Yield prediction method, yield prediction apparatus, yield prediction program, and compound synthesis method

Abstract

To provide a method and the like for efficiently and accurately predicting a yield of a product, and a compound synthesis method using the same.SOLUTION: A yield prediction method includes: a learning model construction step that constructs a learning model trained on a residual spectrum being a combined spectrum residual obtained by excluding portions of partially-overlapping spectral shapes in the combined spectrum obtained by linearly combining, at arbitrary ratios, spectra of respective raw material compounds and a spectrum of a synthesis product in a case where spectral shapes of the respective raw material compounds in all the raw material compounds used in a synthesis reaction and the spectral shape of the synthesis product in the synthesis reaction partially overlap; and a yield prediction step that predicts a yield of a synthesis resultant product by comparing a spectrum of the synthesis resultant product obtained from the synthesis reaction of the respective raw material compounds with the residual spectrum in the learning model that has performed machine learning.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present invention relates to a yield prediction method, a yield prediction device, a yield prediction program, and a compound synthesis method. [Background technology]

[0002] In many industrial sectors, machines are used to produce products. In the field of organic synthesis, delicate and time-consuming operations are often required, making the use of machines suitable for both qualitative and quantitative purposes. In addition to those in industrial production, synthetic chemists working in research and development could also benefit greatly from the use of machines.

[0003] For example, Perera, Richardson, Sack et al. demonstrated 1,500 reactions per day using an automated flow reactor (Non-Patent Document 1). In fact, organic synthesis itself can be accelerated by this automated system, but only a portion of it can be accelerated, and no system has been developed that can predict product yields, etc.

[0004] Therefore, a method for efficiently and accurately predicting the yield of a product and a method for synthesizing a compound using the same have not yet been provided, and there is a strong demand for their prompt provision. [Prior art documents] [Non-patent literature]

[0005] [Non-Patent Document 1] Science 2018 359:429-434 Summary of the Invention [Problem to be solved by the invention]

[0006] The present invention aims to solve the above-mentioned problems in the prior art and to achieve the following object: That is, the present invention aims to provide a method for efficiently and accurately predicting the yield of a product, and a method for synthesizing a compound using the same. [Means for solving the problem]

[0007] As a result of extensive research conducted by the present inventors to achieve the above object, it has been discovered that it is possible to provide a method for efficiently and accurately predicting the yield of a product, and a method for synthesizing a compound using the same.

[0008] The present invention is based on the above findings by the present inventors, and the means for solving the above problems are as follows: <1> In the case where the spectral shape of each raw material compound in all raw material compounds used in a synthesis reaction partially overlaps with the spectral shape of the synthesis product of the synthesis reaction, a learning model construction step of constructing a learning model by machine learning a residual spectrum, which is a remainder of the combined spectrum obtained by linearly combining the spectra of the raw material compounds and the spectrum of the synthetic product in an arbitrary ratio, excluding the overlapping portions of the spectral shapes of the combined spectrum; a yield prediction step of predicting the yield of the synthesis resultant product by comparing the spectrum of the synthesis resultant product obtained by the synthesis reaction of each of the raw material compounds with the remaining spectrum in the machine-learned learning model; The yield prediction method is characterized by comprising: In the learning model construction step, a learning model is constructed by machine learning a residual spectrum, which is a remainder of the combined spectrum obtained by linearly combining the spectra of the raw material compounds and the spectrum of the synthetic product in any ratio, excluding any overlapping spectral shapes. In the yield prediction step, the spectrum of a synthesis result obtained by a synthetic reaction of the raw material compounds is compared with the residual spectrum in the machine-learned learning model to predict the yield of the synthesis result. As a result, the yield of the product can be predicted efficiently and accurately. <2> In the case where the spectral shape of each raw material compound in all raw material compounds used in a synthesis reaction partially overlaps with the spectral shape of the synthesis product of the synthesis reaction, a learning model construction step of constructing a learning model by machine learning a differential residual spectrum obtained by differentiating a combined spectrum obtained by linearly combining the spectra of the raw material compounds and the spectrum of the synthetic product in an arbitrary ratio, excluding a partially overlapping portion of the spectral shape of the combined spectrum; and a yield prediction step of predicting the yield of the synthesis resultant product by comparing the spectrum of the synthesis resultant product obtained by the synthesis reaction of each of the raw material compounds with the differential residual spectrum in the machine-learned learning model; The yield prediction method is characterized by comprising: In the learning model construction step, a learning model is constructed by machine learning a differential remainder spectrum obtained by linearly combining the spectra of the raw material compounds and the spectrum of the synthetic product in an arbitrary ratio, and differentiating the combined spectrum remainder, excluding any overlapping spectral shapes in the combined spectrum. In the yield prediction step, the spectrum of a synthesis result obtained by a synthetic reaction of the raw material compounds is compared with the differential remainder spectrum in the machine-learned learning model to predict the yield of the synthesis result. As a result, product yield prediction can be performed more efficiently and accurately. <3> the spectrum of each raw material compound, the spectrum of the synthesis product, and the spectrum of the synthesis resultant are spectra obtained by optical analysis; <1> or <2> 2. The method for predicting yield according to claim 1. The aforementioned <3> In the yield prediction method described in , the spectrum of each raw material compound, the spectrum of the synthesis product, and the spectrum of the synthesis resultant are obtained by optical analysis, which allows for efficient and accurate prediction of the yield of the product. <4> The spectrum obtained by the optical analysis is at least one selected from an infrared absorption (IR) spectrum, a nuclear magnetic resonance (NMR) spectrum, a near-infrared absorption (NIR) spectrum, and a Raman spectroscopy spectrum. <3> 2. The method for predicting yield according to claim 1. The aforementioned <4> In the yield prediction method described in 2, the spectrum obtained by the optical analysis is at least one selected from an infrared absorption (IR) spectrum, a nuclear magnetic resonance (NMR) spectrum, a near-infrared absorption (NIR) spectrum, and a Raman spectroscopy spectrum, thereby enabling efficient and accurate prediction of the yield of the product. <5> The optical analysis is carried out by an optical analyzer that continuously optically analyzes the synthesis products obtained by the synthesis reaction of the raw material compounds. <3> or <4> 2. The method for predicting yield according to claim 1. The aforementioned <5> In the yield prediction method described in 2., the optical analysis is performed by an optical analyzer that continuously optically analyzes the synthesis products obtained by the synthesis reaction of the raw material compounds. As a result, the yield prediction method can be suitably performed even in a continuous synthesis system using a microreactor or the like. <6> In the case where the spectral shape of each raw material compound in all raw material compounds used in a synthesis reaction partially overlaps with the spectral shape of the synthesis product of the synthesis reaction, a learning model construction step of constructing a learning model by machine learning a residual spectrum, which is a remainder of the combined spectrum obtained by linearly combining the spectra of the raw material compounds and the spectrum of the synthetic product in an arbitrary ratio, excluding the overlapping portions of the spectral shapes of the combined spectrum; a yield prediction step of predicting the yield of the synthesis resultant product by comparing the spectrum of the synthesis resultant product obtained by the synthesis reaction of each of the raw material compounds with the remaining spectrum in the machine-learned learning model; an optimization step of changing the conditions of the synthesis reaction to perform a new synthesis reaction if the predicted yield is less than a predetermined yield value; The present invention relates to a method for synthesizing a compound, comprising: In the learning model construction step, a learning model is constructed by machine learning a residual spectrum, which is the remainder of the combined spectrum obtained by linearly combining the spectra of the raw material compounds and the spectra of the synthetic product in any ratio, excluding the overlapping spectral shape portions of the combined spectrum. In the yield prediction step, the spectrum of the synthesis result obtained by the synthetic reaction of the raw material compounds is compared with the residual spectrum in the machine-learned learning model to predict the yield of the synthesis result. In the optimization step, if the predicted yield is less than a predetermined yield value, the conditions of the synthesis reaction are changed and a new synthesis reaction is performed. As a result, the yield of the product can be predicted efficiently and accurately, and compounds can be synthesized efficiently. <7> In the case where the spectral shape of each raw material compound in all raw material compounds used in a synthesis reaction partially overlaps with the spectral shape of the synthesis product of the synthesis reaction, a learning model construction step of constructing a learning model by machine learning a differential residual spectrum obtained by differentiating a combined spectrum obtained by linearly combining the spectra of the raw material compounds and the spectrum of the synthetic product in an arbitrary ratio, excluding a partially overlapping portion of the spectral shape of the combined spectrum; and a yield prediction step of predicting the yield of the synthesis resultant product by comparing the spectrum of the synthesis resultant product obtained by the synthesis reaction of each of the raw material compounds with the differential residual spectrum in the machine-learned learning model; an optimization step of changing the conditions of the synthesis reaction and performing a new synthesis reaction if the predicted yield is less than a predetermined yield value; The present invention relates to a method for synthesizing a compound, comprising: In the learning model construction step, a learning model is constructed by machine learning a differential residual spectrum obtained by linearly combining the spectra of the raw compounds and the spectra of the synthetic product at any ratio, and differentiating the residual of the combined spectrum, excluding any overlapping spectral shapes. In the yield prediction step, the spectrum of a synthesis result obtained by a synthetic reaction of the raw compounds is compared with the residual spectrum in the machine-learned learning model to predict the yield of the synthesis result. In the optimization step, if the predicted yield is less than a predetermined yield value, the conditions of the synthesis reaction are changed and a new synthesis reaction is performed. As a result, product yields can be predicted efficiently and accurately, and compounds can be synthesized efficiently. <8> In the optimization step, if the predicted yield is less than a predetermined yield value, a spectrum of a new synthesis resultant obtained by performing a new synthesis reaction under different synthesis reaction conditions is compared with the remainder spectrum or the differential remainder spectrum to predict the new yield of the synthesis resultant, and the synthesis reaction is repeated until the newly predicted yield of the synthesis resultant becomes equal to or greater than the predetermined yield value. <6> or <7> This is a method for synthesizing the compound described in The aforementioned <8> In the compound synthesis method described in , if the predicted yield is less than a predetermined yield value in the optimization step, the spectrum of a new synthesis resultant obtained by performing a new synthesis reaction under different synthesis reaction conditions is compared with the remainder spectrum or the differential remainder spectrum to predict the yield of the new synthesis resultant, and the synthesis reaction is repeated until the newly predicted yield of the synthesis resultant becomes equal to or greater than the predetermined yield value. As a result, conditions for obtaining a target compound in high yield can be determined automatically in a short time. <9> In the case where the spectral shape of each raw material compound in all raw material compounds used in a synthesis reaction partially overlaps with the spectral shape of the synthesis product of the synthesis reaction, a learning model construction means for constructing a learning model by machine learning a residual spectrum, which is a remainder of the combined spectrum obtained by linearly combining the spectra of the raw material compounds and the spectrum of the synthetic product in an arbitrary ratio, excluding the overlapping spectral shape portion of the combined spectrum; and a yield prediction means for predicting the yield of the synthesis resultant product by comparing the spectrum of the synthesis resultant product obtained by the synthesis reaction of each of the raw material compounds with the remaining spectrum in the machine-learned learning model; The yield prediction device is characterized by comprising: The learning model construction means constructs a learning model by machine learning a residual spectrum, which is a remainder of the combined spectrum obtained by linearly combining the spectra of the raw material compounds and the spectrum of the synthetic product in an arbitrary ratio, excluding any overlapping spectral shapes. The yield prediction means predicts the yield of the synthetic result by comparing the spectrum of the synthetic result obtained by the synthetic reaction of the raw material compounds with the residual spectrum in the machine-learned learning model. As a result, the yield of the product can be predicted efficiently and accurately. <10> In the case where the spectral shape of each raw material compound in all raw material compounds used in a synthesis reaction partially overlaps with the spectral shape of the synthesis product of the synthesis reaction, a learning model constructing means for constructing a learning model by machine learning a differential residual spectrum obtained by differentiating a residual spectrum obtained by excluding a partially overlapping portion of the spectral shape in a combined spectrum obtained by linearly combining the spectra of the raw material compounds and the spectrum of the synthetic product in an arbitrary ratio; and a yield prediction means for predicting the yield of the synthesis resultant by comparing the spectrum of the synthesis resultant obtained by the synthesis reaction of each of the raw material compounds with the differential residual spectrum in the machine-learned learning model; The yield prediction device is characterized by comprising: The learning model construction means constructs a learning model by machine learning a differential residual spectrum obtained by linearly combining the spectra of the raw material compounds and the spectrum of the synthetic product in an arbitrary ratio, and differentiating the residual of the combined spectrum, excluding any overlapping spectral shapes. The yield prediction means predicts the yield of the synthetic result by comparing the spectrum of the synthetic result obtained by the synthetic reaction of the raw material compounds with the residual spectrum in the machine-learned learning model. As a result, the yield of the product can be predicted more efficiently and accurately. <11> In the case where the spectral shape of each raw material compound in all raw material compounds used in a synthesis reaction partially overlaps with the spectral shape of the synthesis product of the synthesis reaction, The learning model is characterized by being obtained by machine learning a residual spectrum, which is the remainder of the combined spectrum obtained by linearly combining the spectra of the raw material compounds and the spectrum of the synthetic product in any ratio, excluding the partially overlapping spectral shapes of the combined spectrum. The learning model is obtained by machine learning a residual spectrum, which is a remainder of the combined spectrum obtained by linearly combining the spectra of the raw material compounds and the spectrum of the synthetic product in an arbitrary ratio, excluding the overlapping spectral shapes of the combined spectrum, and as a result, an accurate learning model can be obtained. <12> In the case where the spectral shape of each raw material compound in all raw material compounds used in a synthesis reaction partially overlaps with the spectral shape of the synthesis product of the synthesis reaction, The learning model is characterized by being machine-learned using a differential remainder spectrum obtained by differentiating a remainder of the combined spectrum obtained by linearly combining the spectra of the raw material compounds and the spectrum of the synthetic product in any ratio, excluding any overlapping spectral shapes. The learning model is obtained by machine learning a differential residual spectrum obtained by linearly combining the spectra of the raw material compounds and the spectrum of the synthetic product in an arbitrary ratio, and differentiating the residual of the combined spectrum, excluding the overlapping spectral shape portion of the combined spectrum. As a result, an accurate learning model can be obtained. <13> In the case where the spectral shape of each raw material compound in all raw material compounds used in a synthesis reaction partially overlaps with the spectral shape of the synthesis product of the synthesis reaction, a process of constructing a learning model by machine learning a residual spectrum, which is a remainder of the combined spectrum obtained by linearly combining the spectra of the raw material compounds and the spectrum of the synthetic product in an arbitrary ratio, excluding a partially overlapping portion of the spectral shape of the combined spectrum; a process of predicting a yield of the synthesis resultant product by comparing a spectrum of the synthesis resultant product obtained by the synthesis reaction of each of the raw material compounds with the remaining spectrum in the machine-learned learning model; The present invention is a yield prediction program that causes a computer to perform the above. In the yield prediction program, a process of constructing a learning model by machine learning a residual spectrum, which is a remainder of a combined spectrum obtained by linearly combining the spectra of the raw material compounds and the spectrum of the synthetic product in an arbitrary ratio, excluding the overlapping portions of the spectral shapes of the combined spectrum, is performed, and a process of predicting the yield of the synthetic resultant product by comparing the spectrum of the synthetic resultant product obtained by the synthetic reaction of the raw material compounds with the residual spectrum in the machine-learned learning model is performed.As a result, the yield of the product can be predicted efficiently and accurately. <14> In the case where the spectral shape of each raw material compound in all raw material compounds used in a synthesis reaction partially overlaps with the spectral shape of the synthesis product of the synthesis reaction, a process of constructing a learning model by machine learning a differential residual spectrum obtained by differentiating a residual of the combined spectrum obtained by linearly combining the spectra of the raw material compounds and the spectrum of the synthetic product in an arbitrary ratio, excluding a portion of the spectral shape that overlaps the combined spectrum; and a process of predicting a yield of the synthesis resultant product by comparing a spectrum of the synthesis resultant product obtained by the synthesis reaction of each of the raw material compounds with the differential residual spectrum in the machine-learned learning model; The present invention is a yield prediction program that causes a computer to perform the above. In the yield prediction program, a process of constructing a learning model by machine learning a differential residual spectrum obtained by linearly combining the spectra of the raw material compounds and the spectrum of the synthetic product in an arbitrary ratio, and differentiating the residual of the combined spectrum, excluding any overlapping spectral shapes, is performed, and a process of predicting the yield of the synthetic resultant product by comparing the spectrum of the synthetic resultant product obtained by the synthetic reaction of the raw material compounds with the residual spectrum in the machine-learned learning model is performed.As a result, the yield of the product can be predicted more efficiently and accurately. [Effects of the Invention]

[0009] According to the present invention, it is possible to provide a method for efficiently and accurately predicting the yield of a product, and a method for synthesizing a compound using the same. [Brief explanation of the drawings]

[0010] [Figure 1] FIG. 1 is a diagram showing an example of the hardware configuration of a yield prediction device according to the present invention. [Figure 2] FIG. 2 is a diagram showing an example of the functional configuration of the yield prediction device of the present invention. [Figure 3] FIG. 3 shows the FTIR spectra of PhBpin (1), IPhCN (2), and PhPhCN (3). [Figure 4]FIG. 4 is a schematic diagram of the microreactor system used in the examples. [Figure 5] FIG. 5 shows the correlation between actual and predicted yield values. [Figure 6] FIG. 6 is a diagram showing an example of a mimic spectrum. [Figure 7] FIG. 7 is a diagram showing an example of a mimic spectrum. [Figure 8] FIG. 8 is a diagram illustrating an example of a flowchart for constructing a learning model. [Figure 9] FIG. 9 is a diagram illustrating an example of the evaluation result of the model. [Figure 10] FIG. 10 is a diagram showing an example of a flowchart for predicting yield. [Figure 11] FIG. 11 is a schematic diagram of an automatic optimization control system. [Figure 12] FIG. 12 shows an example flowchart of the automated optimization process when using IR. [Figure 13] FIG. 13 shows an example flowchart of the automated optimization process when using NMR. [Figure 14] Figure 14 is a schematic diagram of a flow system equipped with an in-line IR. [Figure 15] FIG. 15 shows the correlation between actual and predicted yield values. [Figure 16] Figure 16 is a schematic diagram of a PLC to which instruments such as a thermometer, a flow meter, and a pressure meter are connected. [Figure 17] FIG. 17 is a diagram showing a log of the PLC. [Figure 18] Figure 18 is a schematic diagram of a flow system equipped with in-line NMR. DETAILED DESCRIPTION OF THE INVENTION

[0011] (Yield prediction method and yield prediction device) The yield prediction method of the present invention includes a learning model construction step and a yield prediction step, and may further include other steps. The yield prediction method of the present invention is a method for predicting a yield when the spectral shape of each raw material compound among all raw material compounds used in a synthesis reaction and the spectral shape of the synthesis product of the synthesis reaction partially overlap. The yield prediction device of the present invention includes a learning model construction means and a yield prediction means, and may further include other means. The yield prediction device of the present invention is a device for predicting yield in the case where the spectral shape of each raw material compound among all raw material compounds used in a synthesis reaction and the spectral shape of the synthesis product of the synthesis reaction partially overlap.

[0012] The yield prediction method of the present invention can be suitably implemented by the yield prediction device of the present invention, and the learning model construction process can be performed by the learning model construction means, the yield prediction process can be performed by the pre-yield prediction process means, and the other processes can be performed by the other means. The yield prediction device of the present invention can be suitably operated by the yield prediction program of the present invention. Hereinafter, the yield prediction device of the present invention will also be described through the description of the yield prediction method of the present invention.

[0013] The number of all raw material compounds is not particularly limited and can be appropriately selected depending on the purpose, but is preferably 2 or more.

[0014] The synthesis reaction is not particularly limited and can be appropriately selected depending on the purpose, but is preferably an organic synthesis reaction. The organic synthesis reaction is not particularly limited and can be appropriately selected depending on the purpose. Examples of the reaction include coupling reaction, lithiation reaction, oxidation-reduction reaction, carbon-carbon bond forming reaction, esterification / amidation reaction, and the like.

[0015] The coupling reaction is not particularly limited and can be appropriately selected depending on the purpose, and examples thereof include the Mizoroki-Heck reaction, Negishi coupling reaction, Migita-Kosugi-Still coupling reaction, Sonogashira coupling reaction, Hiyama coupling reaction, Suzuki-Miyaura coupling reaction, Buchwald-Hartwig amination reaction, Kumada-Tamao-Colew coupling reaction, etc. Among these, the Suzuki-Miyaura coupling reaction is preferred.

[0016] The spectrum of each of the raw material compounds is not particularly limited and can be appropriately selected depending on the purpose, but a spectrum obtained by optical analysis is preferred. The spectrum obtained by the optical analysis is not particularly limited and can be appropriately selected depending on the purpose. Examples include infrared absorption (IR) spectrum, nuclear magnetic resonance (NMR) spectrum, near-infrared absorption (NIR) spectrum, and Raman spectroscopy spectrum. Among these, infrared absorption (IR) spectrum or nuclear magnetic resonance (NMR) spectrum is preferred. The spectra obtained by the optical analysis may be used singly or in combination of two or more.

[0017] The spectrum of the synthesis product is not particularly limited and can be appropriately selected depending on the purpose, but a spectrum obtained by optical analysis is preferred. The spectrum obtained by the optical analysis is not particularly limited and can be appropriately selected depending on the purpose, and examples thereof include infrared absorption (IR) spectrum, nuclear magnetic resonance (NMR) spectrum, near-infrared absorption (NIR) spectrum, Raman spectroscopy spectrum, etc. Among these, infrared absorption (IR) spectrum or nuclear magnetic resonance (NMR) spectrum is preferred.

[0018] The phrase "the spectral shape of each of the raw material compounds among all the raw material compounds used in the synthesis reaction partially overlaps with the spectral shape of the synthesis product of the synthesis reaction" means that the spectral shapes of each of the raw material compounds among all the raw material compounds and the spectral shape of the synthesis product of the synthesis reaction overlap in at least a partial region. When there are a plurality of raw material compounds, the spectral shape of at least one raw material compound and the spectral shape of the synthesis product of the synthesis reaction overlap in at least a partial region.

[0019] The at least a portion of the region is not particularly limited and can be appropriately selected depending on the purpose, but is preferably 1% or more, more preferably 5% or more, even more preferably 10% or more, particularly preferably 25% or more, and most preferably 50% or more of the entire region of the spectrum.

[0020] The overlapping of the spectral shapes means that the spectral intensities are similar. The similarity of the spectral intensity is not particularly limited and can be selected appropriately depending on the purpose, but the spectral intensity relative to the comparison target (overlapping spectrum) is preferably 50% to 200% inclusive, more preferably 70% to 160% inclusive, even more preferably 90% to 120% inclusive, particularly preferably 95% to 110% inclusive, and most preferably 99% to 105% inclusive. Among these, those in which the difference cannot be seen with the naked eye are preferred.

[0021] A specific example of a case where the spectral shape of each of the raw material compounds used in the synthesis reaction partially overlaps with the spectral shape of the synthesis product of the synthesis reaction is, for example, a case where the raw material compounds are 4,4,5,5-tetramethyl-2-phenyl-1,3,2-dioxaborolane (PhBpin) and 4-iodobenzonitrile (IPhCN), and the synthesis product is 4-phenylbenzonitrile (PhPhCN).

[0022] The reaction system for the synthesis reaction is not particularly limited and can be appropriately selected depending on the purpose. It may be a batch synthesis system or a continuous synthesis system, with a continuous synthesis system being preferred. The continuous synthesis system is not particularly limited and can be appropriately selected depending on the purpose. For example, a reaction system using a microreactor (also called a flow microreactor) can be mentioned.

[0023] -Microreactor- The microreactor is not particularly limited and can be appropriately selected depending on the purpose. For example, a microreactor equipped with a mixing means and a flow channel, and further equipped with other means as necessary, can be mentioned. The mixing means and the flow passage may be of an integral type or of a separate type.

[0024] The mixing means is a means capable of mixing two or more types of liquids. The flow passage is a pipe through which a liquid can flow, and is connected to at least one of the mixing means.

[0025] By using the flow microreactor, it is possible to shorten the residence time from the production to the next reaction for a compound with low stability, thereby suppressing side reactions. Furthermore, the flow microreactor has excellent cooling efficiency, and therefore can suppress side reactions caused by heat generated in an exothermic reaction.

[0026] --Integrated flow microreactor-- The mixing means and the flow passage of the integrated flow microreactor may be a substrate-type micromixer.

[0027] The substrate-type micromixer is made of a substrate having passages formed inside or on the surface thereof, and is sometimes called a microchannel. The substrate-type micromixer is not particularly limited and can be appropriately selected depending on the purpose. Examples include the mixer having fine flow channels for mixing described in International Publication No. 96 / 30113; and the mixer described in the literature "Microreactors," Chapter 3, by W. Ehrfeld, V. Hessel, and H. Lowe, published by Wiley-VCH.

[0028] In the substrate-type micromixer, the mixing means and the flow path are configured by minute flow paths that can mix a plurality of liquids.

[0029] In addition to the flow paths, the substrate-type micromixer is preferably formed with inlet paths that communicate with the flow paths and introduce a plurality of liquids into the flow paths. That is, it is preferable that the flow paths are branched on the upstream side depending on the number of the inlet paths.

[0030] The number of the inlet channels is not particularly limited and can be appropriately selected depending on the purpose, but it is preferable that the multiple liquids to be mixed are introduced through separate inlet channels and merged in the flow channel for mixing. Alternatively, one liquid may be pre-loaded in the flow channel, and the other liquids may be introduced through the inlet channels.

[0031] --Separate flow microreactor-- The separate flow microreactor comprises a mixing means and a flow passage connected thereto.

[0032] The mixing means is not particularly limited as long as it can mix two or more liquids, and can be appropriately selected depending on the purpose. For example, a pipe joint type micromixer can be used.

[0033] The pipe joint type micromixer has a flow path formed therein and, if necessary, a connecting member that connects the flow path formed therein with the flow passage. The connection method of the connecting member is not particularly limited and can be appropriately selected from known connection methods depending on the purpose, and examples thereof include a threaded type, a union type, a butt welding type, a plug welding type, a socket welding type, a flange type, a bite type, a flare type, and a mechanical type.

[0034] In addition to the flow paths, it is preferable that inlet paths communicating with the flow paths and introducing multiple liquids into the flow paths are formed inside the pipe joint type micromixer. That is, it is preferable that the flow paths are branched upstream depending on the number of the inlet paths. When the number of the inlet paths is two, the pipe joint type micromixer can be, for example, T-shaped or Y-shaped, and when the number of the inlet paths is three, it can be, for example, cross-shaped. Note that it is also possible to configure the micromixer so that one liquid is pre-loaded into the flow path and the other liquids are introduced through the inlet paths.

[0035] The material of the pipe joint type micromixer is not particularly limited and can be appropriately selected depending on requirements such as heat resistance, pressure resistance, solvent resistance, and ease of processing. Examples include stainless steel, titanium, copper, nickel, aluminum, silicon, fluororesins such as Teflon (registered trademark) and PFA (perfluoroalkoxy resin), and TFAA (trifluoroacetamide).

[0036] As the pipe joint type micro mixer, commercially available products can be used, such as the YM-1 type mixer and YM-2 type mixer manufactured by Azbil Corporation; mixing tees and tees (T-shaped connectors) manufactured by Shimadzu GLC Corporation; the Micro High Mixer developed by Toray Engineering; union tees manufactured by Swagelok Corporation; and a T-shaped micro mixer manufactured by Sanko Seiki Kogyo Co., Ltd.

[0037] The method of mixing two or more raw materials in the mixing means is not particularly limited and can be appropriately selected depending on the purpose, and examples include mixing by laminar flow, mixing by turbulent flow, etc. Among these, mixing by laminar flow (static mixing) is preferred in that it allows for more efficient reaction control and heat removal.

[0038] Since the flow paths in the mixing means are minute, the liquids introduced into the mixing means tend to naturally flow in a laminar manner, and are mixed by diffusing in a direction perpendicular to the flow. In mixing by laminar flow, branching points and confluences may be provided in the flow paths to divide the laminar cross section of the flowing liquids, thereby increasing the mixing speed. Furthermore, when turbulent mixing (dynamic mixing) is performed in the flow path of the mixing means, the laminar flow can be changed to turbulent flow by adjusting the flow rate and the shape of the flow path (the three-dimensional shape of the liquid-contacting portion, the shape of the flow path such as bending, the roughness of the wall surface, etc.). Mixing by turbulent flow has the advantages of better mixing efficiency and faster mixing speed than mixing by laminar flow.

[0039] Here, a smaller inner diameter of the flow path in the mixing means can shorten the diffusion distance of molecules, thereby shortening the time required for mixing and improving the mixing efficiency. Furthermore, a smaller inner diameter of the flow path increases the ratio of surface area to volume, making it easier to control the temperature of the liquid, for example, by removing the heat of reaction. On the other hand, if the inner diameter of the flow path is too small, the pressure loss when the liquid flows increases, and a special high-pressure pump is required for the liquid delivery, which may increase the manufacturing cost. Furthermore, the liquid delivery flow rate is limited, which may also limit the structure of the micromixer.

[0040] The lower limit of the average inner diameter of the flow path in the mixing means (average inner diameter of the micromixer) is not particularly limited and can be appropriately selected depending on the purpose. From the viewpoints of enabling faster mixing, more efficient removal of reaction heat, and reducing pressure loss during liquid transfer, the lower limit is preferably 50 μm or more, more preferably 100 μm or more, even more preferably 250 μm or more, even more preferably 500 μm, particularly preferably 1000 μm or more, and most preferably 1300 μm or more.

[0041] The upper limit of the average inner diameter of the flow path in the mixing means (average inner diameter of the micromixer) is not particularly limited and can be appropriately selected depending on the purpose, but is preferably 4 mm or less, more preferably 3 mm or less, and even more preferably 2.5 mm or less.

[0042] If the average inner diameter is less than 50 μm, the pressure loss may increase, and if the average inner diameter is more than 4 mm, the surface area per unit volume becomes small, which may result in difficulty in rapid mixing and removal of reaction heat.

[0043] The cross-sectional area of ​​the flow channel is not particularly limited and can be appropriately selected depending on the purpose. 2 ~16mm 2 is preferable, and 1,000 μm 2 ~4.0mm 2 is more preferable, and 10,000 μm 2 ~2.1mm 2 is more preferably 190,000 μm 2 ~1mm 2 is particularly preferred.

[0044] The cross-sectional shape of the flow channel is not particularly limited and can be appropriately selected depending on the purpose. Examples include a circle, a rectangle, a semicircle, and a triangle.

[0045] The flow path is not particularly limited as long as it is a tube that is connected to at least one of the mixing means and allows a liquid to flow through it, and can be appropriately selected depending on the purpose. The configuration of the flow path, such as its inner diameter, outer diameter, length, material, etc., can be appropriately selected depending on the desired reaction.

[0046] The flow passage is used, for example, when supplying raw materials to the mixing means. The flow path is used, for example, when supplying a reaction product of two or more substances mixed by the mixing means to the next mixing means, while the reaction may continue to occur within the flow path.

[0047] The flow path can be a commercially available product, such as a stainless steel tube manufactured by GL Sciences Inc. (outer diameter 1 / 16 inch (1.58 mm), inner diameter selectable from 250 μm, 500 μm, and 1,000 μm, tube length adjustable by the user).

[0048] The material of the flow path is not particularly limited, and the materials exemplified as the materials of the mixing means can be suitably used.

[0049] --Other means-- The other means is not particularly limited and can be appropriately selected depending on the purpose. Examples thereof include a liquid delivery means and a temperature control means.

[0050] The liquid delivery means is not particularly limited as long as it can supply various raw materials to the flow passage of the flow microreactor, and can be appropriately selected depending on the purpose. For example, a pump can be used.

[0051] The pump is not particularly limited and can be appropriately selected from those that can be used industrially. However, it is preferable to use a pump that does not generate pulsation during liquid transfer, and examples thereof include a plunger pump, a gear pump, a rotary pump, and a diaphragm pump.

[0052] The temperature control means is not particularly limited as long as it can control the temperatures of the mixing means and the flow channel of the flow microreactor, and can be appropriately selected depending on the purpose.

[0053] The optical analysis is not particularly limited and can be appropriately selected depending on the purpose, but is preferably performed using an optical analyzer that continuously optically analyzes the synthesis product obtained by the synthesis reaction of each of the raw material compounds. The optical analyzer for continuously optically analyzing the synthesis products obtained by the synthesis reaction of the raw material compounds is not particularly limited and can be appropriately selected depending on the purpose. For example, an optical analyzer connected to the microreactor can be used.

[0054] <Learning model construction process> The learning model construction step is a step of constructing a learning model by machine learning a residual spectrum, which is the remainder of the combined spectrum obtained by linearly combining the spectra of the raw material compounds and the spectrum of the synthetic product in any ratio, excluding the partially overlapping spectral shape portions of the combined spectrum.

[0055] The combined spectrum is not particularly limited as long as it is obtained by linearly combining the spectra of the raw material compounds and the spectrum of the synthesis product in any ratio, and can be appropriately selected depending on the purpose. When there are two raw material compounds, the spectra of the two raw material compounds and the spectrum of the synthesis product (a total of three spectra) are linearly combined in any proportion to produce a combined spectrum.

[0056] The remainder spectrum is not particularly limited as long as it is the remainder of the combined spectrum excluding the overlapping spectral shape portion in the combined spectrum, and can be appropriately selected depending on the purpose, but a spectrum in the fingerprint region is preferred.

[0057] The remainder spectrum is preferably a differential remainder spectrum obtained by differentiating the remainder of the combined spectrum. The derivative of the combined spectrum remnant is not particularly limited and can be appropriately selected depending on the purpose, but is preferably the derivative of the spectral intensity value of the combined spectrum remnant.

[0058] The number of data to be subjected to the machine learning is not particularly limited and can be appropriately selected depending on the purpose, but is preferably 1,000 or more, more preferably 5,000 or more, and even more preferably 10,000 or more.

[0059] The learning model for performing the machine learning is not particularly limited and can be appropriately selected depending on the purpose, and examples include those using a neural network, a support vector machine, a Bayesian network, Adaboost, a random forest, active learning, etc. Among these, those using a neural network are preferred.

[0060] <Method for building learning models> The learning model construction means is not particularly limited as long as it can construct a learning model by machine learning a residual spectrum, which is a remainder of the combined spectrum obtained by linearly combining the spectra of the raw material compounds and the spectrum of the synthetic product in any ratio, excluding any overlapping spectral shapes in the combined spectrum, and can be appropriately selected depending on the purpose. Examples of the learning model construction means include known microprocessors, computers, and server devices. The learning model construction means is preferably capable of recording (holding) the learning model (database). The learning model construction means may be one unit or a plurality of units.

[0061] <Yield prediction process> The yield prediction step is a step of predicting the yield of the synthesis resultant by comparing the spectrum of the synthesis resultant obtained by the synthesis reaction of each of the raw material compounds with the residual spectrum in the machine-learned learning model.

[0062] The spectrum of the synthesis product is not particularly limited and can be appropriately selected depending on the purpose, but a spectrum obtained by optical analysis is preferred. The spectrum obtained by the optical analysis is not particularly limited and can be appropriately selected depending on the purpose, and examples thereof include infrared absorption (IR) spectrum, nuclear magnetic resonance (NMR) spectrum, near-infrared absorption (NIR) spectrum, Raman spectroscopy spectrum, etc. Among these, infrared absorption (IR) spectrum or nuclear magnetic resonance (NMR) spectrum is preferred.

[0063] <Yield prediction method> The yield prediction means is not particularly limited as long as it can predict the yield of the synthesis resultant product by comparing the spectrum of the synthesis resultant product obtained by the synthesis reaction of each of the raw material compounds with the residual spectrum in the machine-learned learning model, and can be appropriately selected depending on the purpose. Examples of the yield prediction means include known microprocessors, computers, and server devices. The yield prediction means may be one or a plurality of means.

[0064] (Learning model) The learning model of the present invention is a learning model obtained by machine learning a residual spectrum, which is the remainder of the combined spectrum obtained by linearly combining the spectra of the raw material compounds and the spectrum of the synthetic product in any ratio, excluding the partially overlapping spectral shapes of the combined spectrum.

[0065] The learning model of the present invention is a learning model for the case where the spectral shape of each raw material compound among all raw material compounds used in a synthesis reaction and the spectral shape of the synthesis product of the synthesis reaction partially overlap.

[0066] A preferred embodiment of the learning model of the present invention can be the same as a preferred embodiment of the yield prediction method of the present invention, for example.

[0067] (Yield prediction program) The yield prediction program of the present invention causes a computer to perform the following steps: constructing a learning model by machine learning a residual spectrum, which is a remainder of a combined spectrum obtained by linearly combining the spectra of the raw material compounds and the spectrum of the synthetic product in any ratio, excluding any overlapping spectral shapes of the combined spectrum; and predicting the yield of the synthetic product by comparing the spectrum of a synthesis result obtained by a synthetic reaction of the raw material compounds with the residual spectrum in the machine-learned learning model.

[0068] The yield prediction program of the present invention is a program for predicting yield in the case where the spectral shape of each raw material compound among all raw material compounds used in a synthesis reaction and the spectral shape of the synthesis product of the synthesis reaction partially overlap.

[0069] The yield prediction program of the present invention can be, for example, a program that causes a computer to execute the yield prediction method of the present invention. Furthermore, preferred aspects of the yield prediction program of the present invention can be, for example, the same as preferred aspects of the yield prediction method of the present invention.

[0070] The yield prediction program of the present invention can be created using various known programming languages ​​depending on the configuration of the computer system used and the type and version of the operating system.

[0071] The yield prediction program of the present invention may be recorded on a recording medium such as an internal hard disk or an external hard disk, or on a recording medium such as a CD-ROM, a DVD-ROM, an MO disk, or a USB memory. Furthermore, when the yield prediction program of the present invention is recorded on the above-mentioned recording medium, it can be used directly or by installing it on a hard disk via a recording medium reader possessed by a computer system, as needed. The yield prediction program of the present invention may also be recorded in an external storage area (such as another computer) accessible from the computer system via an information and communication network. In this case, the yield prediction program of the present invention recorded in the external storage area can be used directly or by installing it on a hard disk from the external storage area via an information and communication network, as needed. The yield prediction program of the present invention may be divided into programs for each of the processes and recorded on a plurality of recording media.

[0072] <Computer-readable recording medium> A computer-readable recording medium according to the present invention stores the yield prediction program of the present invention. There are no particular limitations on the computer-readable recording medium related to the present invention and it can be selected appropriately depending on the purpose. Examples include an internal hard disk, an external hard disk, a CD-ROM, a DVD-ROM, an MO disk, and a USB memory. Furthermore, the computer-readable recording medium according to the present invention may be a plurality of recording media on which the yield prediction program of the present invention is recorded in a divided form for each of the processes.

[0073] An example of the technology disclosed in the present invention will be described in more detail below using an example of the configuration of the device. FIG. 1 shows an example of the hardware configuration of the yield prediction device of the present invention. In the yield prediction device 100 of the present invention, for example, a control unit 101, a main memory device 102, an auxiliary memory device 103, an I / O interface 104, a communication interface 105, an input device 106, an output device 107, and a display device 108 are connected via a system bus 109.

[0074] The control unit 101 performs calculations (arithmetic operations, comparison operations, etc.), controls the operation of hardware and software, etc. The control unit 101 may be, for example, a CPU (Central Processing Unit), a part of a machine used in the yield prediction method of the present invention, or a combination thereof. The control unit 101 realizes various functions by executing a program (such as the yield prediction program of the present invention) loaded into the main storage device 102 or the like. The yield prediction process in the yield prediction device 100 of the present invention can be performed by the control unit 101, for example.

[0075] The main memory device 102 stores various programs and also stores data necessary for executing the various programs. The main memory device 102 may include, for example, at least one of a ROM (Read Only Memory) and a RAM (Random Access Memory). The ROM stores various programs such as a BIOS (Basic Input / Output System), etc. There are no particular limitations on the ROM, and it can be selected appropriately depending on the purpose, and examples include mask ROM and PROM (Programmable ROM). The RAM functions as a working area in which various programs stored in, for example, the ROM or the auxiliary storage device 103 are deployed when they are executed by the control unit 101. There are no particular limitations on the RAM, and it can be selected appropriately depending on the purpose, and examples thereof include DRAM (Dynamic Random Access Memory) and SRAM (Static Random Access Memory).

[0076] The auxiliary storage device 103 is not particularly limited as long as it can store various types of information and can be appropriately selected depending on the purpose, and examples include a solid state drive (SSD), a hard disk drive (HDD), etc. The auxiliary storage device 103 may also be a portable storage device such as a CD drive, a DVD drive, or a BD (Blu-ray (registered trademark) Disc) drive. The yield prediction program of the present invention is stored in, for example, the auxiliary storage device 103, loaded into the RAM (main memory) of the main storage device 102, and executed by the control unit 101.

[0077] The I / O interface 104 is an interface for connecting various external devices, and enables input and output of data from, for example, a CD-ROM (Compact Disc ROM), a DVD-ROM (Digital Versatile Disk ROM), an MO disk (Magneto-Optical disk), a USB memory (USB (Universal Serial Bus) flash drive), and the like.

[0078] The communication interface 105 is not particularly limited, and any known interface may be used as appropriate, such as a wireless or wired communication device.

[0079] The input device 106 is not particularly limited as long as it can accept input of various requests and information to the yield prediction device 100 of the present invention, and any known device can be used as appropriate, such as a keyboard, a mouse, a touch panel, a microphone, etc. Furthermore, when the input device 106 is a touch panel (touch display), the input device 106 can also serve as the display device 108.

[0080] The output device 107 is not particularly limited, and any known device can be used as appropriate, such as a printer. The display device 108 is not particularly limited, and any known display device can be used as appropriate, such as a liquid crystal display or an organic EL display.

[0081] FIG. 2 shows an example of the functional configuration of the yield prediction device 100 of the present invention. As shown in FIG. 2, the yield prediction device 100 of the present invention includes a communication function unit 120, an input function unit 130, an output function unit 140, a display function unit 150, a storage function unit 160, and a control function unit 170.

[0082] The communication function unit 120 transmits and receives various data to and from external devices, for example, and may receive data such as information on the spectral shape of a compound from an external device. The input function unit 130 receives, for example, various instructions for the yield prediction apparatus 100 of the present invention. The input function unit 130 also receives, for example, information on the spectral shape of a compound. The output function unit 140 prints out, for example, information regarding the predicted yield of the synthesis result. The display function unit 150 displays, for example, information about the predicted yield of the synthesis result on a display.

[0083] The storage function unit 160 stores various programs, for example, and includes a preprocessing unit 161, a learning unit 162, and a database 163. The preprocessing unit 161 performs preprocessing such as limiting the spectrum range of the information on the spectral shape of the compound and differentiating the data values ​​on the information on the spectral shape of the compound. The learning unit 161 performs machine learning of information on the spectral shape of the compound. The database 163 is a database that stores information on the spectral shapes of the compounds.

[0084] The control function unit 170 includes a generation unit 171, a prediction unit 172, and the like. The generating unit 171 generates a data value from the information on the spectral shape of the compound. The prediction unit 172 predicts the yield of the synthesis result by comparing information on the synthesis result with information in the database. The control function unit 170 executes, for example, various programs stored in the storage function unit 160, and controls the overall operation of the yield prediction device 100 of the present invention.

[0085] (Method for synthesizing compounds) The compound synthesis method of the present invention includes a learning model construction step, a yield prediction step, and an optimization step, and may further include other steps. The method for synthesizing a compound of the present invention is a method for synthesizing a compound in which the spectral shape of each raw material compound used in a synthesis reaction partially overlaps with the spectral shape of the synthesis product of the synthesis reaction.

[0086] The learning model construction step and the yield prediction step are as described above in (Yield prediction method and yield prediction device).

[0087] <Optimization process> The optimization step is a step of changing the conditions of the synthesis reaction to perform a new synthesis reaction if the predicted yield is less than a predetermined yield value.

[0088] The predetermined yield value is not particularly limited and can be appropriately selected depending on the purpose, and examples thereof include 80% or 90%.

[0089] The conditions are not particularly limited and can be appropriately selected depending on the purpose. For example, when the compound synthesis method is a synthesis method using a microreactor, the conditions include the flow rate, temperature, and the like.

[0090] The optimization step can be performed automatically using a control device such as a programmable logic controller (PLC). The PLC is not particularly limited and can be appropriately selected depending on the purpose. For example, the MELSEC iQ-F series manufactured by Mitsubishi Electric Corporation can be used. The software for operating the PLC is not particularly limited and can be appropriately selected depending on the purpose. For example, SoftGOT 2000 manufactured by Mitsubishi Electric Corporation can be used. The PLC can be connected to a spectrometer (spectrophotometer), a pump, a sensor, a heater, a chiller, and the like. When an infrared spectrophotometer (IR) is used as the spectrometer, the software for running the infrared spectrophotometer (IR) is not particularly limited and can be selected appropriately depending on the purpose, and examples include iC IR manufactured by Mettler-Toledo Co., Ltd. The operation of the infrared spectrophotometer (IR) can be performed automatically by mouse simulation. When a nuclear magnetic resonance (NMR) spectrometer is used as the spectrometer, the software for performing the nuclear magnetic resonance (NMR) analysis is not particularly limited and can be appropriately selected depending on the purpose, and examples thereof include dedicated software manufactured by Magritek, Inc. The nuclear magnetic resonance (NMR) can be operated using API (remote control mode). [Example]

[0091] Examples of the present invention will be described below, but the present invention is not limited to these examples in any way.

[0092] 1. Yield prediction method In spectroscopic analysis, the intensity of the peaks depends on the abundance of each molecule, and when several molecules are mixed, their spectrum is displayed as a linear combination of them. They hypothesize that the spectra of the substances involved in a reaction, including the products, can be linearly combined to mimic the spectrum of the reaction, potentially enabling the rapid generation of vast amounts of training data from a small number of samples. As a model reaction, Suzuki-Miyaura cross-coupling (SMC) was chosen due to its importance and ubiquity in organic synthesis. In the SMC, the boryl group and (pseudo)halogeno group are converted into a C—C bond, and this conversion, like most cross-coupling reactions, makes little difference in FTIR (Fourier transform infrared) spectroscopy measurements.

[0093] <Spectral shape of each raw material compound and the spectral shape of the synthesized product> To prepare a training data set, first, the spectral shape of each raw compound and the spectral shape of the synthesized product were measured. A 0.049 M solution of 4,4,5,5-tetramethyl-2-phenyl-1,3,2-dioxaborolane (PhBpin) was prepared by dissolving 2-phenyl-4,4,5,5-tetramethyl-1,3,2-dioxaborolane (230 mg, 1.13 mmol, Tokyo Chemical Industry Co., Ltd.) and potassium hydroxide (85 mg, 1.52 mmol, Fujifilm Wako Pure Chemical Industries Co., Ltd.) in THF / methanol (23 mL, 1:1). A 0.033 M solution of 4-iodobenzonitrile (IPhCN) was prepared by dissolving 4-iodobenzonitrile (217 mg, 0.95 mmol, Tokyo Chemical Industry Co., Ltd.) and potassium hydroxide (106 mg, 1.89 mmol, Fujifilm Wako Pure Chemical Industries Co., Ltd.) in THF / methanol (28 mL, 1:1). A 0.033 M solution of 4-phenylbenzonitrile (PhPhCN) was prepared by dissolving 4-cyanobiphenyl (137 mg, 0.76 mmol, Tokyo Chemical Industry Co., Ltd.) and potassium hydroxide (85 mg, 1.52 mmol, Fujifilm Wako Pure Chemical Industries, Ltd.) in THF / methanol (23 mL, 1:1).

[0094] Each solution was pumped to the IR or NMR instrument at a flow rate of 1.0 mL / min using a syringe pump, and measurements were performed in-line. Five measurements were performed for each solution, taking noise patterns into consideration. To prevent learning of noise patterns, one of the measurements was randomly selected for linear combination. The upper part of FIG. 3 shows the FTIR spectrum of PhBpin (1), the middle part of FIG. 3 shows the FTIR spectrum of IPhCN (2), and the lower part of FIG. 3 shows the FTIR spectrum of PhPhCN (3). These spectra showed little difference to the naked eye.

[0095] <Preparing training data> Assuming that 1.5 equivalents of PhBpin (1) are beneficial for the reaction, a training data set was prepared according to the following equation 1:

number

[0096] <Creating a combined spectrum> Combined spectra were generated by a Python script. In the Python script, substrate, product, and comp1 are all lists containing five spectra each of IPhCN(2), PhPhCN(3), and PhBpin(1). 10,000 sets of data were generated. In the Python script, a random value between 0 and 100 was assigned to the variable (yield), the decomposition rate of comp1 (PhBpin) was randomly determined between 0 and 100, and spectra were randomly selected from the substrate, product, and comp1 and combined. This process was repeated num_data times (here, 10,000 times) to create num_data spectra.

[0097] <Creating a mixed spectrum> The flow microreactor system shown in Figure 4 was used, which consisted of three T-type micromixers (inner diameter 1000 μm, SUS304, manufactured by Sanko Seiki Kogyo Co., Ltd.), three microtube reactors (inner diameter 1000 μm, SUS316, manufactured by GL Sciences), and four pre-cooling units (inner diameter 1000 μm, SUS316, manufactured by GL Sciences).

[0098] A 0.0825 M solution of IPhCN was prepared by dissolving 4-iodobenzonitrile (909 mg, 3.97 mmol, Tokyo Chemical Industry Co., Ltd.) in THF / methanol (48 mL, 1:1). A 0.0825 M solution of PhBpin was prepared by dissolving 2-phenyl-4,4,5,5-tetramethyl-1,3,2-dioxaborolane (1252 mg, 6.13 mmol) in THF / methanol (66 mL, 1:1) and dividing the solution into two portions. A 0.04125 M solution of PhPhCN was prepared by dissolving 4-cyanobiphenyl (357 mg, 1.99 mmol) in THF / methanol (48 mL, 1:1).

[0099] The four prepared solutions were pumped using a syringe pump (flow rates of F1 mL / min, F2 mL / min, F3 mL / min, and F4 mL / min, see Figure 4) and mixed through three T-shaped mixers (total flow rate of 1.0 mL), followed by in-line IR measurement. To mimic each yield condition, F1, F2, F3, and F4 were changed as listed in Table 1. Three spectra were recorded under each condition to consider the noise pattern.

[0100] [Table 1]

[0101] Next, FTIR (Fourier Transform Infrared Spectroscopy) measurements were performed to verify whether the concentration (cyield) of PhPhCN(3) could be predicted from the obtained spectra. The correlation between the actual yield value and the predicted yield value is shown in the left graph (entire range) of Figure 5. Cyield was predicted to be approximately 100 for all spectra, which differed from the actual values.

[0102] <Pretreatment 1 of IR spectrum> To improve the prediction model, Pretreatment 1 of the IR spectrum was performed as follows. The scan setting of ReactIR (FTIR spectrophotometer, manufactured by Metrohm AG) was set to 128 times, and the sensitivity was set to "High" (every 4 wavenumbers). The raw data over the entire range was differentiated using the numpy.gradient function of the NumPy library. No additional processing such as normalization or standardization was performed. Similar to the case without pretreatment, it was verified whether the concentration (cyield) of PhPhCN(3) could be predicted from the obtained spectra. The results are shown in the middle graph (differentiation of the entire range) of Figure 5. The accuracy was improved compared to the case without pretreatment.

[0103] <Pretreatment 2 of IR spectrum> To further improve the prediction model, Pretreatment 2 of the IR spectrum was performed as follows. The scan setting of ReactIR (FTIR spectrophotometer, manufactured by Metrohm AG) was set to 128 times, and the sensitivity was set to "High" (every 4 wavenumbers). Only the raw data in the spectral range of 1692 to 699 cm -1 (434 data points), which is the fingerprint region, was differentiated using the numpy.gradient function of the NumPy library. No additional processing such as normalization or standardization was performed. We verified whether the concentration (yield) of PhPhCN(3) could be predicted from the obtained spectrum in the same way as without pretreatment. The results are shown in the graph on the right of Figure 5 (differential of the fingerprint region). Compared to without pretreatment, accuracy was improved. Furthermore, by using only the data in the fingerprint region, we were able to predict the yield with even higher accuracy.

[0104] Some mimic spectra are shown in Figure 6. The upper left image in Figure 6 mimics the spectrum obtained when the cross-coupling reaction described above gave a 49% yield, the upper right image in Figure 6 mimics the spectrum obtained when the cross-coupling reaction described above gave a 97% yield, the lower left image in Figure 6 mimics the spectrum obtained when the cross-coupling reaction described above gave a 71% yield, and the lower right image in Figure 6 mimics the spectrum obtained when the cross-coupling reaction described above gave a 31% yield.

[0105] <Creating a neural network model (NN model)> A neural network model was created using a Python script. When creating an IR model (434 data points) using an Intel Core i7-12700KF (3.61GHz), it took an average of 0.057 seconds to make one prediction.

[0106] These results show that limiting the spectral range was important for improving the prediction model, and that having more data points is not always advantageous.

[0107] Next, in order to predict the yield of the actual reaction, the reaction was carried out using a column reactor. [ka] 4-Iodobenzonitrile (6.998 g, 0.031 mol, manufactured by Tokyo Chemical Industry Co., Ltd.), 2-phenyl-4,4,5,5-tetramethyl-1,3,2-dioxaborolane (9.356 g, 0.046 mol, manufactured by Tokyo Chemical Industry Co., Ltd.), potassium hydroxide (3.421 g, 0.061 mol, manufactured by Fujifilm Wako Pure Chemical Industries Co., Ltd.), and n-tridecane (2.363 g, internal standard, manufactured by Fujifilm Wako Pure Chemical Industries Co., Ltd.) were dissolved in THF / methanol (926 mL, 1:1). The solution was pumped into the tube reactor using a TPL pump (Takumina Corporation). The tube reactor was immersed in a water bath equipped with a heater and a cooling jacket, and the outlet flow path was immersed in a room temperature water bath, to which an in-line IR device was connected.

[0108] The tube reactor was filled with a polymer material carrying palladium nanoparticles (the polymer material was silica gel or an organic polymer. The former was the one described in Catal. Today 2022 388-389 231-236. DOI: 10.1016 / j.cattod.2020.07.014, and the latter was the one described in JP 2023-078733 A).

[0109] The reaction liquid was sampled and subjected to gas chromatography analysis (GC analysis) by the internal standard method using n-tridecane. For gas chromatography analysis, a SHIMADZU GC-2010 and a capillary column (Rtx (registered trademark)-200, 30 m, 0.25 mm inner diameter, 0.25 μm) were used. The results are shown in Table 2.

[0110] [Table 2]

[0111] The results in Table 2 show that the yield predicted by the yield prediction method of the present invention is almost identical to the yield obtained by GC analysis. Therefore, it was found that the yield prediction method of the present invention can predict the yield under various conditions with high accuracy.

[0112] <Pretreatment of NMR Spectrum> Baseline drift correction and baseline zero correction of raw data were performed. The chemical shift was adjusted by the solvent peak of methanol (2.91 ppm). Only the aromatic region (5.8 - 10.7 ppm, 1000 data points) was used for the training and prediction of the neural network, and no additional processing such as normalization or standardization was performed. Similar to the case when using the IR spectrum, it was verified whether the concentration (cyield) of PhPhCN(3) could be predicted from the obtained spectrum. As in the case when using the IR spectrum, the yield could be predicted with high accuracy.

[0113] Some simulated spectra are shown in Fig. 7. The upper left figure in Fig. 7 is a figure simulating the spectrum when the above cross-coupling reaction had a yield of 33%, the upper right figure in Fig. 7 is a figure simulating the spectrum when the above cross-coupling reaction had a yield of 57%, the lower left figure in Fig. 7 is a figure simulating the spectrum when the above cross-coupling reaction had a yield of 15%, and the lower right figure in Fig. 7 is a figure simulating the spectrum when the above cross-coupling reaction had a yield of 80%.

[0114] <Flow of Learning Model Construction> The flowchart of learning model construction is shown in Fig. 8. As shown in Fig. 8, the learning model construction was executed according to the following flow. · Prepare the data of the yield and the corresponding chart (refer to the simulated spectrum) · Define and compile the neural network model · Execute the learning of the model · Execute the evaluation of the model · If the evaluation value meets the criteria, complete the model construction. If the evaluation value does not meet the criteria, define and compile the neural network model again.

[0115] The neural network model is defined as follows: Number of neurons: 1000 Activation function: ReLU (Rectified Linear Unit) Each of the three hidden layers generates 100 output neurons. Model loss function: Mean Squared Error Model Optimizer: Adaptive Moment Estimation

[0116] An example of the model evaluation results is shown in Figure 9. In Figure 9, the horizontal axis represents the answer on the chart, and the vertical axis represents the predicted yield when the chart is read by the model.

[0117] <Yield prediction flow> The yield prediction flow chart is shown in Figure 10. The yield prediction step used the python code (prediction=model.predict(spectra)). The prediction is a variable (which stores the predicted value obtained from the model), the model is a model class (Keras library), and the spectrum is the measured spectral data.

[0118] The above results demonstrate that the yield prediction method of the present invention makes it possible to build a model (in-line analytical model) using only the spectra of the substrate and product, and to accurately predict reaction yields, a task that is difficult for the human eye to accomplish. This will accelerate the analytical steps in research and development, including the construction of analytical models.

[0119] 2. Automatic optimization control system After building an inline analysis model, we considered building a fully automatic optimization system. A programmable logic controller (PLC) was connected to the pump and heater to control the flow rate and temperature. To verify stabilization, the PLC was connected to a flow meter, pressure gauge, and temperature gauge. The PLC was also connected to a laptop PC and a Fourier transform infrared spectrophotometer (FTIR) to control and monitor the system.

[0120] A schematic diagram of the automatic optimization control system is shown in Figure 11. Overall system control, including file management and data recording, was written in C#.NET. Machine learning (prediction using Bayesian optimization and neural network models) was performed using Python, and scripts were executed in C#.NET. IR measurements using ReactIR (FTIR spectrophotometer, Mettler-Toledo) were performed using software (iC IR, Mettler-Toledo), and the operation was automatically performed using mouse simulation. NMR analysis using Spinsolve was performed using dedicated software (Magritek), which was operated using API (remote control mode). The pump, sensor, heater, and chiller were connected to a programmable logic controller (PLC) (Mitsubishi Electric Corporation, MELSEC iQ-F series), and the PLC was operated using software (Mitsubishi Electric Corporation, SoftGOT 2000). The pump motor speed was controlled by changing the voltage (0-10V). The pressure and flow meter values ​​were obtained from the ammeter (4-20mA). The reaction temperature was obtained from a thermocouple connected to a PLC. These values ​​were recorded every second. The heater and chiller pump were connected to a relay to control the on / off of the power supply.

[0121] <Details of Bayesian optimization> A Bayesian optimization method similar to that described in Front. Chem. Sci. Eng. 2022 3 819752 was used. The script was written in Python using the GPyOpt library. Each element of 'initial_X' corresponds to a previously searched condition, and each predicted yield from these conditions is stored in 'initial_y'. 'random_seed' was fixed for each optimization process.

[0122] <Automated optimization process flow (IR)> A flowchart of the automated optimization process when using IR is shown in Figure 12. The initial conditions were randomly selected from 30 potential conditions (five flow rates and six temperatures). After predicting the reaction and its reaction yield, Bayesian optimization suggested the next candidate condition and ran the next reaction. This cycle was repeated until the selected condition had already run.

[0123] As shown in Figure 12, the automated optimization was performed using the following flow: First, determine the search range for Bayesian optimization (i.e., determine which reaction conditions and from which values ​​to which values ​​to consider. In this experiment, the reaction temperatures were set to 30, 40, 50, 60, 70, and 80°C (discrete), and the flow rates were set to 2, 3, 4, 5, and 6 mL / min (discrete).) If no initial data was available (in the first iteration), the initial conditions were randomly determined. (In subsequent iterations, the conditions suggested by Bayesian optimization were adopted as the next reaction conditions. At that time, the temperature and flow rate were automatically adjusted to achieve the suggested conditions.) Sends commands to the pump and heater to ensure the conditions set out above are met Wait until the temperature and flow rate reach the values ​​specified above (read the values ​​from the flow rate meter and thermometer, and wait until the set values ​​are maintained for the specified time or longer). (Wait for 480 / F seconds (F is flow rate, in mL / min) until the flow condition stabilizes.) After stabilization, perform IR measurement, then obtain a sample for GC analysis if necessary, and then change the flow rate to 1.0 mL / min. Enter standby mode (a state designed to reduce waste until the next reaction conditions are determined) to conserve the reaction solution. (Specifically, keep the flow at a rate of 0.1 mL / min. The pump can be stopped, but stopping the flow while the reactor is heated may cause an increase in internal pressure or clogging due to solid precipitation.) The yield is predicted (calculated) using the measured IR spectrum and a pre-prepared prediction model (the cycle (iteration) ends when the yield is predicted (calculated) from the spectrum, and the next cycle automatically begins). - Bayesian optimization is performed based on reaction conditions and predicted yield data to select conditions with the highest expected value. At this time, if the selected condition is already implemented, the optimization process ends; if it is not implemented, the process moves to the next decision. If the number of trials so far is greater than or equal to the set value (max_iter), the process ends. If it is less than or equal to the set value (max_iter), the conditions selected by Bayesian optimization are considered (looped).

[0124] That is, the process ends when either 1. the condition selected as the better condition has already been tried, or 2. the number of tries reaches the upper limit. The conditions with the highest predicted yield among the trials up to that point (not necessarily the last ones selected) are selected as the optimal conditions.

[0125] <Automated optimization process flow (NMR)> A flow chart of the automated optimization process using NMR is shown in FIG. When NMR was used, shim adjustment was performed before changing the temperature and flow rate (shim adjustment was performed for each cycle (iteration)). As with the use of IR, in-line NMR (Magritek Spinsolve60 ULTRA) was used in Suzuki coupling reactions, and high yield prediction performance and automatic optimization were possible.

[0126] Automatic optimization with inline IR Figure 14 shows a schematic diagram of a flow system equipped with an in-line IR. 4-Iodobenzonitrile (6.998 g, 0.031 mol, Tokyo Chemical Industry Co., Ltd.), 2-phenyl-4,4,5,5-tetramethyl-1,3,2-dioxaborolane (9.356 g, 0.046 mol, Tokyo Chemical Industry Co., Ltd.), potassium hydroxide (3.421 g, 0.061 mol, Fujifilm Wako Pure Chemical Industries Co., Ltd.), and n-tridecane (2.363 g, internal standard, Fujifilm Wako Pure Chemical Industries Co., Ltd.) were dissolved in THF / methanol (926 mL, 1:1). The solution was pumped into a tube reactor using a TPL pump (Takumina Corporation). The tube reactor was immersed in a water bath equipped with a heater and cooling jacket, and the outlet flow path was immersed in a room-temperature water bath, to which an in-line IR instrument was connected.

[0127] The tube reactor was filled with a polymer material carrying palladium nanoparticles (the polymer material was silica gel or an organic polymer. The former was the one described in Catal. Today 2022 388-389 231-236. DOI: 10.1016 / j.cattod.2020.07.014, and the latter was the one described in JP 2023-078733 A).

[0128] In the optimization process, two reactions were run to obtain initial data, and Bayesian optimization was performed from the third run onward. Real-time reaction monitoring by IR was performed every minute (128 scans). Spectroscopic measurements were performed when the reaction conditions were determined to be stable, with an additional wait of 240 seconds and 128 scans. Subsequent sampling was performed with an additional wait of 120 seconds.

[0129] The results of the first optimization trials are shown in Table 3. In the first optimization run, 6.0 mL / min and 50 °C were selected as the initial conditions, and after six reactions, Bayesian optimization suggested a candidate that had already been explored (2.0 mL / min and 80 °C). Therefore, the system concluded that 2.0 mL / min and 80°C were the best conditions as they resulted in the highest predicted yield. Remarkably, this series of experiments required only 2 hours to find the optimal conditions.

[0130] [Table 3]

[0131] The results of the second optimization run, in which different initial conditions were chosen, are shown in Table 4. As shown in Table 4, the second trial also concluded that the same conditions were the best conditions. Although the second trial required three more reactions to complete the optimization, the overall experimental time was still sufficiently short. The above results ensured the speed and reproducibility of the constructed optimization system.

[0132] [Table 4]

[0133] The accuracy of the prediction was tested using several reaction solutions with different conditions. The correlation between actual and predicted yield values ​​is shown in FIG. As shown in Figure 15, the yields predicted by the model were in good agreement with those determined by GC analysis using an internal standard, which is the conventional method (R 2 The value was 0.93442, which was an extremely accurate prediction.

[0134] Real-time reaction monitoring Predictive models were utilized for real-time analysis and process analytical technology (PAT). The temperature, flow, and pressure gauges were connected to the PLC as shown in Figure 16. The connected gauges monitored the temperature, flow, and pressure of the reaction every second. As well as these physical parameters, the yield was also recorded every minute by in-line FTIR. The top panel of Figure 17 shows the yield log recorded by in-line FTIR, and the bottom panel of Figure 17 shows the temperature, flow rate, and pressure logs of the reaction. Sampling was performed at the points indicated by the arrows in the upper part of Figure 17. This shows that FTIR-based yield prediction can function as a PAT system without further preparation of a predictive model. Regions where the physical gauges indicated the reaction was stable were not necessarily considered stable by the yields predicted by FTIR. This demonstrates that the in-line FTIR can detect small differences that do not manifest as physical differences such as temperature.

[0135] Automated optimization using in-line NMR Figure 18 shows a schematic diagram of a flow system equipped with in-line NMR. Since the intensity of the in-line NMR spectrum depends on the flow rate of the sample, the outlet flow path was branched and connected to a Q pump (manufactured by Takumina Corporation). A portion of the reaction solution was introduced into the in-line NMR flow cell at 1.0 mL / min (the manufacturer's recommended value for flow NMR measurements in Spinsolve) and measured. Other setups and conditions were the same as for in-line IR.

[0136] From the above, we have developed a simple method for preparing data for machine learning, and established a method for efficiently and accurately predicting product yields and a method for synthesizing compounds using the same. It was also shown that this yield prediction method has the potential to function as a PAT capable of detecting small differences. The system of the present invention can make a significant contribution to industrial automation. [Explanation of symbols]

[0137] 100 Yield Prediction Device 101 Control section 102 Main storage 103 Auxiliary storage device 104 I / O Interface 105 Communication Interface 106 Input Device 107 Output Device 108 Display device 109 System Bus 120 Communication function unit 130 Input function unit 140 Output function unit 150 Display function section 160 Memory function unit 161 Preprocessing section 162 Learning Department 163 databases 170 Control Function Unit 171 Generation part 172 Prediction Department

Claims

1. In the case where the spectral shape of each raw material compound in all raw material compounds used in a synthesis reaction partially overlaps with the spectral shape of the synthesis product of the synthesis reaction, a learning model construction step of constructing a learning model by machine learning a residual spectrum, which is a remainder of the combined spectrum obtained by linearly combining the spectra of the raw material compounds and the spectrum of the synthetic product in an arbitrary ratio, excluding the overlapping portions of the spectral shapes of the combined spectrum; a yield prediction step of predicting the yield of the synthesis resultant product by comparing the spectrum of the synthesis resultant product obtained by the synthesis reaction of each of the raw material compounds with the remaining spectrum in the machine-learned learning model; A method for predicting a yield, comprising:

2. In the case where the spectral shape of each raw material compound in all raw material compounds used in a synthesis reaction partially overlaps with the spectral shape of the synthesis product of the synthesis reaction, a learning model construction step of constructing a learning model by machine learning a differential residual spectrum obtained by differentiating a combined spectrum obtained by linearly combining the spectra of the raw material compounds and the spectrum of the synthetic product in an arbitrary ratio, excluding a partially overlapping portion of the spectral shape of the combined spectrum; and a yield prediction step of predicting the yield of the synthesis resultant product by comparing the spectrum of the synthesis resultant product obtained by the synthesis reaction of each of the raw material compounds with the differential residual spectrum in the machine-learned learning model; A method for predicting a yield, comprising:

3. 3. The yield prediction method according to claim 1, wherein the spectrum of each raw material compound, the spectrum of the synthesis product, and the spectrum of the synthesis resultant are spectra obtained by optical analysis.

4. 4. The yield prediction method according to claim 3, wherein the spectrum obtained by the optical analysis is at least one selected from an infrared absorption (IR) spectrum, a nuclear magnetic resonance (NMR) spectrum, a near-infrared absorption (NIR) spectrum, and a Raman spectroscopy spectrum.

5. 4. The method for predicting a yield according to claim 3, wherein the optical analysis is carried out by an optical analyzer that continuously optically analyzes the synthesis products obtained by the synthesis reaction of the raw material compounds.

6. In the case where the spectral shape of each raw material compound in all raw material compounds used in a synthesis reaction partially overlaps with the spectral shape of the synthesis product of the synthesis reaction, a learning model construction step of constructing a learning model by machine learning a residual spectrum, which is a remainder of the combined spectrum obtained by linearly combining the spectra of the raw material compounds and the spectrum of the synthetic product in an arbitrary ratio, excluding the overlapping portions of the spectral shapes of the combined spectrum; a yield prediction step of predicting the yield of the synthesis resultant product by comparing the spectrum of the synthesis resultant product obtained by the synthesis reaction of each of the raw material compounds with the remaining spectrum in the machine-learned learning model; an optimization step of changing the conditions of the synthesis reaction to perform a new synthesis reaction if the predicted yield is less than a predetermined yield value; A method for synthesizing a compound, comprising:

7. In the case where the spectral shape of each raw material compound in all raw material compounds used in a synthesis reaction partially overlaps with the spectral shape of the synthesis product of the synthesis reaction, a learning model construction step of constructing a learning model by machine learning a differential residual spectrum obtained by differentiating a combined spectrum obtained by linearly combining the spectra of the raw material compounds and the spectrum of the synthetic product in an arbitrary ratio, excluding a partially overlapping portion of the spectral shape of the combined spectrum; and a yield prediction step of predicting the yield of the synthesis resultant product by comparing the spectrum of the synthesis resultant product obtained by the synthesis reaction of each of the raw material compounds with the differential residual spectrum in the machine-learned learning model; an optimization step of changing the conditions of the synthesis reaction and performing a new synthesis reaction if the predicted yield is less than a predetermined yield value; A method for synthesizing a compound, comprising:

8. 8. The method for synthesizing a compound according to claim 6, wherein, in the optimization step, if the predicted yield is less than a predetermined yield value, the spectrum of a new synthesis resultant obtained by performing a new synthesis reaction under different synthesis reaction conditions is compared with the remainder spectrum or the differential remainder spectrum to predict the yield of the new synthesis resultant, and the synthesis reaction is repeated until the newly predicted yield of the synthesis resultant becomes equal to or greater than the predetermined yield value.

9. In the case where the spectral shape of each raw material compound in all raw material compounds used in a synthesis reaction partially overlaps with the spectral shape of the synthesis product of the synthesis reaction, a learning model construction means for constructing a learning model by machine learning a residual spectrum, which is a remainder of the combined spectrum obtained by linearly combining the spectra of the raw material compounds and the spectrum of the synthetic product in an arbitrary ratio, excluding the overlapping spectral shape portion of the combined spectrum; and a yield prediction means for predicting the yield of the synthesis resultant product by comparing the spectrum of the synthesis resultant product obtained by the synthesis reaction of each of the raw material compounds with the remaining spectrum in the machine-learned learning model; A yield prediction device comprising:

10. In the case where the spectral shape of each raw material compound in all raw material compounds used in a synthesis reaction partially overlaps with the spectral shape of the synthesis product of the synthesis reaction, a learning model construction means for constructing a learning model by machine learning a differential residual spectrum obtained by differentiating a combined spectrum obtained by linearly combining the spectra of the raw material compounds and the spectrum of the synthetic product in an arbitrary ratio, excluding a portion of the spectrum shape that overlaps with the combined spectrum; and a yield prediction means for predicting the yield of the synthesis resultant by comparing the spectrum of the synthesis resultant obtained by the synthesis reaction of each of the raw material compounds with the differential residual spectrum in the machine-learned learning model; A yield prediction device comprising:

11. In the case where the spectral shape of each raw material compound in all raw material compounds used in a synthesis reaction partially overlaps with the spectral shape of the synthesis product of the synthesis reaction, A learning model characterized by being obtained by machine learning a residual spectrum, which is the remainder of the combined spectrum obtained by linearly combining the spectra of the raw material compounds and the spectrum of the synthetic product in any ratio, excluding any overlapping portions of the spectral shapes of the combined spectrum.

12. In the case where the spectral shape of each raw material compound in all raw material compounds used in a synthesis reaction partially overlaps with the spectral shape of the synthesis product of the synthesis reaction, A learning model characterized by being machine-learned using a differential residual spectrum obtained by differentiating a combined spectrum obtained by linearly combining the spectra of the raw material compounds and the spectrum of the synthetic product in any ratio, excluding any overlapping spectral shape portions of the combined spectrum.

13. In the case where the spectral shape of each raw material compound in all raw material compounds used in a synthesis reaction partially overlaps with the spectral shape of the synthesis product of the synthesis reaction, a process of constructing a learning model by machine learning a residual spectrum, which is a remainder of the combined spectrum obtained by linearly combining the spectra of the raw material compounds and the spectrum of the synthetic product in an arbitrary ratio, excluding a portion of the spectral shape that overlaps the combined spectrum; a process of predicting a yield of the synthesis resultant product by comparing a spectrum of the synthesis resultant product obtained by the synthesis reaction of each of the raw material compounds with the remaining spectrum in the machine-learned learning model; A yield prediction program that causes a computer to perform the above.

14. In the case where the spectral shape of each raw material compound in all raw material compounds used in a synthesis reaction partially overlaps with the spectral shape of the synthesis product of the synthesis reaction, a process of constructing a learning model by machine learning a differential residual spectrum obtained by differentiating a residual of the combined spectrum obtained by linearly combining the spectra of the raw material compounds and the spectrum of the synthetic product in an arbitrary ratio, excluding a portion of the spectral shape that overlaps the combined spectrum; and a process of predicting the yield of the synthesis resultant product by comparing the spectrum of the synthesis resultant product obtained by the synthesis reaction of each of the raw material compounds with the differential residual spectrum in the machine-learned learning model; A yield prediction program that causes a computer to perform the above.

Images