relating to a predetermined spectroscopic method are registered for the natural language model 5, and creates explanatory information that explains the analysis results based on the said justification information, and a presentation processing unit 35 presents the explanatory information to the user along with the analysis results.

[0084] This allows the system to present primary sources such as research papers with citations along with the analysis results, and to provide users with explanatory information that explains the background of the analysis process and conclusions along with the analysis process. Since the analysis results are explained with supporting evidence, the burden of specialized knowledge and additional work required of the user is reduced. Furthermore, supporting information can be obtained using natural language search criteria without having to send measurement data or the analysis data obtained from the measurement data to the natural language model 5. Therefore, it is possible to avoid interpreting non-natural language data such as measurement data and analysis data, generate explanatory information through the natural language model 5's strength in interpreting natural language, and reduce the communication costs of measurement data and analysis data.

[0085] In the third embodiment, similar to the first embodiment, the processing module is selected solely by the finely tuned natural language model 5, but the configuration is not limited to this. For example, the configuration of the first embodiment may be modified by adding an extraction processing unit 36a, a search condition generation unit 37a, and an explanatory information creation unit 38, so that explanatory information can be presented to the user along with the analysis results even in the RAG selection method. In this case, for example, two types of RAG databases 6 and 6a would be used.

[0086] Furthermore, in the first to third embodiments, an AI agent method that performs processing autonomously may be adopted. In the AI ​​agent method, the natural language model 5 can sequentially determine "which processing module to select next" by comprehensively referring to the user's requests, past analysis history, and pre-set rules. This determination may be a semi-autonomous operation based on the interpretation of the user's instructions, or it may be an operation by a more advanced fully autonomous agent-type AI. In either case, it is preferable to incorporate guardrails such as mandatory citations and condition verification to ensure the reliability of the analysis. Guardrails include, for example, a rule to perform smoothing if the S / N ratio is not above a certain level in the case of peak extraction, a rule to block the search and regenerate the search conditions if spectral data is included in the search conditions, and a rule to delete the explanatory text or instruct a re-search if the explanatory information does not include supporting information such as excerpts from papers or sources. Within the constraints of these guardrails, re-searches and retries can be autonomously performed as needed.

[0087] Furthermore, in the above-described embodiment, the processing module may be configured as a node-type visual programming environment, with each processing module represented as a node, and dynamically connectable and combinable by the user or the natural language model 5. In this case, the analysis device 1 may provide templates for basic processing nodes and also have an interface that allows the user to develop and add custom nodes containing their own processing logic. Each node has an input port and an output port, and processing is executed as a data flow.

[0088] Furthermore, the analysis device 1 may be equipped with functions to link with external scientific and technological services and databases. For example, it may be configured to connect to protein structure prediction services (such as AlphaFold), scientific paper databases (such as PubMed and Elsevier), material property databases, compound databases, etc., via API or MCP (Model Context Protocol) to automatically acquire reference information necessary for analyzing measurement data. Alternatively, the analysis device 1 can be configured to directly connect to hardware such as measuring instruments and manufacturing equipment via MCP for Machine or MCP for Sensor to acquire and analyze data in real time.

[0089] Furthermore, the analysis device 1 may be equipped with an autonomous orchestration function by an AI agent. This AI agent has an inference model (Reasoning Model) and a memory line function, and automatically constructs and executes the processing flow from the user's request to the final analysis result. The AI ​​agent dynamically determines the processing module to be executed next at each stage of processing, and autonomously performs actions such as calling external services, searching databases, and executing simulations as needed. The inference model may be implemented separately from or integrated with the natural language model 5, and may be configured to be selectable from other AI services (OpenAI, Gemini, etc.).

[0090] The analysis device 1 may also have a function to build and utilize knowledge bases specific to the organization or project. Users can register their company's past measurement data, analysis results, know-how, internal documents, etc., as a dedicated customer knowledge base, and the natural language model 5 and AI agent can refer to this information to perform analysis. This allows the system to learn organization-specific analysis patterns and judgment criteria, and provide more accurate analysis results.

[0091] As the execution environment for the processing module, the analysis device 1 may utilize distributed processing and cloud computing functions. When large-scale data processing or complex simulations are required, processing can be distributed across multiple servers or cloud instances for parallel execution, and the results can be integrated. High-speed processing using dedicated computing devices such as GPUs and TPUs, and optimization calculations in cooperation with quantum computers are also possible. The distribution of processing may be automatically determined and controlled by an AI agent.

[0092] The analysis device 1 may include design pipeline functionality and simulation pipeline functionality. The design pipeline functionality automatically generates design proposals for new materials and compounds based on the analysis results of measurement data. The simulation pipeline functionality performs simulations such as property prediction and reaction prediction on the proposed designs, predicting results before experimentation. These pipeline functions are implemented as processing modules and can be used in combination with other processing modules.

[0093] Furthermore, the analysis device 1 may also have a function to determine and convert the output format of the analysis results. Specifically, the analysis device 1 reads user information (job title, field of expertise, past usage history, etc.) from the storage unit 18 and determines the output format based on pre-set determination rules. For example, the analysis device 1 may perform processes such as converting numerical data into bar graphs, line graphs, scatter plots, etc. using a graph plotting library, converting tabular data into HTML or CSV format, integrating multiple analysis results to generate a report file in presentation or PDF format, and embedding analysis results into a paper template of a predetermined format. In addition, the analysis device 1 may perform processes such as periodically executing analysis processing at set times to generate a report file and sending it to a registered email address, generating a warning message and notifying the user when the measurement data exceeds a pre-set threshold, and displaying the analysis results in real time on a dashboard in a browser using WebSocket or Server-Sent Events.

[0094] As a service delivery model for analysis device 1, multiple models may be offered depending on the scale of use and functionality. For example, a tiered service level can be set, such as a basic model that allows the use of basic analysis functions and a cloud AI agent, a professional model that includes large-scale data processing and automatic report generation functions, and a premium model that includes learning support and dedicated agent development support. Each model may be priced on a project basis, per user, or per company-wide deployment basis and may be provided in a SaaS (Software as a Service) format.

[0095] Although one embodiment of the present invention has been described above, the present invention is not limited to the embodiments described above, and any modifications, improvements, etc. that can achieve the objectives of the present invention are included in the present invention.

[0096] Furthermore, the series of processes described above can be executed by hardware or by software. In other words, the functional configuration described above is merely illustrative and not particularly limiting. That is, it is sufficient that the analysis device 1 is equipped with a function that can execute the series of processes described above as a whole, and the type of functional block used to realize this function is not particularly limited to the example above. Also, the location of the functional block is not particularly limited and can be arbitrary. For example, the functional block of the analysis device 1 may be transferred to another device, etc. Conversely, the functional block of another device may be transferred to a server, etc. Also, a single functional block may be composed of hardware alone, software alone, or a combination of both.

[0097] When a series of processes are executed by software, the programs that make up that software are installed on a computer or other device from a network or storage medium. The computer may be a computer built into dedicated hardware. Alternatively, the computer may be a computer capable of performing various functions by installing various programs, such as a server, a general-purpose smartphone, or a personal computer.

[0098] Such recording media containing programs may consist not only of removable media (not shown) distributed separately from the main device to provide the programs, but also of recording media provided pre-installed in the main device. Since programs can be distributed via a network, the recording media may be installed on or accessible from a computer connected to or capable of connecting to a network.

[0099] In this specification, the step of describing a program to be recorded on a recording medium includes not only processes that are performed chronologically in that order, but also processes that are not necessarily performed chronologically, but are executed in parallel or individually. Furthermore, in this specification, the term "system" refers to an overall system composed of multiple devices, means, etc.

[0100] 1, 1a, 1b Analysis device 5 Natural language model 6, 6a RAG database 31 Measurement data acquisition unit 32 User request reception unit 33 Module selection unit 34 Processing execution unit 35 Presentation processing unit 36, 36a Extraction processing unit 37, 37a Search condition generation unit 38 Explanation information creation unit 100 Analysis system

Claims

1. An analysis apparatus comprising: a measurement data acquisition unit that acquires measurement data including the spectrum of a substance to be analyzed measured by a predetermined spectroscopic method; a user request acceptance unit that accepts user requests regarding the analysis of the measurement data in natural language; a module selection unit that performs a selection process to cause a natural language model trained to interpret the natural language to select a processing module corresponding to the user's request from among a plurality of processing modules that process the measurement data; a processing execution unit that causes the processing module selected by the selection process to execute processing; and a presentation processing unit that presents the analysis results based on the processing of the processing module to the user.

2. The analysis apparatus according to claim 1, wherein the processing module includes at least one of the processing of analyzing the spectrum contained in the measurement data and processing related to displaying the waveform of the spectrum.

3. The analysis apparatus according to claim 1 or 2, further comprising: an extraction processing unit for extracting feature information indicating waveform characteristics from the measurement data; and a search condition generation unit for generating search conditions based on the feature information, wherein the module selection unit performs processing to cause the natural language model to select the processing module based on the user's request, the search conditions, and a rule base which pre-defines rules corresponding to the characteristics of the measurement data.

4. The analysis apparatus according to claim 1 or 2, further comprising: an extraction processing unit that extracts feature information indicating the characteristics of a waveform from the measurement data; a search condition generation unit that generates search conditions based on the feature information; and an explanatory information creation unit that obtains justification information based on the search conditions from a related information registration unit in which a plurality of related pieces of information relating to a predetermined spectroscopic method are registered for the natural language model, and causes the creation of explanatory information that explains the analysis results based on the justification information, wherein the presentation processing unit presents the explanatory information to the user along with the analysis results.

5. An analysis method performed by an information processing device, comprising: a measurement data acquisition step of acquiring measurement data including the spectrum of a substance to be analyzed measured by a predetermined spectroscopic method; a user request acceptance step of receiving a user's request regarding the analysis of the measurement data in natural language; a module selection step of causing a natural language model trained to interpret the natural language to select a processing module corresponding to the user's request from among a plurality of processing modules that perform processing on the measurement data; a processing execution step of causing the processing module selected by the selection process to perform processing; and a presentation processing step of presenting the analysis results based on the processing of the processing module to the user.

6. A program for causing a computer to execute the following steps: a measurement data acquisition step of acquiring measurement data including the spectrum of a substance to be analyzed measured by a predetermined spectroscopic method; a user request acceptance step of receiving the user's requests regarding the analysis of the measurement data in natural language; a module selection step of causing a natural language model that has been trained to interpret the natural language to select a processing module from among several types of processing modules that process the measurement data and that corresponds to the user's requests; a processing execution step of causing the processing module selected by the selection process to execute processing; and a presentation processing step of presenting the analysis results based on the processing of the processing module to the user.

Images