Physical property prediction system

Abstract

To provide a physical property prediction method and system that allow anyone to simply and accurately predict physical properties of an organic compound.SOLUTION: A physical property prediction method and system of an organic compound are provided that includes the steps of: learning a correlation between a molecular structure and physical properties of an organic compound; and predicting a target physical property value from a molecular structure of a target material based on the learning results and simultaneously uses a plurality of kinds of fingerprint methods as a method for describing the molecular structure of the organic compound.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] One embodiment of the present invention relates to a method and an apparatus for predicting properties of an organic compound. [Background technology]

[0002] In the past, the physical properties of organic compounds could only be known by synthesizing the target substance and measuring it directly. However, these properties are determined by the molecular structure of the organic compound. Therefore, it is difficult to know the approximate values ​​of the physical properties of an organic compound with a certain molecular structure. Nowadays, with the amount of data accumulated, it is easy for an experienced person to spot whether a particular In recent years, it has become possible to calculate Prediction is also possible by

[0003] In research and development using organic compounds, we select the appropriate physical properties according to the required characteristics. Therefore, it is possible to select and use organic compounds having the above properties without actually synthesizing them. Accurately predicting and selecting organic compounds with the required physical properties from unknown or unknown substances If this could be achieved, it is expected that the speed of development could be greatly improved.

[0004] However, not everyone can make accurate predictions like those mentioned above, and at present, On the other hand, the candidate organic compounds are Since there are always many organic compounds, it is necessary to develop a method that allows anyone to easily and quickly predict the properties of the target organic compound. and systems are desired.

[0005] In recent years, methods for classification, estimation, prediction, etc. using machine learning and other methods have undergone significant advances. In particular, selection by deep learning using convolutional neural networks The performance of classification and prediction has improved significantly, and has achieved excellent results in a variety of fields. However, in the field of organic compounds, it is difficult to make a computer understand the structure of the compound without any discrepancies. Moreover, it is possible to accurately extract features related to physical properties, and to obtain a manageable amount of information. The current situation is that there are still almost no adequate methods for describing an organic compound. Therefore, this method allows anyone to easily and accurately predict the properties of organic compounds. And the system has not yet been realized.

[0006] Patent Document 1 discloses a method for searching for new substances using machine learning and an apparatus for the same. There are. [Prior art documents] [Patent documents]

[0007] [Patent Document 1] JP 2017-91526 A Summary of the Invention [Problem to be solved by the invention]

[0008] In one aspect of the present invention, the physical properties of unknown organic compounds can be predicted easily and accurately by anyone. The present invention aims to provide a method for predicting the physical properties of organic compounds. The objective of the present invention is to provide a system for predicting physical properties that enables anyone to easily and accurately predict the properties of a material. Let us assume that. [Means for solving the problem]

[0009] One aspect of the present invention includes a step of learning the correlation between the molecular structure and physical properties of an organic compound, and predicting a desired physical property from the molecular structure of the target substance based on the result. As a method for expressing the molecular structure of a compound, it is useful to use multiple types of fingerprinting methods simultaneously. This is a method for predicting the physical properties of organic compounds.

[0010] Another aspect of the present invention is a method for learning a correlation between a molecular structure and a physical property of an organic compound, comprising: and predicting a desired physical property from the molecular structure of the target substance based on the results of said learning. As a method for expressing the molecular structure of the organic compound, two types of fingerprint methods are used simultaneously. This is a method for predicting the physical properties of organic compounds.

[0011] Another aspect of the present invention is a method for learning a correlation between a molecular structure and a physical property of an organic compound, comprising: and predicting a target physical property from the molecular structure of the target substance based on the result of the learning. Three types of fingerprinting methods are used simultaneously to describe the molecular structure of the organic compound. This is a method for predicting the physical properties of organic compounds.

[0012] In addition, in the above-mentioned configuration, another aspect of the present invention is a method for detecting a fingerprint using At om Pair type, Circular type, Substructure key type and P This is a property prediction method including at least one of the ath-based types.

[0013] In addition, another aspect of the present invention is a method for detecting a fingerprint of a subject in the above configuration, wherein the plurality of fingerprint methods are: Atom Pair type, Circular type, Substructure key type and This is a property prediction method selected from among the path-based and the nonlinear type.

[0014] Another aspect of the present invention is a method for detecting a fingerprint using the fingerprint detection method according to the above configuration. m A property prediction method including Pair type and Circular type.

[0015] In addition, another aspect of the present invention is a method for detecting a fingerprint using a Ci This is a property prediction method that includes the rcular type and the substructure key type.

[0016] In addition, another aspect of the present invention is a method for detecting a fingerprint using a Ci These are property prediction methods including ricular and path-based types.

[0017] In addition, in the above-mentioned configuration, another aspect of the present invention is a method for detecting a fingerprint using At om Pair type and Substructure key type are included in the property prediction method. .

[0018] In addition, in the above-mentioned configuration, another aspect of the present invention is a method for detecting a fingerprint using At These are property prediction methods including Pair type and Path-based type.

[0019] In addition, another aspect of the present invention is a method for detecting a fingerprint of a subject having a cellular environment, comprising: tom Pair type, Substructure key type and Circular type The present invention relates to a method for predicting physical properties.

[0020] In addition, another aspect of the present invention is a method for detecting a fingerprint of a subject in the above configuration, When the circular type is used, it is a property prediction method in which r is 3 or more.

[0021] In addition, another aspect of the present invention is a method for manufacturing a circular finger in the above-mentioned configuration. The -print method is a physical property prediction method in which r is 5 or more.

[0022] Another aspect of the present invention is a method for detecting a fingerprint of a subject having a fingerprint, comprising: When the molecular structure of each organic compound to be trained is represented using 1, the representation of each organic compound is All of these are different methods for predicting physical properties.

[0023] Another aspect of the present invention is a method for detecting a fingerprint of a subject having a fingerprint, comprising: The first is a property prediction method that can express structural information that characterizes the property to be predicted.

[0024] Another aspect of the present invention is a method for detecting a fingerprint of a subject having a fingerprint, comprising: The substituent, the substitution position of the substituent, the functional group, the number of elements, the type of element, the valence of the element, the bond This is a property prediction method that can express at least one of the compound order and atomic coordinates.

[0025] In another aspect of the present invention, in the above-mentioned configuration, the physical properties are an emission spectrum, a half-width , emission energy, excitation spectrum, absorption spectrum, transmission spectrum, reflection spectrum , molar extinction coefficient, excitation energy, transient emission lifetime, transient absorption lifetime, S1 level, T1 level, Sn level, Tn level, Stokes shift value, luminescence quantum yield, oscillator strength, oxidation potential, reduction Electric potential, HOMO level, LUMO level, glass transition point, melting point, crystallization temperature, decomposition temperature, boiling point , sublimation temperature, carrier mobility, refractive index, orientation parameter, mass-to-charge ratio and NMR measurements Spectrum, chemical shift value and number of elements or coupling constant, ESR measurement The physical property prediction is one or more of the spectrum, g factor, D value, or E value in the determination. This is a measurement method.

[0026] Another aspect of the present invention is a method for making a computer-readable recording medium comprising: an input unit; a data server; and data stored in the data server. A learning means for learning the correlation between the molecular structure and physical properties of the organic compound, and based on the result of the learning, A prediction means for predicting a desired physical property from the molecular structure of a target substance inputted from the input means. and an output unit for outputting the predicted physical property value, As a method for this, we developed a system for predicting the physical properties of organic compounds using multiple fingerprinting methods simultaneously. It is a stem.

[0027] Another aspect of the present invention is a method for making a computer-readable recording medium comprising: an input unit; a data server; and data stored in the data server. A learning means for learning the correlation between the molecular structure and physical properties of the organic compound, and a learning means for learning the correlation between the molecular structure and physical properties of the organic compound based on the results of the learning. A prediction method for predicting a desired physical property from a molecular structure of a target substance inputted from the input means. and an output unit for outputting the predicted physical property value, A system for predicting the physical properties of organic compounds that uses two different fingerprint methods simultaneously. It is a stem.

[0028] Another aspect of the present invention is a method for making a computer-readable recording medium comprising: an input unit; a data server; and data stored in the data server. A learning means for learning the correlation between the molecular structure and physical properties of the organic compound, and a learning means for learning the correlation between the molecular structure and physical properties of the organic compound based on the results of the learning. A prediction method for predicting a desired physical property from a molecular structure of a target substance inputted from the input means. and an output unit for outputting the predicted physical property value, A system for predicting the physical properties of organic compounds that uses three different fingerprint methods simultaneously. It is a stem.

[0029] In addition, in the above-mentioned configuration, another aspect of the present invention is a method for detecting a fingerprint using At om Pair type, Circular type, Substructure key type and P The physical property prediction system includes at least one of the ath-based types.

[0030] In addition, another aspect of the present invention is a method for detecting a fingerprint of a subject in the above configuration, tom Pair type, Circular type, Substructure key type and This is a property prediction system selected from among path-based types.

[0031] Another aspect of the present invention is a method for detecting a fingerprint using the fingerprint detection method according to the above configuration. m A physical property prediction system including Pair type and Circular type.

[0032] Another aspect of the present invention is a method for detecting a fingerprint using a Cir It is a property prediction system that includes the cular type and the substructure key type. .

[0033] In addition, another aspect of the present invention is a method for detecting a fingerprint using a Ci It is a property prediction system that includes real-type and path-based types.

[0034] In addition, in the above-mentioned configuration, another aspect of the present invention is a method for detecting a fingerprint using At om Pair type and / or Substructure key type It is a stem.

[0035] In addition, in the above-mentioned configuration, another aspect of the present invention is a method for detecting a fingerprint using At om Pair-type and / or Path-based property prediction systems .

[0036] In addition, another aspect of the present invention is a method for detecting a fingerprint of a subject having a cellular environment, comprising: tom Pair type, Substructure key type and Circular type It is a physical property prediction system including

[0037] In addition, another aspect of the present invention is a method for detecting a fingerprint of a subject in the above configuration, When the circular type is used, it is a property prediction system in which r is 3 or more.

[0038] In addition, another aspect of the present invention is a circular fin in the above configuration. The Garprint method is a property prediction system with an r of 5 or more.

[0039] Another aspect of the present invention is a method for detecting a fingerprint of a subject having a fingerprint, comprising: When the molecular structure of each organic compound to be trained is represented using 1, the representation of each organic compound is All of these are different property prediction systems.

[0040] Another aspect of the present invention is a method for detecting a fingerprint of a subject having a fingerprint, comprising: The first is a property prediction system that can express structural information that characterizes the properties to be predicted. do.

[0041] Another aspect of the present invention is a method for detecting a fingerprint of a subject having a fingerprint, comprising: The substituent, the substitution position of the substituent, the functional group, the number of elements, the type of element, the valence of the element, the bond This is a property prediction system that can express at least one of the compound order and atomic coordinates.

[0042] In another aspect of the present invention, in the above-mentioned configuration, the physical properties are an emission spectrum, a half-width , emission energy, excitation spectrum, absorption spectrum, transmission spectrum, reflection spectrum , molar extinction coefficient, excitation energy, transient emission lifetime, transient absorption lifetime, S1 level, T1 level, Sn level, Tn level, Stokes shift value, luminescence quantum yield, oscillator strength, oxidation potential, reduction Electric potential, HOMO level, LUMO level, glass transition point, melting point, crystallization temperature, decomposition temperature, boiling point , sublimation temperature, carrier mobility, refractive index, orientation parameter, mass-to-charge ratio and NMR measurements Spectrum, chemical shift value and number of elements or coupling constant, ESR measurement The physical property prediction is one or more of the spectrum, g factor, D value, or E value in the determination. It is a measurement system. Effect of the Invention

[0043] In one aspect of the present invention, the physical properties of unknown organic compounds can be predicted easily and accurately by anyone. Furthermore, the present invention provides a method for predicting the properties of organic compounds that can be easily understood by anyone. It is possible to provide a property prediction system that can easily and accurately predict the properties of a material. [Brief description of the drawings]

[0044] [Figure 1] 1 is a flowchart illustrating one embodiment of the present invention. [Diagram 2] A diagram showing a method for converting molecular structures using fingerprinting. [Diagram 3] FIG. 1 is a diagram for explaining types of fingerprinting methods. [Figure 4] A diagram explaining the conversion from SMILES notation to fingerprint notation. [Diagram 5] A diagram explaining the types of fingerprinting methods and overlapping notations. [Figure 6] FIG. 1 is a diagram for explaining an example of a molecular structure represented using multiple fingerprinting methods. [Figure 7] FIG. 1 is a diagram for explaining the configuration of a neural network. [Figure 8] FIG. 1 is a diagram showing a property prediction system according to one embodiment of the present invention. [Figure 9] FIG. 1 is a diagram for explaining the configuration of a neural network. [Figure 10] 1A to 1C are diagrams illustrating examples of the configuration of a semiconductor device having a function of performing arithmetic operations. [Figure 11] 1A and 1B are diagrams illustrating a specific configuration example of a memory cell. [Figure 12] FIG. 2 is a diagram for explaining an example of the configuration of an offset circuit OFST. [Figure 13] 10 is a timing chart illustrating an example of the operation of a semiconductor device. [Figure 14] Graph showing the results of physical property prediction. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0045] Hereinafter, the embodiment of the present invention will be described in detail with reference to the drawings. The present invention is not limited to the above description, and the form and details are not limited thereto without departing from the spirit and scope of the present invention. It will be readily understood by those skilled in the art that various modifications may be made. It should not be construed as being limited to the description of the embodiment modes.

[0046] (Embodiment 1) The method for predicting physical properties according to one embodiment of the present invention can be shown, for example, in a flowchart as shown in FIG. According to FIG. 1, the method for predicting physical properties according to one embodiment of the present invention includes: The correlation of physical properties is learned (S101).

[0047] In order to machine-learn the correlation between molecular structure and physical properties, it is necessary to describe the molecular structure in a mathematical formula. For the mathematical formulation of molecular structures, we used the open source cheminformatics toolkit A RDKit is available that can generate SMIL files of the molecular structure you input. ES notation (Simplified molecular input line ent ry specification syntax) by fingerprinting It can be converted into formula data.

[0048] In the fingerprint method, for example, as shown in Figure 2, partial structures (fragment structures) of a molecular structure are identified. The molecular structure is represented by assigning a bit (bit) to each bit, and if the corresponding partial structure exists in the molecule, If it is, the bit is set to "1" and if it is not, the bit is set to "0". By using this method, we can obtain a formula that extracts the characteristics of the molecular structure. The molecular structure formulas expressed by the fingerprint method are hundreds to tens of thousands of bits long, making them easy to handle. In order to express the molecular structure as a mathematical formula of 0 and 1, the fingerprint method is used. By using this, it is possible to realize extremely high-speed calculation processing.

[0049] In addition, there are many types of fingerprinting methods (differences in bit generation algorithms, atomic tag Considering the type, bond type, aromaticity condition, etc., dynamically bit-for-bit using hash function There are various types of scalars (such as those that generate long lengths), each with their own characteristics.

[0050] Representative types of fingerprinting methods are as shown in Figure 3. 1) Circul ar type (atoms surrounding the starting atom up to a specified radius are treated as partial structures), 2 ) Path-based type (The length of the path from the starting atom (Path len gth) as a substructure), 3) Substructure keys type ( 4) Atom pair type (all the structures in the molecule are specified by bit) The RDKit has these A fingerprint for each type is implemented.

[0051] Figure 4 shows the molecular structure of an organic compound, expressed as a mathematical formula using the fingerprint method. In this way, the molecular structure is first converted to SMILES notation and then the finger is - It can be converted into a print.

[0052] In addition, when expressing the molecular structure of an organic compound using the fingerprint method, In some cases, the formulas obtained for different organic compounds are the same. There are several different types of fingerprinting methods, but they are not the same. The tendency of the compounds to become like this is shown in Figure 5, 1) Circular type (Morgan Fing type). erprint), 2)Path-based type (RDK Fingerprint), 3) Substructure keys type (Avalon Fingerprint) 4) As shown in Atom pair type (Hash atom pair), In Figure 5, the molecules in each double arrow are the same. Therefore, the fingerprint method used for learning is When the molecular structure of each organic compound to be trained is represented using at least 1, It is preferable to use a fingerprinting method in which all notations are different. In Figure 5, Atom It can be seen that compounds with different pair types can be represented without duplication. Depending on the population of organic compounds, other notation methods may be possible without duplication. do.

[0053] In one embodiment of the present invention, the organic compounds to be learned are represented by a fingerprint method. The method is characterized in that a plurality of different types of fingerprinting methods are used. There can be any number of types, but two or three types are preferable because they are easier to handle in terms of data volume. When training with multiple fingerprinting methods, Formulas written by other types of fingerprinting methods are shown after the formula written by can be used in combination, or there can be multiple different formulas for one organic compound. In Fig. 6, multiple fingerprints with different types are used for classification. An example of a method for describing a child structure is shown below.

[0054] A fingerprint is a method to describe the presence or absence of a substructure; the information of the entire molecular structure is not lost. However, if multiple fingerprints of different types are used to mathematically model the molecular structure, For example, different substructures are generated for each fingerprint type, and these substructures The presence or absence of a specific molecule can provide information about the entire molecular structure. In cases where a characteristic that cannot be expressed has a large effect on the physical property value, or where the physical property between compounds that have such a characteristic is different, If it affects the value difference, it will be complemented by other fingerprints, so that the difference between different types It is effective to describe a molecular structure using multiple fingerprints.

[0055] In addition, when notating using the two types of fingerprinting methods, Atom Pair type and The circular type is the preferred configuration because it allows accurate prediction of physical properties. be.

[0056] In addition, when using the three types of fingerprinting methods, the Atom Pair type and The circular type and the substructure keys type are more accurate. This is a preferable configuration because it allows for easy prediction of physical properties.

[0057] In addition, when using the circular fingerprint method, the radius r must be 3 or more. It is preferable that the radius r is 5 or more, and more preferable that the radius r is 5 or more. The element that is the first to occur is counted as 0, and the number of elements counted from that element in the chain.

[0058] In addition, when selecting a fingerprinting method to be used, as mentioned above, When expressing the molecular structure of each organic compound, we tried to reduce the number of organic compounds that are all written differently. It is preferable to select one of both.

[0059] Fingerprints are trained by increasing the bit length (number of bits) they represent. This reduces the possibility of generating completely identical descriptions between organic compounds. If the bit length is too large, the computational costs and database management costs will increase. On the other hand, multiple fingerprints can be used simultaneously to express By doing so, it is possible to find multiple molecular structures that match the notation of a given fingerprint type. However, by combining different fingerprint types, the overall notation is perfectly matched. As a result, it is necessary to use a fingerprint with as small a bit length as possible. It is possible to generate a state in which multiple organic compounds with the same molecular structure are not generated. Since features are extracted using multiple methods, learning is efficient and overfitting is unlikely to occur. There is no particular restriction on the bit length of the fingerprint to be used, but it may be difficult to determine the computational cost or database size. Considering the management cost, if each molecular weight is up to about 2000, the finger print is The bit length for each type of input must be 4096 or less, preferably 2048 or less, and in some cases 102 4 or less, the fingerprints between molecules do not match perfectly, and the learning efficiency is A good fingerprint can be generated.

[0060] In addition, the bit length of the fingerprint generated by each fingerprint type is It is not necessary to unify the methods, but it is possible to adjust them appropriately by considering the characteristics of the type and the overall molecular structure to be studied. For example, the bit length is 1024 bits for Atom Pair type and 1024 bits for Circular The type may be expressed as 2048 bits and then concatenated.

[0061] Any machine learning method can be used, but a neural network is It is preferable to use a neural network with a structure as shown in FIG. The programming language can be Python, for example, and the machine learning framework can be Chainer or similar can be used for the framework. To achieve this, it is necessary to use a portion of the physical property data for testing and the rest for learning. stomach.

[0062] Examples of physical properties that are learned in association with molecular structure include emission spectrum, half-width, emission Light energy, excitation spectrum, absorption spectrum, transmission spectrum, reflection spectrum, absorption coefficient, excitation energy, transient emission lifetime, transient absorption lifetime, S1 level, T1 level, Sn Level, Tn level, Stokes shift value, luminescence quantum yield, oscillator strength, oxidation potential, reduction potential , HOMO level, LUMO level, glass transition point, melting point, crystallization temperature, decomposition temperature, boiling point, Fuel temperature, carrier mobility, refractive index, orientation parameter, mass-to-charge ratio and NMR measurements Spectra, chemical shift values ​​and their element numbers or coupling constants and ESR measurements Examples of the spectrum, g factor, D value, or E value in the determination can be given.

[0063] These may be obtained by measurement or by simulation. The measurement object can be selected from among solutions, thin films, powders, etc. It is preferable to learn the physical property values ​​obtained under the measurement conditions, simulation conditions, and units. If the conditions cannot be unified, some of the training data (at least two types of compounds) should be , preferably 1% or more, more preferably 3% or more) and the same compound under each measurement condition. Measure or simulate the physical properties of the It is preferable to learn the correlation of values. And to learn the information of the condition itself. It is preferable to incorporate it into the data at the same time.

[0064] The physical property values ​​to be learned and predicted may be one or more. There is a correlation between the physical property values. In this case, learning multiple types of physical properties at the same time will increase the learning efficiency and improve the prediction accuracy. Even if there is no correlation or only a low correlation between the physical properties, multiple physical properties can be measured simultaneously. Predictable, efficient and desirable.

[0065] Physical properties that are effective to learn in combination are those based on the same or similar characteristics. For example, physical properties related to optical properties, chemical properties, electrical properties, etc. It is advisable to have the object learn by appropriately combining physical property values ​​belonging to the physical property values ​​related to optical properties. The physical properties related to the characteristics include absorption peak, absorption edge, molar absorption coefficient, emission peak, emission spectrometer, etc. For example, the emission peak of a solution and that of a thin film are The emission peaks, the emission peaks measured at room temperature and at low temperatures, and the simulation The S1 level (lowest singlet excited level), T1 level (lowest triplet excited level), and Sn level obtained by the These include the Tn level (higher singlet excited level) and the Tn level (higher triplet excited level). It is preferable to combine two or more of these for learning.

[0066] The physical properties to be learned and predicted can be selected appropriately. For example, for organic EL elements, The physical properties obtained by the measurement method or simulation are preferable. This article will explain the above.

[0067] The emission spectrum is learned by calculating the emission intensity for each wavelength in a fixed wavelength range. In this case, the absolute value may be used, but it is better to normalize the maximum value in order to obtain a better spectrum. If you want to compare absolute values, use the maximum intensity or luminescence quantum yield as appropriate. should be written in parallel.

[0068] Measurements have been made in the form of solutions, thin films, powders, etc. The solution values ​​are It is preferable to predict the emission color of the host. (The difference in the relative dielectric constant between the solvent and the actual device is preferably within 10, and preferably in absolute value.) It is preferable to measure in a solvent (preferably within about 5). For example, the solvent is toluene. In the case of a solution, there is no intermolecular interaction. Similarly, the concentration is approximately 10 -4 ~10 -6 M is preferred. Even a thin film is preferable for predicting the emission color of the dopant. In this case, the doping concentration is also related to the element. The same is preferable, and about 0.5w% to 30w% is preferable. Also, the emission spectrum is There are fluorescent and phosphorescent spectra. Phosphorescent spectra are produced by the emission of heavy atoms such as iridium complexes. If the device is equipped with a liquid nitrogen gas, the measurement can be performed at room temperature in an oxygen-free state. The temperature can be kept low (100K to 10K) using liquid helium or other materials, and measurements can be made. The spectrum can be measured using a fluorescence spectrophotometer. This is the spectral width when the intensity is half of the original intensity.

[0069] The luminous energy is learned to have a value that suits the purpose. If there are multiple maximum values, for example, organic E In predicting the emission color of a dopant in an L element, it is preferable to obtain the maximum intensity value. It is preferable that the host material and the carrier transport layer have a maximum energy on the shortest wavelength side. Also, the rising value on the short wavelength side (70 to 50% of the maximum intensity on the shortest wavelength side) is plotted. The value of the intersection of the tangent and the baseline at the point where the line crosses the line may be used. Alternatively, the tangent line may be drawn at the point where the line is maximum.

[0070] Absorption, transmission, and reflection spectra are wavelength-specific in a fixed wavelength range. All you need to do is calculate the absorbance, absorption rate, transmittance, and reflectance of the object and let it learn those values. You can learn using absolute or normalized values, and if you want to compare spectral shapes, If you want to compare absolute values, you can use the absolute values. If the conditions such as concentration and film thickness are not standardized, the absolute values ​​of those conditions and intensity For example, the effect of light extraction efficiency in an organic EL element is When making a prediction, it is preferable to learn the transmittance and the thickness of the thin film in parallel. In order to predict the efficiency of energy transfer from the host to the dopant in an organic electroluminescence device, It is preferable to use the molar absorption coefficient of the dopant as the intensity. It can be measured.

[0071] The excitation energy can be determined from the absorption spectrum. You can learn the wavelengths where the absorption peaks occur and their intensities, or the intensities at any wavelength. For example, in the plot of 70-50% of the absorption maximum intensity on the longest wavelength side, The absorption peak at the longest wavelength can be calculated from the intersection point of the tangent line at the The tangent line at the point where the derivative (negative value) is minimum on the curve where absorption decays from high to low You may also subtract.

[0072] The Stokes shift value can be determined by the difference between the maximum excitation wavelength and the maximum emission wavelength. It may be the difference between the absorption wavelength and the maximum emission wavelength. For example, in the case of a luminescent material, the Stokes shift value It is preferable to learn the energy (eV) of the electron. The smaller this value, the shorter the time from excitation to emission. It is believed that the structural relaxation at this temperature is small, and therefore the luminescence quantum yield is high.

[0073] The transient emission lifetime is the time (lifetime) for the emission intensity to decay when a pulsed excitation light is irradiated onto a sample. In this case, the emission intensity for each time period in a certain time range and the It is advisable to learn the life value appropriately. In the case of waveforms, it is preferable to normalize them. The initial integrated intensity may be normalized, and the intensity of each wavelength may be a relative value. For example, In this case, the faster the decay (the shorter the lifetime), the higher the luminescence quantum yield. It can be measured using a fluorescence (luminescence) lifetime measuring device. In this case, electrical excitation may be used instead of optical excitation. Alternatively, a pressure may be applied and the time it takes for the emission intensity to decay (lifetime) may be measured. The time it takes for the emission intensity to decrease to 1 / e is usually used as an indicator of the time (lifetime) that the emission This is often the case.

[0074] The S1 level is the absorption edge of the absorption spectrum, the maximum on the long wavelength side, and the maximum of the excitation spectrum. It can be determined from the value, the maximum value of the emission spectrum, and the rising edge value on the short wavelength side. The T1 level is the absorption edge of the absorption spectrum obtained by transient absorption measurement, the maximum value on the long wavelength side, The maximum value of the phosphorescence spectrum, the peak wavelength on the short wavelength side of the phosphorescence spectrum, and the rise It can be calculated from the value of the rise of the absorption edge or the emission spectrum. The method for determining is as described above. The S1 and T1 levels are also obtained from simulations. For example, density functionals such as Gaussian in quantum chemistry calculation programs can be obtained. After performing a numerical optimization of the ground state (S0) structure, the excitation energy was calculated using the time-dependent density functional method. Similarly, the Sn level (singlet level above S1) and the Tn level The transition probability (the triplet level above T1) can also be calculated. For example, in the case of a luminescent material, the higher the oscillator strength, the better the quasi-luminescence. It is believed that this position is favorable for emitting light. In addition, the structure optimization of S0 obtained by density functional theory The difference between the potential energy obtained by the optimization of the structure of T1 and the potential energy obtained by the optimization of the structure of T2 is , it may be the T1 level.

[0075] The luminescence quantum yield can be determined by an absolute quantum yield measurement device.

[0076] Oxidation and reduction potentials can be measured by cyclic voltammetry (CV) The HOMO and LUMO levels are also related to the oxidation / reduction potential energies (eV ) is based on the redox potential of a standard sample (e.g., ferrocene) whose redox potential is known. On the other hand, the HOMO level can be measured in solid (thin film or powder) and exposed to air. It can also be measured by photoelectron spectroscopy (PESA). In this case, the LUMO is determined by the absorption spectrum. The band gap was calculated from the absorption edge of the PESA and the HOMO level was calculated using the PESA method. For example, in the case of an organic EL element, there is an excitation between two molecules. To estimate the emission energy when a complex occurs, the higher HOMO level (H The HOMO level of a molecule with a shallow OMO level and the LUMO level of a molecule with a small LUMO level The energy difference between the HOMO level obtained by CV and the other molecule (the deeper one) is calculated. It is preferable to use the LUMO level and the quantum chemical calculation program Gaussia n, the HOMO level and LUMO level, and the HOMO-n level (HOMO The occupied orbital levels below LUMO and LUMO+n (unoccupied orbital levels above LUMO) are calculated. It is possible.

[0077] The glass transition point, melting point, and crystallization temperature can be determined using a differential scanning calorimetry (DSC) device. It is preferable to measure the temperature at a constant rate of 10 to 50°C / min. The viscosity, boiling point, and sublimation temperature can be measured using a thermogravimetric-differential thermal analyzer (TG-DTA) device. It is advisable to use the results measured at atmospheric pressure or reduced pressure as appropriate. Values ​​measured under reduced pressure are used for sublimation purification. The temperature and deposition temperature can be used as a reference, and the value at which the weight is reduced by about 5-20% should be used. It is preferable to perform the measurement while keeping the temperature rise rate constant at 10 to 50° C. / min.

[0078] The carrier mobility was determined by the time-of-flight (TOF) method using transient photocurrent. In the TOF method, the sample film is sandwiched between electrodes and a direct current voltage is applied. In this state, carriers are generated by pulsed light excitation, and the transit time of the generated carriers (current transient In this case, the film thickness is preferably 3 μm or more. Alternatively, the current-voltage characteristics of the sample film can be determined by the space charge limited current (SCLC) If the current-voltage characteristic follows the equation of SCLC, the following can be obtained by fitting the current-voltage characteristic to the SCLC equation: The mobility can be obtained by impedance spectroscopy. A method for determining the mobility from the frequency dependence of the charge or capacitance has also been reported. In either method, the mobility at a certain voltage (electric field strength) can be calculated. This can be used as a physical property value. In addition, the mobility dependence on electric field strength can be plotted. By extrapolating, the mobility μ0 in the absence of an electric field can be calculated, and this is used as a physical property value. You can do that.

[0079] The refractive index and orientation parameters can be determined using a spectroscopic ellipsometry device. In the case of an organic EL element, a lower refractive index in the visible range is preferable because it improves the light extraction efficiency. There are several reports on the orientation parameters. For example, in the case of organic EL elements, The parameter S is often used. The orientation parameter S is determined by spectroscopic ellipsometry. It can be calculated by measuring the absorption anisotropy. The wavelengths corresponding to the absorption due to the phase (S1) where S is closer to -0.5 are the light extraction surfaces such as substrates. The transition dipole moment is considered to be more horizontal, resulting in higher light extraction efficiency. In the case of phosphorescent materials, it is preferable to focus on the absorption of the lowest triplet excited state (T1). When S is 0, the orientation is random, and when S is 1, the orientation is perpendicular. When the transition dipole moment is divided into a component horizontal to the substrate and a component perpendicular to the substrate, The percentage of the vertical component may also be used. This parameter is determined by the photoluminescence (PL ) or electroluminescence (EL), and the angular dependence of p-polarized light intensity was investigated. It can be found by fitting.

[0080] The mass-to-charge ratio (m / z) is calculated by calculating the detection intensity per unit within a certain fixed range of mass-to-charge ratio numbers. Depending on the purpose, it is possible to learn the absolute values ​​or normalized values. If you want to compare the spectral shapes, normalize them to an arbitrary wavelength, such as the m / z of the parent ion. If you want to compare absolute values, you can train the absolute values. m / z is , and can be measured using a mass spectrometer. The ionization method is electron ionization, chemical ionization, Electrochemical ionization, fast atom bombardment, matrix-assisted laser desorption / ionization, electrospinning There are various methods, such as pre-ionization, atmospheric pressure chemical ionization, and inductively coupled plasma. The child (parent molecule) decomposes (dissociates bonds) and the fragments (daughter ions) are detected at the same time. The detected m / z and the detection intensity ratio to the parent ion indicate the characteristics of the molecule. For example, molecules with the same substituents will have the same m / z fragments. Therefore, the m / z of the parent ion and the fragment and their detection intensity ratios are studied. By training the system, it is possible to predict the m / z of fragments of other compounds and the detection intensity ratio with respect to the parent ion. Generally, when the ionization energy is strong, the fragment generation ratio The rate will be higher.

[0081] NMR (Nuclear Magnetic Resonance) spectrum is a measurement of chemical shifts in a fixed chemical shift range. The signal intensity for each value can be calculated and learned as a value. The integral value of the intensity (number of elements), J value (coupling constant), etc. are displayed in parallel. In this case, it is preferable to express the sum of the integral values ​​of the numerators as the number of elements to be measured. NMR measurements can analyze the molecular structure of a substance at the atomic level. For example, molecules with the same substituents tend to show similar spectra at similar chemical shift values. The spectrum can be measured using an NMR device.

[0082] ESR (Electron Spin Resonance) spectra are obtained by measuring the magnetic field strength over a fixed range or magnetic flux density (Tes The detection intensity for each unit in the range and rotation angle can be calculated and learned as a value. Also, the g value ( It can also be expressed by g factor, square of g value, spin amount, spin density, etc. When a sample containing paired electrons is placed in a magnetic field, the spin transition of the unpaired electrons causes the absorption of microwaves. Therefore, ESR is used to measure paramagnetic substances that have unpaired electrons. It is effective for observing triplet states, for example, at low temperatures (100K to If ESR measurements are performed while irradiating excitation light at 10 K, the spin state information of the triplet excited state can be obtained. In this case, the D value (a quantity that represents the magnitude of the interaction between the two electron spins) It can also be expressed in terms of the E value (the amount by which the electron orbit deviates from axial symmetry). The spectrum can be measured with an ESR instrument.

[0083] After the learning stage is completed, the molecular structure of the input target substance is calculated based on the learned results. The target physical property values ​​are predicted from the structure (S102).

[0084] Finally, the predicted physical property values ​​are output (S103).

[0085] In this manner, one embodiment of the present invention is capable of predicting various physical properties, and is capable of predicting the molecular weight of an organic compound. Using multiple fingerprints to learn structures allows for more accurate predictions This is a method for predicting the physical properties of organic compounds.

[0086] (Embodiment 2) In embodiment 2, a system for predicting physical properties of an organic compound according to one embodiment of the present invention will be described. do. <Configuration example> The physical property prediction system 10 according to one embodiment of the present invention includes an input means, a learning means, a prediction means, an output means, and The system has at least a data server and a data server. These can exchange data with each other. They may be integrated into one device, or they may be separate devices. Some of them may be built into the same device, and the data server may be a cloud. However, these are collectively referred to as property prediction systems.

[0087] FIG. 8 shows a system having an input unit, a learning unit, a prediction unit, and an output unit as one embodiment of the present invention. The following explains the properties prediction system, which consists of an information terminal and a data server. The terminal 20 has an input unit, a learning means, a prediction means, and an output unit, and is connected to a data server provided separately. Data can be exchanged with the server.

[0088] The information terminal 20 mainly comprises an input unit 21, a calculation unit 22, and an output unit 25. The calculation unit 22 simultaneously serves as a learning means and a prediction means. The data provided by the data server is preferably a neural network. This data is used for learning or prediction by the rule network circuit 26. By using the trained learning method as both validation data and training data, It is possible to update the weight coefficients in the neural network circuit and generate learned weight coefficients. This can improve the accuracy of predictions.

[0089] In FIG. 8, the signal flow is shown in the order of the input unit 21, the calculation unit 22, the data server 30, and the output unit 25. In this specification, the term "signal" may be interpreted as "data" or "information" as appropriate. It is possible.

[0090] The data server 30 inputs the structure and physical properties of the organic compound to be learned into the learning means of the calculation unit 22. The structure of the provided organic compound is represented using two or more fingerprints. The learning means of the calculation unit 22 may have a neural network circuit. preferable.

[0091] The input unit 21 has a function for a user to input information. keyboard, mouse, touch panel, pen tablet, microphone, camera, etc. Any input means can be used.

[0092] Input information D in is the data output from the input unit 21 to the calculation unit 22. i n is information input by the user. For example, when the input unit 21 is a touch panel, is information obtained by inputting characters by operating a touch panel. Alternatively, the input unit 21 In the case of a microphone, the information is obtained by voice input by the user. In the case of a camera, the information is obtained by processing captured image data.

[0093] Input information D in is information about the structure of the organic compound whose physical properties you want to predict. If the structure image, substance name, or other notation other than the fingerprint notation is entered, The signal is then input to the prediction means in the calculation unit 22 via an appropriate conversion means. Based on the results of learning performed by the learning means, predictions are made for the properties of the input organic compounds. Do the following.

[0094] The result of the prediction is output via the output unit.

[0095] In addition, when the calculation unit has a neural network circuit, the neural network circuit It is preferable that the circuit has a multiply-and-accumulate circuit capable of executing multiply-and-accumulate processing. The path preferably has a memory circuit for storing the weight data. The memory element includes a transistor and a capacitor. The transistor has a channel forming The semiconductor layer having the region is provided with an oxide semiconductor. It is preferable that the OS transistor is an OS transistor. Therefore, the leakage current flowing in the off-state of the OS transistor is extremely small. By utilizing the property of being able to hold an electric charge by changing the state of the material, data can be stored. The configuration of the neural network circuit will be described in detail in the third embodiment.

[0096] In addition, a fingerprint that is concatenated or parallelized using multiple fingerprint types is created. A control program and control software that can generate lint, perform machine learning, and predict physical properties is available. A recorded recording medium is also one aspect of the present invention.

[0097] (Embodiment 3) In this embodiment, the neural network circuit (hereinafter referred to as a semiconductor An example of the configuration of a semiconductor device that can be used in a semiconductor device (hereinafter referred to as a semiconductor device) will be described.

[0098] In this specification, a semiconductor device is a device that can function by utilizing semiconductor characteristics. In other words, it refers to a neural network circuit that uses transistors that utilize the properties of semiconductors. is a semiconductor device.

[0099] As shown in Figure 9(A), the neural network NN has an input layer IL, an output layer OL, and an intermediate layer OL. It can be composed of an input layer IL, an output layer OL, and an intermediate layer HL. Each layer has one or more neurons (units). A neural network having two or more hidden layers HL may have one or more hidden layers. The network can also be called DNN (deep neural network), and deep nucleus Learning using neural networks can also be called deep learning.

[0100] Input data is input to each neuron in the input layer IL, and previous data is input to each neuron in the hidden layer HL. The output signal of the neurons in the layer OL or the back layer is input, and each neuron in the output layer OL receives the output signal of the neurons in the front layer. Each neuron receives the output signal of all the neurons in the previous and next layers. It may be connected to all neurons (fully connected) or it may be connected to only a portion of neurons.

[0101] FIG. 9(B) shows an example of a computation by a neuron. Here, we consider a neuron N and The figure shows two neurons in the front layer that output signals to neuron N. Neuron N has The output of the neuron x1 and the output of the neuron in the previous layer x2 are input. In step N, the multiplication result (x1w1) of output x1 and weight w1 and the multiplication result (x2w1) of output x2 and weight w2 are After the sum of the results (x2w2), x1w1+x2w2, is calculated, the bias b is added if necessary. The value a is then multiplied by the activation function h. and the output signal y=h(a) is output from neuron N.

[0102] In this way, the operation of a neuron involves adding up the products of the outputs of the neurons in the previous layer and the weights. This multiplication and addition operation is called multiplication and addition (x1w1+x2w2 above). This may be done on software using a program, or on hardware. When the multiplication and accumulation is performed by hardware, a multiplication and accumulation circuit can be used. The multiplication and accumulation circuit may be a digital circuit or an analog circuit. When an analog circuit is used for the sum-of-products operation circuit, the circuit scale of the sum-of-products operation circuit can be reduced. The aim is to improve processing speed and reduce power consumption by reducing the number of memory accesses. can be done.

[0103] The multiply-and-accumulate circuit is a transistor that contains silicon (such as single crystal silicon) in the channel formation region. The channel forming region may be a silicon transistor. The transistors are made of oxide semiconductors (hereinafter referred to as OS transistors). In particular, since the off-state current of OS transistors is extremely small, The Si transistor and the OS transistor are suitable for use as transistors that constitute a memory. The sum-of-products circuit may be constructed by using both the 1st and 2nd transistors. A configuration example of the semiconductor device will be described.

[0104] <Example of semiconductor device configuration> FIG. 10 shows an example of the configuration of a semiconductor device MAC having a function for performing neural network calculations. The semiconductor device MAC stores first data corresponding to the connection strength (weight) between neurons. The first data corresponds to the input data, and the second data corresponds to the input data. The first data and the second data are analog data or multi-valued data (discrete data), The semiconductor device MAC can utilize the data obtained by the multiplication and accumulation operation. It has the function of converting by a property function.

[0105] The semiconductor device MAC includes a cell array CA, a current source circuit CS, a current mirror circuit CM, and a circuit WDD, circuit WLD, circuit CLD, offset circuit OFST, and activation function circuit ACT Has V.

[0106] The cell array CA has a plurality of memory cells MC and a plurality of memory cells MCref. In the cell array CA shown in FIG. 10, memory cells MC (MC [1,1] to [m,n]), and m memory cells MCref (MCref[1] to [ The memory cell MC has a function of storing first data. The memory cell MCref has a function of storing reference data used in the multiplication and accumulation operation. The reference data may be analog data or multi-valued data.

[0107] A memory cell MC[i,j] (i is an integer between 1 and m, and j is an integer between 1 and n) is arranged as It is connected to line WL[i], wiring RW[i], wiring WD[j], and wiring BL[j]. In addition, the memory cell MCref[i] is connected to the wiring WL[i], the wiring RW[i], the wiring WDr ef and the wiring BLref. Here, the memory cell MC[i,j] and the wiring BL [j] is the current flowing between MC[i,j] and the memory cell MCref[i] and the wiring The current flowing between BLref and I MCref[i] It is written as follows.

[0108] A specific example of the configuration of the memory cell MC and the memory cell MCref is shown in FIG. As representative examples, memory cells MC[1,1], [2,1] and memory cell MCref[1] , [2] is shown, but the other memory cells MC and memory cells MCref have similar configurations. The memory cell MC and the memory cell MCref can each be a transistor. Here, the transistors Tr11 and Tr12 and the capacitance element C11 are A case where the transistor Tr11 and the transistor Tr12 are n-channel transistors will be described.

[0109] In the memory cell MC, the gate of the transistor Tr11 is connected to the line WL, and the source One of the drains is connected to the gate of the transistor Tr12 and the first electrode of the capacitance element C11. The other of the source and drain is connected to the wiring WD. One of the source or drain of 12 is connected to the wiring BL, and the other of the source or drain is connected to the wiring A second electrode of the capacitance element C11 is connected to the line VR. A second electrode of the capacitance element C11 is connected to the line RW. The wiring VR has a function of supplying a predetermined potential. A case where a low power supply potential (such as a ground potential) is supplied from the VR will be described.

[0110] One of the source and drain of the transistor Tr11, the gate of the transistor Tr12, and A node connected to the first electrode of the capacitor C11 and the second electrode of the capacitor C12 is referred to as a node NM. The nodes NM in the cells MC[1,1] and MC[2,1] are called nodes NM[1,1] and NM[2 ,1].

[0111] The memory cell MCref has a similar configuration to the memory cell MC. Cref is connected to the wiring WDref instead of the wiring WD, and the wiring BL is connected to the wiring BL ref. Also, in memory cells MCref[1] and [2], One of the source or drain of the transistor Tr11, the gate of the transistor Tr12, and a capacitance The nodes connected to the first electrode of the element C11 are called nodes NMref[1] and NMref[2], respectively. ] is written as

[0112] The nodes NM and NMref are the storage nodes of the memory cells MC and MCref, respectively. The node NM holds the first data, and the node NMref holds the second data. The reference data is stored in the memory cell MC[1,1], [2 ,1] transistor Tr12 has a current I MC[1,1] , I MC[2,1] Also, the transistors of the memory cells MCref[1] and [2] are connected to the wiring BLref. The current I MCref[1] , I MCref[2] flows.

[0113] The transistor Tr11 has a function of holding the potential of the node NM or the node NMref. Therefore, it is preferable that the off-state current of the transistor Tr11 is small. It is preferable to use an OS transistor with extremely low off-state current as the transistor Tr11. This makes it possible to suppress fluctuations in the potential of the node NM or the node NMref, improving the accuracy of the calculation. In addition, the potential of the node NM or the node NMref can be refreshed. This makes it possible to keep the frequency of flash operations low, thereby reducing power consumption.

[0114] The transistor Tr12 is not particularly limited and may be, for example, a Si transistor or an OS transistor. When an OS transistor is used for the transistor Tr12, Transistor Tr12 can be manufactured using the same manufacturing equipment as transistor Tr11. This allows the manufacturing cost to be reduced. The insulating layer 14 may be of either a p-channel type or a p-channel type.

[0115] The current source circuit CS is connected to the wirings BL[1] to BL[n] and the wiring BLref. The current source circuit CS has the function of supplying current to the wirings BL[1] to [n] and the wiring BLref. Note that the current values ​​supplied to the wirings BL[1] to BL[n] and the current value supplied to the wiring BLref are The current values ​​may be different. Here, the current source circuit CS is connected to the wirings BL[1] to BL[ n] is I C The current supplied from the current source circuit CS to the wiring BLref is I Cref It is written as follows.

[0116] The current mirror circuit CM has wiring IL[1] to IL[n] and a wiring ILref. The lines IL[1] to IL[n] are connected to the wirings BL[1] to BL[n], respectively, and the wiring ILre f is connected to the wiring BLref. Here, the wirings IL[1] to IL[n] and the wiring B The connection points of L[1] to L[n] are represented as nodes NP[1] to NP[n]. The connection point between Lref and the wiring BLref is denoted as node NPref.

[0117] The current mirror circuit CM generates a current I CM Wiring ILre f and this current I CM It also has the function of passing the signal through the wiring IL[1] to IL[n]. In 10, a current I CM is discharged, and wiring BL[1] A current I flows from the wiring IL[1] to [n] CM This shows an example of the discharge of In addition, current flows from the current mirror circuit CM to the cell array CA via the wiring BL[1] to [n]. The current that flows through the B The current mirror circuit CM is connected to the The current flowing through the cell array CA via BLref is I Bref It is written as follows.

[0118] The circuit WDD is connected to the wirings WD[1] to WD[n] and the wiring WDref. DD supplies a potential corresponding to the first data stored in the memory cell MC to the wiring WD[1]. The circuit WDD has the function of supplying the data stored in the memory cell MCref to the The circuit WLD has a function of supplying a potential corresponding to the reference data to the wiring WDref. The circuit WLD is connected to the wirings WL[1] to WL[m]. A signal for selecting the memory cell MC or memory cell MCref is applied to the wiring WL[1]. The circuit CLD is connected to the wirings RW[1] to RW[m]. The circuit CLD supplies a potential corresponding to the second data to the wirings RW[1] to RW[m]. It has the function of supplying

[0119] The offset circuit OFST includes wirings BL[1] to [n] and wirings OL[1] to [n]. The offset circuit OFST is connected to the offset terminals BL[1] to BL[n]. The amount of current flowing in the circuit OFST and / or the amount of offset current from the wirings BL[1] to BL[n] The offset circuit OFST has a function of detecting the change in the current flowing through the offset circuit OFST. T has a function of outputting the detection result to wiring OL[1] to [n]. The circuit OFST may output a current corresponding to the detection result to the wiring OL, or may The current corresponding to the cell array CA may be converted to a voltage and output to the wiring OL. The current flowing between the OFST and I α These are written as [1] through [n].

[0120] An example of the configuration of the offset circuit OFST is shown in FIG. has circuits OC[1] to OC[n]. Also, the circuits OC[1] to OC[n] each have Transistor Tr21, transistor Tr22, transistor Tr23, and capacitance element C21 The connections of the elements are as shown in FIG. A node connected to the first electrode of the element C21 and the first terminal of the resistor element R1 is referred to as a node N a. The second electrode of the capacitance element C21, the source or drain of the transistor Tr21 A node connected to one of the inputs and the gate of the transistor Tr22 is referred to as a node Nb. do.

[0121] The wiring VrefL has a function of supplying a potential Vref, and the wiring VaL supplies a potential Va. The wiring VbL has a function of supplying a potential Vb. The wiring VSSL has a function of supplying a potential VSS. Now, let us consider the case where the potential VDD is the high power supply potential and the potential VSS is the low power supply potential. The wiring RST supplies a potential for controlling the conduction state of the transistor Tr21. The transistor Tr22, the transistor Tr23, the wiring VDDL, the wiring The line VSSL and the wiring VbL form a source follower circuit.

[0122] Next, an example of the operation of the circuits OC[1] to OC[n] will be described. An example of the operation of OC[1] will be described, but circuits OC[2] to [n] can also be operated in the same manner. First, when a first current flows through the wiring BL[1], the potential of the node Na becomes equal to the first potential. The potential depends on the current and the resistance value of the resistor R1. The transistor Tr21 is in an on state, and the potential Va is supplied to the node Nb. It becomes the off state.

[0123] Next, when a second current flows through the wiring BL[1], the potential of the node Na increases by a factor of the second current and the resistance The potential changes according to the resistance value of element R1. At this time, transistor Tr21 is in the off state. Since node Nb is in a floating state, the potential of node Na changes. The potential of the node Nb changes due to capacitive coupling. Here, the change in the potential of the node Na is Δ V Na If the capacitance coupling coefficient is 1, the potential of node Nb is Va+ΔV Na It becomes. Then, the threshold voltage of the transistor Tr22 is V th Then, the potential from wiring OL[1] Va+ΔV Na -V th is output, where Va=V th By doing so, the wiring Potential ΔV from L[1] Na can be output.

[0124] Potential ΔV Na is the change in current from the first current to the second current, the resistance element R1, and the potential Vre f. Here, since the resistance element R1 and the potential Vref are known, the potential ΔV N a From this, the amount of change in the current flowing through the wiring BL can be obtained.

[0125] The amount of current and / or the change in current detected by the offset circuit OFST as described above. The signal corresponding to the amount is input to the activation function circuit ACTV via wiring OL[1] to [n]. will be done.

[0126] The activation function circuit ACTV includes wirings OL[1] to [n] and wirings NIL[1] to [ n]. The activation function circuit ACTV receives input from the offset circuit OFST. The function to perform calculations to transform the received signal according to a predefined activation function is The activation function can be, for example, a sigmoid function, a tanh function, or a softmax function. The activation function circuit ACTV can be used. The signal converted by is output to the wirings NIL[1] to [n] as output data. .

[0127] <Example of operation of semiconductor device> Using the semiconductor device MAC, a multiplication and accumulation operation of first data and second data can be performed. An example of the operation of the semiconductor device MAC when performing a product-sum operation will be described below.

[0128] FIG. 13 shows a timing chart of an operation example of the semiconductor device MAC. In the diagram, wiring WL[1], wiring WL[2], wiring WD[1], wiring WDref, and node NM [1,1], node NM[2,1], node NMref[1], node NMref[2] , the transition of the potential of the wire RW[1] and the wire RW[2], and the current I B [1]-I α [1], and current I Bref The graph shows the transition of the current I B [1]-I α [1] is wiring BL This corresponds to the sum of the currents flowing from [1] to memory cells MC[1,1] and [2,1].

[0129] As a representative example, memory cells MC[1,1], [2,1] and The operation will be explained focusing on memory cells MCref[1] and MCref[2]. The memory cell MCref can be operated in the same manner.

[0130] [Storage of the first data] First, at time T01-T02, the potential of the wiring WL[1] is at a high level (High). The potential of the wire WD[1] becomes V higher than the ground potential (GND). PR -V W[1,1] big The potential of the wiring WDref becomes V higher than the ground potential. PR The potential is large. The potential of the line RW[1] and the wiring RW[2] is the reference potential (REFP). W[1,1] is a potential corresponding to the first data stored in the memory cell MC[1,1]. In addition, the potential V PR is a potential corresponding to the reference data. [1,1] and the transistor Tr11 of the memory cell MCref[1] are turned on. The potential of node NM[1,1] becomes V PR -V W[1,1] , node NMref[1] The potential of V PR It becomes.

[0131] At this time, a current flows from the wiring BL[1] to the transistor Tr12 of the memory cell MC[1,1]. The current I MC[1,1],0 can be expressed as follows, where k is the number of transistors A constant determined by the channel length, channel width, mobility, and capacitance of the gate insulating film of Tr12 Also, Vth is the threshold voltage of transistor Tr12.

[0132] I MC[1,1],0 =k(V PR -V W[1,1] -V th ) 2 (E1)

[0133] Also, current flows from the wiring BLref to the transistor Tr12 of the memory cell MCref[1]. current I MCref[1],0 can be expressed by the following formula:

[0134] I MCref[1],0 =k(V PR -V th ) 2 (E2)

[0135] Next, at time T02-T03, the potential of the wiring WL[1] becomes low level (Low). As a result, the transistors in the memory cell MC[1,1] and the memory cell MCref[1] The transistor Tr11 is turned off, and the nodes NM[1,1] and NMref[1 ] potential is maintained.

[0136] As described above, it is preferable to use an OS transistor as the transistor Tr11. This makes it possible to suppress the leakage current of the transistor Tr11, and the node NM[ 1,1] and the potential of node NMref[1] can be accurately maintained.

[0137] Next, at time T03-T04, the potential of the wiring WL[2] becomes high level, and the potential of the wiring W The potential of D[1] is V higher than the ground potential. PR -V W[2,1] This results in a large potential, and the wiring WDr The potential of ef is V higher than the ground potential. PR The potential V W[2,1] Memo This potential corresponds to the first data stored in the memory cell MC[2,1]. The transistor Tr11 included in the memory cell MC[2,1] and the memory cell MCref[2] is turned on, and the potential of node NM[1,1] becomes V PR -V W[2,1] , node NM The potential of ref[1] is V PR It becomes.

[0138] At this time, a current flows from the wiring BL[1] to the transistor Tr12 of the memory cell MC[2,1]. The current I MC[2,1],0 can be expressed by the following formula:

[0139] I MC[2,1],0 =k(V PR -V W[2,1] -V th ) 2 (E3)

[0140] In addition, current flows from the wiring BLref to the transistor Tr12 of the memory cell MCref[2]. current I MCref[2],0 can be expressed by the following formula:

[0141] I MCref[2],0 =k(V PR -V th ) 2 (E4)

[0142] Next, at time T04-T05, the potential of the wiring WL[2] becomes low level. As a result, the transistors of the memory cell MC[2,1] and the memory cell MCref[2] Tr11 is turned off, and the potentials of nodes NM[2,1] and NMref[2] are Retained.

[0143] Through the above operations, the first data is stored in the memory cells MC[1,1] and MC[2,1]. Reference data is stored in memory cells MCref[1] and [2].

[0144] Here, at time T04-T05, the current flowing through the wiring BL[1] and the wiring BLref A current is supplied to the wiring BLref from the current source circuit CS. The current flowing through ref is the current mirror circuit CM and the memory cell MCref[1],[2] The current supplied from the current source circuit CS to the wiring BLref is I Cref ,wiring The current flowing from BLref to the current mirror circuit CM is I CM,0 Then, the following equation holds true.

[0145] I Cref -I CM,0 =I MCref[1],0 +I MCref[2],0 (E 5)

[0146] A current is supplied to the wiring BL[1] from the current source circuit CS. The current flowing through the current mirror circuit CM is drained to the memory cells MC[1,1] and MC[2,1]. In addition, a current flows from the wiring BL[1] to the offset circuit OFST. The current supplied from the circuit CS to the wire BL[1] is I C,0 , offset from wire BL[1] The current flowing through the circuit OFST is I α,0 Then, the following equation holds:

[0147] I C -I CM,0 =I MC[1,1],0 +I MC[2,1],0 +I α,0 (E 6)

[0148] [Multiply and add the first data and the second data] Next, at time T05-T06, the potential of the wiring RW[1] becomes V X[1] At this time, the memory cell MC[1,1] and the memory cell MCref[ Each capacitance element C11 in X[1] is supplied to the transformer through capacitive coupling. The potential of the gate of transistor Tr12 rises. Note that the potential V x[1] is the memory cell MC[1 , 1] and a potential corresponding to the second data supplied to the memory cell MCref[1]. .

[0149] The change in the potential of the gate of the transistor Tr12 is proportional to the change in the potential of the wiring RW. The capacitance coupling coefficient is multiplied by the capacitance element C1. 1, the gate capacitance of transistor Tr12, and the parasitic capacitance. For the sake of convenience, in the following, the change in the potential of the wiring RW and the change in the potential of the gate of the transistor Tr12 are In this explanation, we will assume that the capacitance is the same, i.e., the capacitive coupling coefficient is 1. In reality, the capacitive coupling coefficient is Considering the potential V x It is sufficient to determine the following.

[0150] A potential V X[1 ] When the voltages of the nodes NM[1] and NMref[1] are V X [1] to rise.

[0151] Here, at time T05-T06, the line BL[1] of the memory cell MC[1,1] Current I flowing through transistor Tr12 MC[1,1],1 can be expressed as follows: .

[0152] I MC[1,1],1 =k(V PR -VW[1,1] +V X[1] -V th ) 2 ( E7)

[0153] That is, the potential V X[1] By supplying The current flowing through the transistor Tr12 of the memory cell MC[1,1] is ΔI MC[1,1] =I MC[1,1],1 -I MC[1,1],0 Increase.

[0154] In addition, at time T05-T06, the line BLref of the memory cell MCref[1] Current I flowing through transistor Tr12 MCref[1],1 can be expressed by the following formula: do.

[0155] I MCref[1],1 =k(V PR +V X[1] -V th ) 2 (E8)

[0156] That is, the potential V X[1] By supplying The current flowing through the transistor Tr12 of the memory cell MCref[1] is ΔI MCref[ 1] =I MCref[1],1 -I MCref[1],0 Increase.

[0157] Next, consider the current flowing through the wiring BL[1] and the wiring BLref. The current I Cref Also, the current flowing through the wiring BLref The current is discharged to the current mirror circuit CM and memory cells MCref[1] and [2]. The current flowing from the wiring BLref to the current mirror circuit CM is I CM,1 Then, The following equation holds true.

[0158] I Cref -I CM,1 =I MCref[1],1 +I MCref[2],0 (E 9)

[0159] The wire BL[1] carries the current I C Also, wiring BL[1] The current flowing through the current mirror circuit CM is drained to the memory cells MC[1,1] and MC[2,1]. In addition, a current flows from the wiring BL[1] to the offset circuit OFST. The current flowing from BL[1] to the offset circuit OFST is I α,1 Then, the following formula is obtained: stand.

[0160] I C -I CM,1 =I MC[1,1],1 +I MC[2,1],1 +I α,1 (E 10)

[0161] Then, from equations (E1) to (E10), the current I α,0 and current I α,1 (Differential current ΔI α ) can be expressed by the following formula:

[0162] ΔI α =I α,0 -I α,1 =2kV W[1,1] V X[1] (E11)

[0163] Thus, the differential current ΔI α is the potential V W[1,1] and V X[1] The value depends on the product of .

[0164] After that, at time T06-T07, the potential of the wiring RW[1] becomes the ground potential, and the node The potentials of NM[1,1] and node NMref[1] are the same as those at times T04 and T05.

[0165] Next, at time T07-T08, the potential of the wiring RW[1] becomes V X[1] The potential of the wiring RW[2] becomes V higher than the reference potential. X[2] It becomes a large electric potential As a result, the memory cells MC[1,1] and MCref[1] The potential V X[1] is supplied to nodes NM[1,1] and and node NMref[1] are V X[1] Also, the memory cell MC The potential V X[ 2] is supplied, and the voltages of the nodes NM[2,1] and NMref[2] are increased by capacitive coupling. The 1st place is V X[2] to rise.

[0166] Here, at time T07-T08, the line BL[1] of the memory cell MC[2,1] Current I flowing through transistor Tr12 MC[2,1],1 can be expressed as follows: .

[0167] I MC[2,1],1 =k(V PR -V W[2,1] +V X[2] -V th ) 2 ( E12)

[0168] That is, the potential V X[2] By supplying The current flowing through the transistor Tr12 of the memory cell MC[2,1] is ΔI MC[2,1] =I MC[2,1],1 -I MC[2,1],0 Increase.

[0169] In addition, at time T05-T06, the line BLref of the memory cell MCref[2] Current I flowing through transistor Tr12 MCref[2],1 can be expressed by the following formula: do.

[0170] I MCref[2],1 =k(V PR +V X[2] -V th ) 2 (E13)

[0171] That is, the potential V X[2] By supplying The current flowing through transistor Tr12 of memory cell MCref[2] is ΔI MCref[ 2] =I MCref[2],1 -I MCref[2],0 Increase.

[0172] Next, consider the current flowing through the wiring BL[1] and the wiring BLref. The current I Cref Also, the current flowing through the wiring BLref The current is discharged to the current mirror circuit CM and memory cells MCref[1] and [2]. The current flowing from the wiring BLref to the current mirror circuit CM is I CM,2 Then, The following equation holds true.

[0173] I Cref -I CM,2 =I MCref[1],1 +I MCref[2],1 (E 14)

[0174] The wire BL[1] carries the current I C Also, wiring BL[1] The current flowing through the current mirror circuit CM is drained to the memory cells MC[1,1] and MC[2,1]. In addition, a current flows from the wiring BL[1] to the offset circuit OFST. The current flowing from BL[1] to the offset circuit OFST is I α,2 Then, the following formula is obtained: stand.

[0175] I C -I CM,2 =I MC[1,1],1 +I MC[2,1],1 +I α,2 (E 15)

[0176] From the formulas (E1) to (E8) and the formulas (E12) to (E15), the current I α,0 and current I α,2 Difference between the current (ΔI α ) can be expressed by the following formula:

[0177] ΔI α =I α,0 -I α,2 =2k(V W[1,1] V X[1] +V W[2,1] V X[ 2] ) (E16)

[0178] Thus, the differential current ΔI α is the potential V W[1,1] and potential V X[1] and the potential V W [2,1] and potential V X[2] The value is determined by adding the product of and .

[0179] After that, at time T08-T09, the potential of the wiring RW[1] and [2] becomes the ground potential. The potentials of the nodes NM[1,1], [2,1] and the nodes NMref[1], [2] are It will be the same as T04-T05.

[0180] As shown in equations (E9) and (E16), the difference input to the offset circuit OFST is Minute current ΔI α is the potential V corresponding to the first data (weight) X and the second data (input data The potential V corresponding to W The value corresponds to the sum of the products of the above. In other words, the differential current ΔI α By measuring the offset of the first data and the second data by the offset circuit OFST, The result of the multiply-and-accumulate operation can be obtained.

[0181] In the above, memory cells MC[1,1], [2,1] and memory cell MCref[ We focused on [1] and [2], but the number of memory cells MC and memory cells MCref can be set arbitrarily. The number of rows m of the memory cells MC and the memory cells MCref is set to an arbitrary number. The differential current ΔIα in this case can be expressed by the following equation.

[0182] ΔI α =2kΣ i V W[i,1] V X[i] (E17)

[0183] In addition, by increasing the number of columns n of the memory cells MC and the memory cells MCref, The number of multiply-add operations performed can be increased.

[0184] As described above, by using the semiconductor device MAC, the multiplication of the first data and the second data can be performed. The memory cell MC and the memory cell MCref shown in FIG. By using the configuration shown in 1, a multiply-and-accumulate circuit can be constructed using a small number of transistors. Therefore, the circuit scale of the semiconductor device MAC can be reduced.

[0185] When the semiconductor device MAC is used for calculations in a neural network, the memory cell MC The number of rows m corresponds to the number of input data supplied to one neuron, and the number of columns of memory cells MC The number n can correspond to the number of neurons. For example, the hidden layer HL Consider a case where a multiply-and-accumulate operation is performed using a semiconductor device MAC in the memory cell The number of rows m of MC is the number of input data provided by the input layer IL (the number of neurons in the input layer IL). The number of columns n of the memory cells MC is set to the number of neurons in the hidden layer HL. can be done.

[0186] The structure of the neural network to which the semiconductor device MAC is applied is not particularly limited. For example, the semiconductor device MAC is a convolutional neural network (CNN), a recurrent neural network (RNN), Recurrent neural networks (RNNs), autoencoders, Boltzmann machines (restricted Boltzmann machines) It can also be used for other purposes, such as

[0187] As described above, by using the semiconductor device MAC, the product-sum operation of the neural network can be realized. Furthermore, the cell array CA includes the memory cells MC and the memory By using the resistor MCref, the calculation accuracy can be improved, the power consumption can be reduced, or the circuit size can be reduced. It is possible to provide an integrated circuit IC that can reduce the size. EXAMPLES

[0188] In this embodiment, an example of predicting the physical properties of an organic compound will be described in detail. The T1 level was selected as the physical property to be predicted in relation to the molecular structure of the The value of the T1 level is the emission peak on the short wavelength side in the phosphorescence spectrum obtained by low-temperature PL measurement. The value was calculated from the wavelength. The total number of data is 420, with 380 for training and 380 for testing. The validity of the predictive model was evaluated using 40 of these.

[0189] The open-source cheminformatics toolkit R was used to model molecular structures. We used DKit. RDKit creates fingerprints from the SMILES representation of the molecular structure. The fingerprinting method can convert the data into mathematical formula data. cular type and Atom Pair type were used.

[0190] The input values ​​for predicting physical properties are the formulas written in circular format only, Ato We used a formula that expresses the m Pair type alone, and also a formula that connects the two. The radius of the lar type is set to 4, and the path length of the Atom Pair type is set to 30. The bit length of the fingerprint is 2048. Note that the radius of the circular type and A The path length of the tom Pair type is the length of the element that is connected from the starting point, which is set to 0. This is the number of elements counted.

[0191] In addition, when the circular type is used alone, the formula of the 420 organic compounds is There were two pairs that were identical. On the other hand, Atom Pair type alone or Circula When the r-type and Atom Pair-type are combined, the formula can be expressed as It was confirmed that all of the above were different and not identical.

[0192] The machine learning method used was neural network. The programming language used was Python. We used hon as the machine learning framework and Chainer as the machine learning framework. The work structure has two hidden layers. The number of neurons in each layer is 2048 (C number of bits for single axial type or single Atom Pair type) or 4096 (Ci (number of bits for the concatenation of the rcular type and the Atom Pair type), the first hidden layer and The hidden layer was set to 500 and the output layer was set to 1. The activation function for the hidden layer was the ReLU function. Ta.

[0193] Machine learning was performed under the above conditions, and the mean squared error between the training data and the test data was estimated. The transition was calculated up to 500 learning times. The results are shown in Figure 14. Note that Figure 14(A) shows the Circu As a result of learning using formulas written in lar format only, Figure 14(B) shows the Atom Pai As a result of learning using a formula written in r-type only, Fig. 14(C) shows the circular type and These are the results of learning using formulas written by concatenating Atom Pair types.

[0194] From the above results, the fingerprints of Circular and Atom Pair types When formulas written in this format are concatenated and used together, they perform better than when each formula is used alone. The mean square error of the test data was also reduced, and the prediction accuracy of the T1 level was improved.

[0195] From the above, different substructures are generated for each fingerprint type, and the Since the presence / absence information can provide information about the entire molecular structure, different types of finger patterns can be A method for describing molecular structures using multiple Lint methods is effective for predicting physical properties using machine learning. It can be seen that.

[0196] In addition, in this case, when there are different compounds that are represented identically by one fingerprinting method, By concatenating other fingerprints, the resulting formula can be made different. It is easy to use only one fingerprint pattern, eliminating compounds with the same name. Rather than increasing the number of bits to 1, it is better to combine two or more fingerprints. However, the generated formulas are unlikely to be identical, and differences in compounds are expressed with as few bits as possible. This is preferable because it allows the computational load in machine learning to be kept small. [Explanation of symbols]

[0197] T01-T02: Time, T02-T03: Time, T03-T04: Time, T04-T05 :Time, T05-T06:Time, T06-T07:Time, T07-T08:Time, T08 -T09: time, Tr11: transistor, Tr12: transistor, Tr21: transistor Tr22: transistor, Tr23: transistor, 20: information terminal, 21: input input unit, 22: calculation unit, 25: output unit, 30: data server

Claims

1. A property prediction system having an information terminal and a data server, The information terminal has an input unit, an output unit, and a calculation unit, The calculation unit is A first step of generating a first formula using a circular type fingerprinting method for a molecular structure of an organic compound; A second step of generating a second formula using the Atom Pair type of fingerprinting method for the molecular structure of the organic compound; a third step of generating a third formula using the molecular structure of the organic compound by a fingerprinting method using SubstructureKeys; a fourth step of generating a fourth formula by concatenating the first formula, the second formula, and the third formula; A fifth step of learning a correlation between the fourth formula and physical properties; A sixth step of predicting a target physical property from the molecular structure of the target substance based on the result of the learning, A physical property prediction system, wherein the first step, the second step, and the third step are performed simultaneously.

2. In claim 1, The arithmetic unit has a memory element, the memory element includes a transistor and a capacitor; The transistor has an oxide semiconductor in a channel formation region.

Images