Material property prediction system

Abstract

Property prediction that allows anyone to easily and accurately predict the physical properties of organic compounds. This provides a method for anyone to easily and accurately predict the physical properties of organic compounds. We provide a system for predicting possible physical properties. [Solution] A step of learning the correlation between the molecular structure and physical properties of organic compounds, and the results of the learning The process includes the step of predicting the desired physical properties from the molecular structure of the target substance, and the organic compound Organic morphology is a method of representing the molecular structure of substances by simultaneously using multiple types of fingerprinting methods. This invention provides a method and system for predicting the physical properties of a compound.

Description

[Technical Field]

[0001] One aspect of the present invention relates to a method and apparatus for predicting the physical properties of organic compounds. [Background technology]

[0002] Historically, the physical properties of organic compounds could only be determined by synthesizing the target substance and directly measuring it. However, those properties are determined by the molecular structure of the organic compound in question. Therefore, the approximate value of the physical properties possessed by an organic compound with a certain molecular structure is In today's world, with all the accumulated data, experts can usually pinpoint what it indicates. This has become possible. In recent years, calculations have also been performed using first-principles simulation theory, etc. Prediction is also possible by doing so.

[0003] In research and development using organic compounds, the corresponding physical properties are determined according to the required characteristics. Organic compounds possessing these compounds are selected and used. Therefore, known compounds are used without actually synthesizing them. Accurately predicting, selecting, and using organic compounds with the required physical properties based on their characteristics and unknown properties. If this can be achieved, it is expected to significantly improve the speed of development.

[0004] However, accurate predictions like those mentioned above are not something everyone can do, and currently, simulations are... The process would incur enormous costs and time. On the other hand, candidate organic compounds are non Because they are always abundant, there is a method that anyone can easily and quickly use to predict the physical properties of a target organic compound. A calling system is desired.

[0005] In recent years, methods for classification, estimation, and prediction using machine learning and other techniques have undergone significant advancements. In particular, deep learning using convolutional neural networks is being used for selection. The performance of predictions and forecasts has improved significantly, and excellent results are being achieved in various fields. However, in the field of dealing with organic compounds, it is necessary to ensure that their structure is understood by a computer without discrepancies. Furthermore, it is possible to accurately extract features related to physical properties, and with a manageable amount of information. Currently, there are very few adequate methods for describing certain organic compounds. Therefore, a method for predicting the physical properties of organic compounds that anyone can easily and accurately predict. And the system has not yet been realized.

[0006] Patent Document 1 discloses a novel material discovery method and apparatus using machine learning. Yes, they are. [Prior art documents] [Patent Documents]

[0007] [Patent Document 1] Japanese Patent Publication No. 2017-91526 [Overview of the project] [Problems that the invention aims to solve]

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

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

[0010] Another aspect of the present invention includes the step of learning the correlation between the molecular structure and physical properties of an organic compound, The process includes a step of predicting the desired physical properties from the molecular structure of the target substance based on the results of the aforementioned learning. As a method for representing the molecular structure of the aforementioned organic compound, two types of fingerprinting methods are used simultaneously. This is a method for predicting the physical properties of organic compounds used.

[0011] Another aspect of the present invention includes the step of learning the correlation between the molecular structure and physical properties of an organic compound, The process includes the step of predicting the desired physical properties from the molecular structure of the target substance based on the results of the aforementioned learning. As a method for representing the molecular structure of the aforementioned organic compound, three types of fingerprinting methods are used simultaneously. This is a method for predicting the physical properties of organic compounds.

[0012] Furthermore, in another aspect of the present invention, in the above configuration, the fingerprinting method is At Pair type, Circular type, Substructure key type and P This is a material property prediction method that includes at least one of the ATH-based types.

[0013] Furthermore, in another aspect of the present invention, in the above configuration, the plurality of fingerprinting methods are Atom Pair type, Circular type, Substructure key type, and This is a method of predicting material properties selected from among path-based types.

[0014] Furthermore, in another aspect of the present invention, in the above configuration, the fingerprinting method is Ato This is a method for predicting physical properties, including Pair-type and Circular-type methods.

[0015] Furthermore, in another aspect of the present invention, in the above configuration, the fingerprinting method is Ci This is a method for predicting physical properties, including rcular and substructure key types.

[0016] Furthermore, in another aspect of the present invention, in the above configuration, the fingerprinting method is Ci This is a method for predicting physical properties, including rcular and path-based methods.

[0017] Furthermore, in another aspect of the present invention, in the above configuration, the fingerprinting method is At This is a method for predicting physical properties, including the om Pair type and Substructure key type. .

[0018] Furthermore, in another aspect of the present invention, in the above configuration, the fingerprinting method is At This is a method for predicting physical properties, including both pair-type and path-based methods.

[0019] Another aspect of the present invention is, in the above configuration, the fingerprinting method is A tom Pair type, Substructure key type, and Circular type This is a method for predicting physical properties.

[0020] Furthermore, in another aspect of the present invention, in the above configuration, the fingerprinting method is as described above. When the circular type is used, it is a material property prediction method where r is 3 or greater.

[0021] Furthermore, in another aspect of the present invention, in the above configuration, the circular type of the finger The print method is a method for predicting material properties where r is 5 or greater.

[0022] Furthermore, in another aspect of the present invention, in the above configuration, at least the fingerprinting method When the molecular structure of each organic compound used for learning is written using 1, the notation for each organic compound is These are all different methods for predicting material properties.

[0023] Furthermore, in another aspect of the present invention, in the above configuration, at least the fingerprinting method Method 1 is a material property prediction method that can represent structural information that characterizes the material properties to be predicted.

[0024] Furthermore, in another aspect of the present invention, in the above configuration, at least the fingerprinting method 1 is a substituent, the substitution position of the substituent, a functional group, the number of elements, the type of element, the valence of the element, and This is a method for predicting material properties that can represent at least one of the aggregate order and atomic coordinates.

[0025] Furthermore, in another aspect of the present invention, in the above configuration, the physical properties are emission spectrum, full width at half maximum. , emission energy, excitation spectrum, absorption spectrum, transmission spectrum, reflectance spectrum Molar extinction coefficient, excitation energy, transient emission lifetime, transient absorption lifetime, S1 level, T1 level, Sn level, Tn level, Stokes shift value, emission quantum yield, oscillator strength, oxidation potential, reduction Potential, HOMO level, LUMO level, glass transition temperature, melting point, crystallization temperature, decomposition temperature, boiling point For sublimation temperature, carrier mobility, refractive index, orientation parameters, mass-to-charge ratio, and NMR measurements. Spectra, chemical shift value and its elemental number or coupling constant, ESR measurement One or more of the following physical properties in the determination: spectrum, g-factor, D-value, or E-value. This is the measurement method.

[0026] Another aspect of the present invention comprises an input means, a data server, and data stored on the data server. A learning method for learning the correlation between the molecular structure and physical properties of organic compounds, and based on the results of the learning, Prediction means for predicting the desired physical properties from the molecular structure of the target substance input from the aforementioned input means. The system has an output means for outputting the predicted physical properties, and a table of the molecular structure of the organic compound. As a method of notation, a system for predicting the physical properties of organic compounds that simultaneously uses multiple types of fingerprinting methods. It is a stem.

[0027] Another aspect of the present invention includes an input means, a data server, and data stored on the data server. A learning method for learning the correlation between the molecular structure and physical properties of a given organic compound, and based on the results of the learning A predictive tool that predicts the desired physical properties from the molecular structure of the target substance input from the input means. The device has a stage and an output means for outputting the predicted physical properties, and the molecular structure of the organic compound As a notation method, a system for predicting the physical properties of organic compounds that simultaneously uses two types of fingerprinting methods. It is a stem.

[0028] Another aspect of the present invention includes an input means, a data server, and data stored on the data server. A learning method for learning the correlation between the molecular structure and physical properties of a given organic compound, and based on the results of the learning A predictive tool that predicts the desired physical properties from the molecular structure of the target substance input from the input means. The device has a stage and an output means for outputting the predicted physical properties, and the molecular structure of the organic compound As a notation method, the prediction of the physical properties of organic compounds uses three types of fingerprinting methods simultaneously. It is a stem.

[0029] Furthermore, in another aspect of the present invention, in the above configuration, the fingerprinting method is At Pair type, Circular type, Substructure key type and P This is a physical property prediction system that includes at least one ATH-based type.

[0030] Another aspect of the present invention is that in the above configuration, the plurality of fingerprinting methods are A tom Pair type, Circular type, Substructure key type and This is a material property prediction system selected from among path-based systems.

[0031] Furthermore, in another aspect of the present invention, in the above configuration, the fingerprinting method is Ato This is a material property prediction system that includes both Pair-type and Circular-type systems.

[0032] Furthermore, in another aspect of the present invention, in the above configuration, the fingerprinting method is Cir This is a material property prediction system that includes cular type and substructure key type. .

[0033] Furthermore, in another aspect of the present invention, in the above configuration, the fingerprinting method is Ci This is a material property prediction system that includes both rcular and path-based types.

[0034] Furthermore, in another aspect of the present invention, in the above configuration, the fingerprinting method is At Property prediction system including Pair type and / or Substructure key type It is a stem.

[0035] Furthermore, in another aspect of the present invention, in the above configuration, the fingerprinting method is At This is a material property prediction system that includes Pair-type and / or Path-based types. .

[0036] Another aspect of the present invention is, in the above configuration, the fingerprinting method is A tom Pair type, Substructure key type, and Circular type It is a system that predicts physical properties.

[0037] Furthermore, in another aspect of the present invention, in the above configuration, the fingerprinting method is as described above. When the circular type is used, it is a material property prediction system where r is 3 or greater.

[0038] Furthermore, in another aspect of the present invention, the circular type fin in the above configuration is The Garprint method is a material property prediction system where r is 5 or greater.

[0039] Furthermore, in another aspect of the present invention, in the above configuration, at least the fingerprinting method When the molecular structure of each organic compound used for learning is written using 1, the notation for each organic compound is These are all different material property prediction systems.

[0040] Furthermore, in another aspect of the present invention, in the above configuration, at least the fingerprinting method The first is a material property prediction system that can represent structural information that characterizes the material properties to be predicted. ru.

[0041] Furthermore, in another aspect of the present invention, in the above configuration, at least the fingerprinting method 1 is a substituent, the substitution position of the substituent, a functional group, the number of elements, the type of element, the valence of the element, and This is a material property prediction system capable of representing at least one of the aggregate order and atomic coordinates.

[0042] Furthermore, in another aspect of the present invention, in the above configuration, the physical properties are emission spectrum, full width at half maximum. , emission energy, excitation spectrum, absorption spectrum, transmission spectrum, reflectance spectrum Molar extinction coefficient, excitation energy, transient emission lifetime, transient absorption lifetime, S1 level, T1 level, Sn level, Tn level, Stokes shift value, emission quantum yield, oscillator strength, oxidation potential, reduction Potential, HOMO level, LUMO level, glass transition temperature, melting point, crystallization temperature, decomposition temperature, boiling point For sublimation temperature, carrier mobility, refractive index, orientation parameters, mass-to-charge ratio, and NMR measurements. Spectra, chemical shift value and its elemental number or coupling constant, ESR measurement One or more of the following physical properties in the determination: spectrum, g-factor, D-value, or E-value. It is a measurement system. [Effects of the Invention]

[0043] In one aspect of the present invention, the physical properties of an unknown organic compound can be easily and accurately predicted by anyone. This provides a method for predicting physical properties that allows anyone to predict the physical properties of organic compounds. This system can provide a material property prediction system that can easily and accurately predict physical properties. [Brief explanation of the drawing]

[0044] [Figure 1] A flowchart illustrating one aspect of the present invention. [Figure 2] A diagram illustrating the method of transforming molecular structures using the fingerprinting technique. [Figure 3] A diagram explaining the different types of fingerprinting methods. [Figure 4] A diagram illustrating the conversion from SMILES notation to fingerprint notation. [Figure 5] A diagram explaining the types of fingerprinting methods and the overlap in their notation. [Figure 6] A diagram illustrating examples of molecular structures represented using multiple fingerprinting methods. [Figure 7] A diagram illustrating the structure of a neural network. [Figure 8] A diagram illustrating a physical property prediction system according to one aspect of the present invention. [Figure 9] A diagram illustrating the structure of a neural network. [Figure 10] A diagram illustrating an example configuration of a semiconductor device that performs calculations. [Figure 11] A diagram illustrating a specific example of a memory cell configuration. [Figure 12] A diagram illustrating an example configuration of an offset circuit (OFST). [Figure 13] A diagram showing a timing chart for an example of semiconductor device operation. [Figure 14] A diagram showing the results of material property prediction. [Modes for carrying out the invention]

[0045] The embodiments of the present invention will be described in detail below with reference to the drawings. However, the present invention is as follows Not limited to the description, the form and details thereof may be described without departing from the spirit and scope of the present invention. Those skilled in the art will readily understand that the invention can be modified in various ways. Therefore, the present invention is as follows: This should not be interpreted as being limited to the contents described in the embodiments.

[0046] (Embodiment 1) One embodiment of the present invention, a method for predicting physical properties, can be illustrated, for example, by a flowchart like the one shown in Figure 1. According to Figure 1, first, one aspect of the present invention is a method for predicting physical properties, which involves the molecular structure of an organic compound and Learn about the correlations between material properties (S101).

[0047] In this case, to use machine learning to determine the correlation between molecular structure and physical properties, it is necessary to describe the molecular structure using mathematical formulas. Yes, it is. For formulating molecular structures mathematically, open-source cheminformatics toolkits are available. A certain RDKit can be used. With RDKit, the SMIL of the input molecular structure can be processed. ES notation (Simplified molecular input line ent The number of ry specification syntax (by fingerprinting) It can be converted into formula data.

[0048] In the fingerprinting method, for example, as shown in Figure 2, a partial structure (fragment) of the molecular structure is obtained. The molecular structure is represented by assigning (t) to each bit, and if a corresponding substructure exists in the molecule, If the action is performed, the bit is set to "1"; otherwise, it is set to "0". In other words, the fingerprint method. By using this, it is possible to obtain mathematical formulas that extract the characteristics of the molecular structure. Also, generally, The molecular structure formulas represented by the Wingerprint method have bit lengths ranging from several hundred to tens of thousands, making them easy to handle. It is of a certain size. Furthermore, in order to represent the molecular structure using mathematical formulas of 0s and 1s, the fingerprinting method is used. By using this method, it becomes possible to achieve extremely fast computational processing.

[0049] Furthermore, there are many types of fingerprinting methods (differences in bit generation algorithms, atomic data). Considering the type, bond type, and aromaticity conditions, a hash function is used to dynamically generate bits. There are various types (such as those that generate length), each with its own unique characteristics.

[0050] Typical types of fingerprinting methods include, as shown in Figure 3, 1) Circul ar type (a substructure consisting of surrounding atoms up to a specified radius, with the starting atom at the center), 2 )Path-based type (Path length from the starting atom to the specified path length) 3) Substructure keys type (with atoms up to gth) 4) Atom pair type (all substructures are defined for each bit), 4) Atom pair type (all substructures in the molecule) Examples include (using atomic pairs generated for each atom as a substructure). RDKit includes these Each type's fingerprint is implemented.

[0051] Figure 4 shows the molecular structure of a certain organic compound, represented mathematically using the fingerprinting method. This is an example. In this way, the molecular structure is first converted to SMILES notation and then fingered - It can be converted to a printout.

[0052] Furthermore, when representing the molecular structure of an organic compound using the fingerprinting method, similar structures are found to exist. In some cases, the resulting mathematical formulas may be identical across different organic compounds. There are several types of fingerprinting methods, depending on the notation, but they are the same. The tendency of compounds to become like this is shown in Figure 5, 1) Circular type (Morgan Fing erprint), 2) Path-based type (RDK Fingerprint), 3) Substructure keys type (Avalon Fingerprint) 4) As shown in Atom pair type (Hash atom pair), It varies depending on the law. Note that in Figure 5, the molecules within each double arrow are identical. The formula (notation) is shown. Therefore, as a fingerprinting method used for learning, it is At the very least, when writing the molecular structure of each organic compound to be trained using 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, but training Depending on the population of organic compounds, it may be possible to express them without duplication using other notation methods. ru.

[0053] In one aspect of the present invention, the organic compound to be learned is represented by the fingerprinting method. The method is characterized by the use of multiple different types of fingerprinting techniques. Any number of types is acceptable, but two or three types are preferable in terms of manageability in terms of data volume. When learning using multiple types of fingerprinting methods, one type of fingerprinting method The formula written by [this method] is followed by the formula written by [this method]. You can use them in combination, or multiple different mathematical formulas exist for each organic compound. It is acceptable to train the system assuming that it exists. Figure 6 shows how to differentiate using multiple fingerprints of different types. Here is an example of how to describe a child structure.

[0054] Fingerprinting is a method of describing the presence or absence of substructures, and information about the entire molecular structure is not lost. However, if you use multiple fingerprints of different types to mathematically represent the molecular structure... Then, different substructures are generated for each fingerprint type, and these substructures Information about the presence or absence of a certain feature can be used to supplement information related to the entire molecular structure. This refers to cases where characteristics that cannot be fully expressed significantly affect the physical properties, or where the physical properties of compounds with such characteristics differ. If it affects the value difference, it will be complemented by other fingerprints, so different types The method of describing molecular structure using multiple fingerprints is effective.

[0055] When using two types of fingerprinting methods for representation, the Atom Pair type and the other are used. Using a circular type allows for accurate prediction of material properties, making it a preferred configuration. be.

[0056] Furthermore, when using the three types of fingerprinting methods for notation, the Atom Pair type and Using Circular and Substructure Keys types provides higher precision. This configuration is preferable because it allows for the prediction of physical properties.

[0057] Furthermore, when using the circular fingerprinting method, the radius r must be 3 or greater. It is preferable that it be 5 or more. The radius r is the starting point This is the number of elements counted by linking them together, with the element that is currently at the beginning of the sequence being set to 0.

[0058] Furthermore, when selecting the fingerprinting method to use, as mentioned earlier, the training method When describing the molecular structure of each organic compound, we aim to minimize instances where the description of each organic compound is different. It is preferable to choose one of them.

[0059] Fingerprints are learned by increasing the bit length (number of bits) used for representation. This can reduce the likelihood of generating descriptions that perfectly match across different organic compounds. If the bit length is made too large, the computational cost and database management cost will increase significantly. A trade-off arises. On the other hand, it is possible to express using multiple fingerprints simultaneously. By doing so, there are multiple molecular structures whose notation is a perfect match for a certain fingerprint type. By combining different fingerprint shapes, the overall representation can be perfectly matched. There is a possibility that this will not happen. As a result, the fingerprint table will be made with the smallest possible bit length. It is possible to create a state in which no multiple organic compounds with a perfect match in the description are formed. Furthermore, the characteristics of the molecular structure Because features are extracted using multiple methods, learning efficiency is high and overfitting is less likely. There are no particular restrictions on the bit length of the fingerprint, but the computational cost and database Considering management costs, for molecules with molecular weights up to about 2000, finger preservatives are suitable. The bit length for each bit type is 4096 or less, preferably 2048 or less, and in some cases 102 Below 4, the intermolecular fingerprints do not perfectly match, and the learning efficiency is It is possible to generate a good fingerprint.

[0060] Furthermore, the bit length of the fingerprint generated for each fingerprint type is It is sufficient to adjust the settings as needed, taking into account the characteristics of the type and the overall molecular structure being studied; there is no need to unify them. For example, the bit length is 1024 bits for Atom Pair type, and Circular The data type can be represented using 2048 bits, and these can be concatenated.

[0061] Any machine learning method can be used, but a neural network is recommended. It is preferable to use it. Learning by neural networks is, for example, as shown in Figure 7. You can build a structure and do it. For the programming language, for example, Python, and the machine learning framework Chainer and similar tools can be used for the workflow. The validity of the predictive model can be evaluated. To achieve this, you can use some of the material property data for testing and the rest for training. stomach.

[0062] Examples of physical properties that can be learned in relation to molecular structure include emission spectra, full width at half maximum, and emission. Light energy, excitation spectrum, absorption spectrum, transmission spectrum, reflectance spectrum, mo Absorption coefficient, excitation energy, transient emission lifetime, transient absorption lifetime, S1 level, T1 level, Sn Level, Tn level, Stokes shift value, emission quantum yield, oscillator strength, oxidation potential, reduction potential HOMO level, LUMO level, glass transition temperature, melting point, crystallization temperature, decomposition temperature, boiling point, rising In measurements of nucleus temperature, carrier mobility, refractive index, orientation parameters, mass-to-charge ratio, and NMR, The spectrum, chemical shift value, and the number of elements or coupling constants and ESR measurement Examples of parameters that can be cited include the spectrum, g-factor, D-value, or E-value in the determination.

[0063] These can be determined by measurement or by simulation. This is also fine. The object to be measured can be appropriately selected from solutions, thin films, powders, etc. However, the same applies to each. It is preferable to train the model using physical property values ​​obtained under measurement conditions, simulation conditions, and units. If the conditions cannot be standardized, use some of the training data (at least two types of compounds). (preferably 1% or more, more preferably 3% or more) and the same compound under each measurement condition Measure or simulate the physical properties of and perform measurements and simulations under different conditions. It is preferable to enable learning of the correlation between values. And it is desirable to learn the information of the conditions themselves. It is preferable to incorporate this into the data simultaneously.

[0064] The physical properties to be learned and predicted can be one type or multiple types. There can be correlations between the physical properties. In this case, learning multiple types of physical properties simultaneously increases learning efficiency and predictive accuracy. Therefore, it is preferable. Even if there is no correlation or a low correlation between physical properties, multiple physical properties can be measured simultaneously. It is predictable, efficient, and desirable.

[0065] For material properties that are effective to learn in combination, use the same or similar characteristics as a basis. Examples of determined physical properties include physical properties related to optical properties, chemical properties, and electrical properties. It is best to combine and train the model with appropriate physical properties, such as those related to properties of the material. Physical properties related to characteristics include absorption peak, absorption edge, molar extinction coefficient, emission peak, and emission coefficient. Examples include the full width at half maximum of the vector and the emission quantum yield. For example, the emission peak of a solution and a thin film. Emission peaks, emission peaks measured at room temperature and emission peaks measured at low temperatures, simulations The S1 level (lowest singlet excitation level), T1 level (lowest triplet excitation level), and Sn level were determined using the method. Examples include the Tn level (a higher singlet excitation level) and the Tn level (a higher triplet excitation level). It is preferable to combine two or more of these selected methods for learning.

[0066] The physical properties to be learned and predicted can be selected as appropriate, but for organic EL elements, for example, the following: Physical properties obtained through measurement methods or simulations like the one described above are preferred. I will explain it.

[0067] The emission spectrum is learned by determining the emission intensity for each wavelength within a fixed wavelength range. This is sufficient. While absolute values ​​are acceptable, it is better to normalize the maximum and local maximum values. This is a favorable prediction. If you want to compare absolute values, you can use maximum intensity or luminescence quantum yield as appropriate. You can simply describe them in parallel.

[0068] Measurements have been taken in various states, including solution, thin film, and powder. The solution values ​​are based on the doping of organic EL elements. This is preferable for predicting the emission color of the pantograph. At this time, the polarity of the host used in the actual device should be (Preferably the difference in relative permittivity between the solvent and the actual device is within 10, preferably in absolute value) It is preferable to measure in a solvent (preferably within 5). Chromoform, dichloromethane, etc. are preferred. In the case of a solution, there are no intermolecular interactions. The concentration is approximately 10 -4 ~10 -6 M is preferable. Doping organic matter such as a host is preferable. Even thin films are preferable for predicting the emission color of the dopant. In this case, the dope concentration is also important for the device. Similar values ​​are preferred, and generally 0.5 w% to 30 w% is preferred. Furthermore, the emission spectrum includes: There are fluorescence spectra and phosphorescence spectra. Phosphorescence spectra are found in heavy atoms such as iridium complexes. If using this method, the measurement can be performed at room temperature under deoxygenated conditions. Otherwise, liquid nitrogen is required. The temperature can be lowered to 100K or 10K using liquid helium or similar materials for measurement. The luminescence intensity can be measured with a fluorescence spectrophotometer. The full width at half maximum (FWHM) is the maximum value of the emission intensity. This refers to the spectral width when the intensity is halved.

[0069] The luminescence energy is trained to learn a value that suits the purpose. If there are multiple maximum values, for example, organic E For predicting the emission color of a dopant in an L element, it is best to find the value with the highest intensity within that range. It seems so. The energy of the host material and carrier transport layer is the maximum value at the shortest wavelength. Also, the rise time on the short wavelength side (plotted at 70-50% of the maximum intensity on the shortest wavelength side) The value of the intersection of the tangent and the baseline is also acceptable. Alternatively, the derivative of the rising edge on the short wavelength side. You can also find the value by drawing a tangent line at the point where the value is maximized.

[0070] Absorption spectra, transmission spectra, and reflection spectra are obtained at each wavelength within a fixed wavelength range. You can learn the absorbance, absorption rate, transmittance, and reflectance of a given element by calculating these values. Depending on the purpose, You can learn using absolute or normalized values, and if you want to compare spectral shapes, You can simply train the model to use values ​​normalized by the desired wavelength. If you want to compare absolute values, keep them as absolute values. To train the model. If conditions such as concentration and film thickness are not standardized, the absolute values ​​of those conditions and the intensity are used. It is preferable to list them in parallel. For example, the effect of light extraction efficiency in an organic EL element, etc. If you want to make a prediction, it is preferable to learn the transmittance and film thickness of the thin film in parallel. Also, for example If you want to predict the energy transfer efficiency from host to dopant in an organic EL element, The degree is preferably determined using the molar extinction coefficient of the dopant. The spectrum is obtained using an absorbance spectrophotometer. It can be measured.

[0071] The excitation energy can be determined from the absorption spectrum. This involves determining the wavelength of the absorption edge and the absorbance. You can learn the wavelength and intensity at which the maximum value occurs, as well as the intensity at any given wavelength, as appropriate. One way to determine this is to plot the 70-50% of the absorption maximum intensity on the longest wavelength side. It can be determined from the value at the intersection of the tangent line and the baseline. Also, the absorption electrode on the longest wavelength side. In a curve where absorption decreases from large, the tangent line is at the point where its derivative (negative value) is minimized. You may subtract it.

[0072] The Stokes shift value can be determined by the difference between the maximum excitation wavelength and the maximum emission wavelength. The difference between the absorption wavelength and the maximum emission wavelength can also be used. For example, in the case of light-emitting materials, the Stokes shift value. It is preferable to train the system using energy (eV). The smaller this value, the longer it takes from excitation to emission. The structural relaxation at this stage is considered to be small, and therefore the luminescence quantum yield is thought to be high.

[0073] Transient emission lifetime is the time (lifetime) it takes for the emission intensity to decay when a sample is irradiated with pulsed excitation light. It can be determined from this. At this time, the light intensity at each time in a certain time range and can be determined from this. It is advisable to appropriately learn the lifetime values. For waveforms, normalization is preferable. Also, all wavelengths The initial integrated intensity can be normalized, and the intensity of each wavelength can be a relative value. For example, a light-emitting material In this case, the faster the decay (the shorter the lifetime), the higher the emission quantum yield is considered to be. It can be measured with a fluorescence (luminescence) lifetime measuring device. Note that the transient luminescence lifetime of the light-emitting element can be measured. In this case, electrical excitation may be performed instead of photoexcitation. That is, pulsed electrical excitation may be applied to the light-emitting element. You can also apply pressure and measure the time it takes for the luminescence intensity to decay (lifetime). The time (lifetime) is usually measured using the time it takes for the luminescence intensity to become 1 / e. There are many such cases.

[0074] The S1 level is the absorption edge of the absorption spectrum, the maximum value on the long wavelength side, and the maximum maximum of the excitation spectrum. It can be determined from the value, the maximum value of the emission spectrum, and the rise time on the short wavelength side. The T1 level is determined by the absorption edge of the absorption spectrum obtained by transient absorption measurements, or by the maximum value on the longer wavelength side. Maximum value of the phosphorescence spectrum, peak wavelength on the short-wavelength side of the phosphorescence spectrum, rise on the short-wavelength side. It can be determined from the rising value. Note that this also includes the absorption edge and the rising value of the emission spectrum. The method for determining this is as described above. Furthermore, the S1 and T1 levels can also be obtained from simulations. It can be calculated. For example, density generalizations such as those found in quantum chemistry calculation programs like Gaussian. After optimizing the structure of the ground state (S0) using numerical methods, the excitation energy was calculated using time-dependent density functional theory. It can be calculated as follows. Similarly, the Sn level (singlet level above S1) and the Tn level The level (triplet level above T1) can also be determined. In this case, the transition probability is oscillatory. The oscillator intensity can also be determined simultaneously. For example, in the case of light-emitting materials, a higher oscillator intensity indicates a higher relative strength. It is considered that light emission is more likely at this position, which is desirable. Furthermore, the structure of S0 obtained by density functional theory was optimized. The difference between the potential energy obtained and the potential energy obtained by optimizing the structure of T1 is It can also be used as a T1 level.

[0075] The emission quantum yield can be determined using an absolute quantum yield measuring device.

[0076] Oxidation potential and reduction potential can be measured by cyclic voltammetry (CV). The potential energy (eV) for oxidation / reduction is also for the HOMO and LUMO levels. Using the oxidation-reduction potential of a standard sample (e.g., ferrocene) whose oxidation-reduction potential is known as a reference, CV It can be determined by measurement. On the other hand, the HOMO level is in the air in the solid state (thin film or powder). It can also be measured using photoelectron spectroscopy (PESA). In this case, LUMO is the absorption spectrum. The band gap is determined from the absorption edge of the PESA, and its energy is set to the HOMO level determined by PESA. It can be calculated by adding the values. For example, in the case of an organic EL element, the excitatory space between two molecules To estimate the emission energy when a plex occurs, use the higher of the HOMO levels (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 other molecules (the deeper one) is determined. At this time, the HOMO level determined by CV and It is preferable to use the LUMO level. Also, the quantum chemical calculation program Gaussia In density functional theory such as n, the HOMO level and LUMO level, or the HOMO-n level (HOMO The energy levels of occupied orbits below the LUMO (energy levels of unoccupied orbits above the LUMO) can be determined. It is possible.

[0077] The glass transition temperature, melting point, and crystallization temperature can be determined using a differential scanning calorimetry (DSC) system. It is possible to measure the temperature rise rate at a constant rate of 10-50°C / min. Decomposition temperature The temperature, boiling point, and sublimation temperature can be determined using a thermogravimetric / differential thermal analysis (TG-DTA) instrument. It is advisable to use the results measured at atmospheric pressure or under reduced pressure as appropriate. Values ​​measured under reduced pressure should be used for sublimation purification. This can be used as a reference for temperature and deposition temperature, and the value should be one in which the weight has decreased by about 5-20%. This is preferable. It is preferable to keep the heating rate constant at 10-50°C / min and measure it.

[0078] Carrier mobility is determined using the time-of-flight (TOF) method, which utilizes transient photocurrent. This can be done. In the TOF method, the sample film is sandwiched between electrodes and a DC voltage is applied. In this state, carriers are generated by pulsed light excitation, and the travel time of the generated carriers (current transient) This method estimates mobility from the response. In this case, a film thickness of 3 μm or more is preferable. Another method is to determine if the current-voltage characteristics of the sample film are space charge-limited current (SCLC). If it follows this, then by fitting its current-voltage characteristics with the SCLC formula, Mobility can be determined. Furthermore, conductan can be obtained from impedance spectroscopy measurements. Methods for determining mobility from the frequency-dependent characteristics of capacitance have also been reported. In either method, it is possible to determine the mobility at a given voltage (electric field strength). This can be used as a physical property value. Furthermore, the dependence of mobility on the electric field strength can be plotted. By extrapolating, the mobility μ0 in the absence of an electric field can be determined and used as a material property. You can.

[0079] The refractive index and orientation parameters can be determined using a spectroscopic ellipsometry instrument. For example, In the case of EL elements, a lower refractive index in the visible range is preferable because it improves light extraction efficiency. There are several reported examples regarding orientation parameters, but for example, in the case of organic EL elements, orientation Parameter S is often used. The orientation parameter S is determined by light ellipsometry. This can be calculated by measuring absorption anisotropy. In the case of fluorescent materials, the lowest singlet excitation state When S is close to -0.5 at the wavelength corresponding to absorption originating from state (S1), the light extraction surface of the substrate, etc. The transition dipole moment is considered to be more horizontal, resulting in higher light extraction efficiency. This is preferable. In the case of phosphorescent materials, it is sufficient to focus on the absorption of the lowest triplet excited state (T1). Oh, when S is 0, it's random orientation, and when S is 1, it's vertical orientation. Also, with other orientation parameters... Therefore, when the transition dipole moment is divided into a component horizontal to the substrate and a component perpendicular to the substrate, The proportion of the vertical component may also be used. This parameter is related to photoluminescence (PL). ) or investigate the angular dependence of the p-polarization intensity of electroluminescence (EL), This can be determined by fitting the values.

[0080] The mass-to-charge ratio (m / z) is determined by calculating the detection intensity per unit within a fixed range of mass-to-charge ratios. You can train the model using values. Depending on the purpose, you can train it using absolute values ​​or normalized values. If you want to compare spectral shapes, normalize to an arbitrary wavelength such as the m / z of the parent ion. You just need to train the values. If you want to compare absolute values, train the model using the absolute values. m / z is It can be measured with a mass spectrometer, and ionization methods include electron ionization and chemical ionization. Electrolytic ionization, fast atomic bombardment, matrix-assisted laser desorption / ionization, electros These include pre-ionization, atmospheric pressure chemical ionization, and inductively coupled plasma methods. The child (parent molecule) decomposes (the bond breaks), and the fragment (daughter ion) is detected at the same time. In some cases, the detected m / z and the detection intensity ratio with the parent ion indicate the characteristics of the molecule. For example, fragments with the same m / z can be detected between molecules that have the same substituent. This is possible. Therefore, the ratio of the parent ion, the m / z of the fragment, and its detection intensity is studied. With practice, it can predict the m / z of fragments of other compounds and the detection intensity ratio with the parent ion. This becomes possible. Generally speaking, a stronger ionization energy leads to a higher fragment generation ratio. The rate will increase.

[0081] NMR (nuclear magnetic resonance) spectra show the chemical shift within a fixed chemical shift range. The signal intensity for each t value can be calculated and used as a learning value. Also, the chemical shift value of the peak. The integral value of its intensity (number of elements), the J value (coupling constant), etc., are expressed in parallel. It is also acceptable to do so. In this case, it is preferable to express the sum of the integral values ​​of the molecules as equal to the number of elements of the measured element. Furthermore, NMR measurement can analyze the molecular structure of a substance at the atomic level. Therefore, molecules with the same substituents will exhibit similar chemical shift values ​​and similar spectra. The spectrum can be measured using an NMR spectrometer.

[0082] ESR (Electron Spin Resonance) spectra are obtained over a fixed magnetic field strength range or magnetic flux density (Tes). (r) You can learn the detection intensity for each unit within the range and rotation angle by calculating the values. It may also be expressed as the g-factor, the square of the g-value, spin amount, spin density, etc. Note that ESR measurement is not A sample containing paired electrons absorbs microwaves in a magnetic field due to the spin transition of unpaired electrons. It observes resonance phenomena. Therefore, ESR is used to measure paramagnetic materials that have unpaired electrons. It is effective for this purpose. It can also be used to observe triplet states, for example, at low temperatures (100K~ If ESR measurements are performed while irradiating with excitation light at 10K, information about the spin state of the triplet excited state can be obtained. The report is obtained. At this time, the D value (a quantity representing the magnitude of the interaction between two electron spins), It can also be expressed using the E value (a quantity that represents how much the electron orbital deviates from axial symmetry). The vector can be measured using an ESR instrument.

[0083] Once the learning phase is complete, the next step is to use the learned results to determine the molecular structure of the input target substance. The desired physical properties are predicted from the material's structure (S102).

[0084] Finally, the predicted physical properties are output (S103).

[0085] Thus, one aspect of the present invention can predict various physical properties of organic compounds. By using multiple fingerprints when training the model to learn structures, more accurate predictions can be made. This is a method for predicting the physical properties of organic compounds.

[0086] (Embodiment 2) Embodiment 2 describes an organic compound property prediction system, which is one aspect of the present invention. do. <Example Configuration> A physical property prediction system 10 according to one aspect of the present invention includes an input means, a learning means, a prediction means, and an output means. It has at least a data server. These can each exchange data. Ideally, they could be integrated into a single device, or they could be separate devices. Furthermore, some parts may be incorporated into the same device, and the data server may be in the cloud. However, these will be collectively referred to as material property prediction systems.

[0087] Figure 8 shows an embodiment of the present invention comprising an input means, a learning means, a prediction means, and an output means. We will explain using a physical property prediction system consisting of an information terminal and a data server as an example. Terminal 20 has an input unit, a learning means, a prediction means, and an output unit, and a separately installed data storage A "user" allows for data exchange.

[0088] The information terminal 20 mainly consists of an input unit 21, a calculation unit 22, and an output unit 25. Unit 22 simultaneously handles both learning and prediction. Furthermore, the processing unit 22 is a neural network. It is preferable that it has a neural circuit. The data provided from the data server is neural This data will be used to train or predict in the network circuit 26. By using this as validation data and training data for a trained learning method, It is possible to update the weight coefficients within the multi-network circuit and generate pre-trained weight coefficients. This allows for improved prediction accuracy.

[0089] In Figure 8, the signal flow is shown in the following order: input unit 21, calculation unit 22, data server 30, and output unit 25. This is illustrated with arrows. In this specification, signals may be interpreted as data or information as appropriate. It is possible.

[0090] The data server 30 provides the learning means of the calculation unit 22 with information on the structure and physical properties of the organic compounds to be learned. Provided. The structure of the provided organic compound is represented using two or more fingerprints. It is what has been obtained. The learning means of the arithmetic unit 22 preferably has a neural network circuit. Preferably.

[0091] The input unit 21 has a function for the user to input information. Specific examples of the input unit 21 include a keyboard, a mouse, a touch panel, a tablet, a microphone, a camera, or other input means. can be mentioned. Preferably.

[0092] Input information D in is data output from the input unit 21 to the arithmetic unit 22. The input information D i n is information input by the user. For example, when the input unit 21 is a touch panel, it is information obtained by character input by operating the touch panel. Alternatively, when the input unit 21 is a microphone, it is information obtained by voice input by the user. Alternatively, when the input unit 21 is a camera, it is information obtained by image processing of imaging data. Preferably.

[0093] Input information D in is information regarding the structure of the organic compound for which physical properties are to be predicted. If it is input in a form other than fingerprint notation, such as a structural formula, an image of the structure, a substance name, etc., it is input to the prediction means in the arithmetic unit 22 via appropriate conversion means. The prediction means makes a prediction about the physical properties of the input organic compound based on the results previously learned by the learning means. Preferably. The results of the prediction are output via the output unit. Preferably. Preferably.

[0094] Preferably.

[0095] When the arithmetic unit has a neural network circuit, it is preferable that the neural network circuit has a multiply-accumulate circuit capable of executing multiply-accumulate operation processing. Also, the multiply-accumulate circuit Preferably. The path preferably has a memory circuit for storing weight data. The memory element has a transistor and a capacitive element, and the transistor is channel forming The semiconductor layer having the region contains an oxide semiconductor. It is preferable that the transistor is an OS transistor. The leakage current that flows when the OS transistor is in the off state is extremely small. By utilizing the property that it can retain electric charge when in a certain state, data can be stored. The configuration of the neural network circuit will be described in detail in Embodiment 3.

[0096] Furthermore, these multiple fingerprint shapes are linked or arranged in parallel. A control program and control software that can generate lint, perform machine learning, and predict material properties. 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 semiconductor) described in the above embodiment is used. This section describes an example configuration of a semiconductor device that can be used in a semiconductor device (referred to as a semiconductor body).

[0098] In this specification, a semiconductor device refers to a device that can function by utilizing semiconductor properties. This refers to a neural network circuit that has transistors that utilize semiconductor properties. It is a semiconductor device.

[0099] As shown in Figure 9(A), the neural network NN consists of an input layer IL, an output layer OL, and an intermediate layer. It can be constructed with layers (hidden layers) HL. Input layer IL, output layer OL, hidden layer HL Each has one or more neurons (units). Note that the intermediate layer HL can be one layer or two or more layers. A neural network having two or more intermediate layers HL can also be called a DNN (Deep Neural Network), and learning using a deep neural network can also be called deep learning.

[0100] Input data is input to each neuron in the input layer IL, and the output signals of neurons in the previous layer or the subsequent layer are input to each neuron in the intermediate layer HL, and the output signals of neurons in the previous layer are input to each neuron in the output layer OL. Note that each neuron may be connected to all neurons in the previous and subsequent layers (fully connected) or may be connected to some neurons.

[0101] FIG. 9(B) shows an example of an operation by a neuron. Here, a neuron N and two neurons in the previous layer that output signals to the neuron N are shown. To the neuron N, the output x1 of the neuron in the previous layer and the output x2 of the neuron in the previous layer are input. And in the neuron N, after calculating the sum x1w1 + x2w2 of the multiplication result (x1w1) of the output x1 and the weight w1 and the multiplication result (x2w2) of the output x2 and the weight w2, a bias b is added as necessary and a value a = x1w1 + x2w2 + b is obtained. Then, the value a is converted by an activation function h and an output signal y = h(a) is output from the neuron N. Thus, the operation by a neuron includes an operation of adding the products of the outputs of the neurons in the previous layer and the weights, that is, a sum-of-products operation (the above x1w1 + x2w2). This sum-of-products operation may be performed on software using a program or may be performed by hardware.

[0102] ​ This is also acceptable. When performing multiply-accumulate operations in hardware, a multiply-accumulate circuit can be used. This multiply-accumulate circuit may use a digital circuit or an analog circuit. This is also good. When using analog circuits in the sum-accumulate circuit, the circuit size of the sum-accumulate circuit can be reduced, This aims to improve processing speed and reduce power consumption by reducing the number of memory accesses. It is possible.

[0103] The multiply-accumulate circuit is a transistor that contains silicon (such as single-crystal silicon) in the channel formation region. It may be constructed using a (hereinafter also called a Si transistor), or in the channel formation region It is composed of transistors containing oxide semiconductors (hereinafter also called OS transistors) This is also good. In particular, because OS transistors have an extremely small off-current, the analog of the sum-of-accumulate circuit It is suitable as a transistor for constituting a memory. Note that Si transistors and OS transistors A sum-of-accumulate circuit may be constructed using both transistors. The following describes the functions of a sum-of-accumulate circuit. An example of a semiconductor device configuration will be described.

[0104] <Example of semiconductor device configuration> Figure 10 shows an example configuration of a semiconductor device MAC with the function of performing neural network calculations. This shows that the semiconductor device MAC displays the first data corresponding to the connection strength (weight) between neurons. It has the function of performing a sum-of-products operation on a second data corresponding to the input data. The first and second data sets are either analog data or multi-level data (discrete data), respectively. This can be done. Furthermore, the semiconductor device MAC can utilize the data obtained by the sum-accumulate operation. It has the function of transforming using a characterization function.

[0105] The semiconductor device MAC consists of 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 It has V.

[0106] The cell array CA has multiple memory cells MC and multiple memory cells MCref. (Figure) In 10, the cell array CA has m rows and n columns (where m and n are integers greater than or equal to 1) and memory cells MC (MC [1,1] to [m,n]) and m memory cells MCref(MCref[1] to [ This shows an example configuration having m). The memory cell MC has the function of storing the first data. It has. In addition, the memory cell MCref is a device that stores reference data used in multiply-accumulate operations. It has the capability to perform this function. The reference data can be analog data or multi-level data.

[0107] Memory cell MC[i,j] (where i is an integer between 1 and m, and j is an integer between 1 and n) It is connected to wire WL[i], wiring RW[i], wiring WD[j], and wiring BL[j]. Furthermore, the memory cell MCref[i] is connected to the wiring WL[i], wiring RW[i], and wiring WDR. ef is connected to wiring BLref. Here, memory cell MC[i,j] and wiring BL [j] The current flowing between them is I MC[i,j] This is written as and the memory cell MCref[i] is wired The current flowing between BLref is I MCref[i] This is how it is written.

[0108] Figure 11 shows specific configuration examples of memory cell MC and memory cell MCref. Representative examples include memory cells MC[1,1], [2,1] and memory cell MCref[1]. [2] is shown, but other memory cell MCs and memory cell MCrefs have a similar configuration. It can be used. Memory cell MC and memory cell MCref are transient It has transistors Tr11 and Tr12, and a capacitive element C11. Here, transistor Tr11 and Next, we will explain the case where transistor Tr12 is an n-channel type transistor.

[0109] In the memory cell MC, the gate of transistor Tr11 is connected to wiring WL, and the source Alternatively, one of the drains is connected to the gate of transistor Tr12 and the first electrode of capacitive element C11. It is connected to the source or drain, and the other end is connected to the wiring WD. Transistor Tr One of the 12 sources or drains is connected to wiring BL, and the other source or drain is connected to wiring It is connected to wire VR. The second electrode of capacitive element C11 is connected to wiring RW. A wiring VR is a wire that has the function of supplying a predetermined potential. Here, as an example, This section explains the case where a low power potential (such as ground potential) is supplied from the VR.

[0110] Either the source or drain of transistor Tr11, the gate of transistor Tr12, and The node connected to the first electrode of the capacitive element C11 is denoted as node NM. The nodes NM of cells MC[1,1] and [2,1] are respectively node NM[1,1] and [2 It is written as ,1].

[0111] Memory cell MCref has the same configuration as memory cell MC. However, memory cell M Cref is connected to wiring WDref instead of wiring WD, and wiring BL is connected instead of wiring BL. It is connected to ref. Also, in memory cell MCref[1], [2], Either the source or drain of transistor Tr11, the gate of transistor Tr12, and the capacitance. The nodes connected to the first electrode of the element C11 are denoted as node NMref[1] and [2 , respectively.

[0112] The node NM and the node NMref each function as a holding node of the memory cell MC and the memory cell MCref. The first data is held in the node NM, and the reference data is held in the node NMref. Also, currents I and I flow through the transistors Tr12 of the memory cells MC[1,1] and [2 ,1] from the wiring BL[1], respectively. Also, currents I MC[1,1] and I MC[2,1] flow through the transistors Tr12 of the memory cells MCref[1] and [2] from the wiring BLref, respectively. Also, currents I MCref[1] and I MCref[2] flow through the transistors Tr12 of the memory cells MCref[1] and [2], respectively.

[0113] Since the transistor Tr11 has a function of holding the potential of the node NM or the node NMref, it is preferable that the off-current of the transistor Tr11 is small. Therefore, it is preferable to use an OS transistor having an extremely small off-current as the transistor Tr11. Thereby, fluctuations in the potential of the node NM or the node NMref can be suppressed, and the calculation accuracy can be improved. Also, the frequency of the operation of refreshing the potential of the node NM or the node NMref can be kept low, and the power consumption can be reduced.

[0114] The transistor Tr12 is not particularly limited, and for example, a Si transistor or an OS transistor etc. can be used. When an OS transistor is used for the transistor Tr12, the transistor Tr12 can be fabricated using the same manufacturing apparatus as the transistor Tr11. ​​​​​This allows for reduced manufacturing costs. Note that transistor Tr12 is n-channel. It can be either a 1-channel type or a p-channel type.

[0115] The current source circuit CS is connected to wiring BL[1] to [n] and wiring BLref. The current source circuit CS has the function of supplying current to wiring BL[1] to [n] and wiring BLref. It has. Furthermore, the current values ​​supplied to wiring BL[1] to [n] and the current values ​​supplied to wiring BLref The current values ​​may be different. Here, the current source circuit CS is connected to the wiring BL[1] to [ The current supplied to n] is I C The current supplied from the current source circuit CS to the wiring BLref is I Cref This is how it is written.

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

[0117] A current mirror circuit CM generates a current I corresponding to the potential of node NPref. CM Wiring ILre The function of flowing current f, and this current I CM It also has the function of passing through wiring IL[1] to [n]. In 10, current I flows from wiring BLref to wiring ILref. CM The wires are discharged and the wiring BL[1] Current I from [n] to wiring IL[1] to [n] CM This shows an example of what is being discharged. Furthermore, current flows from the current mirror circuit CM to the cell array CA via wiring BL[1] to [n]. The current that is applied is I B [1] to [n] is used as denotation. Also, wiring from the current mirror circuit CM. The current flowing through BLref to the cell array CA is I Bref This is how it is written.

[0118] Circuit WDD is connected to wiring WD[1] through [n] and wiring WDref. Circuit W DD controls the potential corresponding to the first data stored in the memory cell MC, via the wiring WD[1] It has the function of supplying to [n]. In addition, the circuit WDD is stored in the memory cell MCref It has the function of supplying the potential corresponding to the reference data to the wiring WDref. It is connected to wiring WL[1] to [m]. Circuit WLD performs data writing. A signal for selecting either memory cell MC or memory cell MCref is wired WL[1] It has the function of supplying to [m]. Circuit CLD is connected to wiring RW[1] to [m]. The circuit CLD supplies the potential corresponding to the second data to the wiring RW[1] to [m]. It has the function of supplying.

[0119] The offset circuit OFST is connected to wiring BL[1] to [n] and wiring OL[1] to [n]. It is connected. The offset circuit OFST is offset from wiring BL[1] through [n]. The amount of current flowing through circuit OFST, and / or the offset from wiring BL[1] to [n] It has the function of detecting the change in the current flowing through the OFST circuit. T has the function of outputting the detection result to wiring OL[1] to [n]. Note the offset. Circuit OFST may output a current corresponding to the detection result to wiring OL, or it may output a current corresponding to the detection result The corresponding current may be converted to a voltage and output to the wiring OL. Cell array CA and offset circuit The current flowing between paths OFST is I α [1] Or [n].

[0120] Figure 12 shows an example configuration of an offset circuit OFST. (Figure 12 shows the OFST offset circuit) It has circuits OC[1] to [n]. Also, circuits OC[1] to [n] are each, Transistor Tr21, Transistor Tr22, Transistor Tr23, Capacitor C21 It also has a resistor R1. The connection relationship of each element is as shown in Figure 12. The node connected to the first electrode of element C21 and the first terminal of resistor element R1 is node N. Let a be the second electrode of the capacitive element C21, and the source or drain of the transistor Tr21. Let one of the inputs and the node connected to the gate of transistor Tr22 be called node Nb. ru.

[0121] Wiring VrefL has the function of supplying potential Vref, and wiring VaL supplies potential Va. Wiring VbL has the function of supplying potential Vb. Wiring VDDL has the function of supplying potential It has the function of supplying VDD, and the wiring VSSL has the function of supplying potential VSS. Here Now, let's explain the case where the potential VDD is the high power supply potential and the potential VSS is the low power supply potential. Furthermore, the wiring RST provides a potential to control the conduction state of transistor Tr21. It has a supply function. Transistor Tr22, Transistor Tr23, Wiring VDDL, Distribution The source follower circuit is formed by the line VSSL and the wiring VbL.

[0122] Next, we will explain examples of the operation of circuits OC[1] to [n]. Note that the circuit shown here is a representative example. An example of OC[1] operation will be explained, but circuits OC[2] through [n] can be operated in the same way. Yes, it is possible. First, when the first current flows through wiring BL[1], the potential of node Na is the same as the potential of the first current. The potential is determined by the current and the resistance of the resistive element R1. Also, at this time, transistor Tr21 It is in the ON state, and potential Va is supplied to node Nb. Subsequently, transistor Tr21 It will switch to the off state.

[0123] Next, when a second current flows through wiring BL[1], the potential of node Na is equal to 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. Yes, and since node Nb is in a floating state, it reacts with changes in the potential of node Na. First, the potential of node Nb changes due to capacitive coupling. Here, the change in the potential of node Na is Δ V Na Assuming a capacitive coupling coefficient of 1, the potential at node Nb is Va + ΔV Na This is how it will be. Then, the threshold voltage of transistor Tr22 is set to V th Therefore, the potential from wiring OL[1] Va+ΔV Na -V th The output is: Here, Va = V th By doing so, wiring O From L[1], the potential ΔV Na It can output.

[0124] Potential ΔV Na This is the change in current from the first to the second, the resistance of the resistor R1, and the potential Vre. It is determined by f. Here, since the resistive element R1 and the potential Vref are known, the potential ΔV N a From this, we can determine the change in the current flowing through wiring BL.

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

[0126] The activation function circuit ACTV consists of wiring OL[1] to [n] and wiring NIL[1] to [ It is connected to n]. The activation function circuit ACTV is input from the offset circuit OFST. A function that performs calculations to transform a given signal according to a predefined activation function. It has. Examples of activation functions include the sigmoid function, the tanh function, and the softmax function. Functions such as ReLU functions and threshold functions can be used. Activation function circuit ACTV The converted signal is output as output data to wiring NIL[1] through [n]. .

[0127] <Example of semiconductor device operation> Using the above semiconductor device MAC, it is possible to perform a sum-of-products operation on the first data and the second data. Yes, it is possible. The following describes an example of the operation of the semiconductor device MAC when performing a multiply-accumulate operation.

[0128] Figure 13 shows a timing chart of an example of semiconductor device MAC operation. Figure 13 is shown in Figure 11. Wiring WL[1], Wiring WL[2], Wiring WD[1], Wiring WDref, Node NM [1,1], node NM[2,1], node NMref[1], node NMref[2] The potential changes of wiring RW[1] and wiring RW[2], and current I B [1]-I α [1], and current I Bref This shows the trend of the value of current I. B [1]-I α [1] is wiring BL This corresponds to the sum of the currents flowing from [1] to the memory cells MC[1,1] and [2,1].

[0129] Here, as representative examples, we have the memory cells MC[1,1], [2,1] and shown in Figure 11. The operation will be explained focusing on the Morisel MCref[1], [2], but other memory cell MCs and The memory cell MCref can be operated in the same way.

[0130] [Storage of the first data] First, at times T01-T02, the potential of wiring WL[1] is high. As a result, the potential of wiring WD[1] is higher than the ground potential (GND). PR -V W[1,1] big The potential becomes such that the potential of the wiring WDref is higher than the ground potential. PR The potential becomes large. The potential of line RW[1] and wiring RW[2] becomes the reference potential (REFP). Note that potential V W[1,1] This is the potential corresponding to the first data stored in the memory cell MC[1,1]. Also, the potential V PR This is the potential corresponding to the reference data. [1,1] and the transistor Tr11 of memory cell MCref[1] are in the ON state. As a result, the potential at node NM[1,1] is V PR -V W[1,1] ,NodeNMref[1] The potential is V PR This is the result.

[0131] At this time, current flows from wiring BL[1] to transistor Tr12 of memory cell MC[1,1]. Current I MC[1,1],0 This can be expressed by the following equation, where k is a transistor. Constants determined by the channel length, channel width, mobility of Tr12, and the capacitance of the gate insulating film. It is. Also, Vth This 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] Furthermore, the current flows from the wiring BLref to the transistor Tr12 of the memory cell MCref[1]. current I MCref[1],0 This can be expressed by the following formula:

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

[0135] Next, at times T02-T03, the potential of wiring WL[1] becomes low. This allows the memory cell MC[1,1] and memory cell MCref[1] to have Rangitator Tr11 is turned off, and nodes NM[1,1] and NMref[1 The potential of ] is maintained.

[0136] As mentioned above, it is preferable to use an OS transistor as transistor Tr11. This allows the leakage current of transistor Tr11 to be suppressed, and node NM[ The potentials of [1,1] and node NMref[1] can be accurately maintained.

[0137] Next, at times T03-T04, the potential of wiring WL[2] becomes high, and wiring W The potential of D[1] is greater than the ground potential. PR -V W[2,1] The potential becomes high, and the wiring WDr The potential of ef is greater than the ground potential. PR The potential becomes large. W[2,1] This is a memo. This is the potential corresponding to the first data stored in the MC[2,1]. The transistor Tr11 in the Morissel MC[2,1] and memory cell MCref[2] When it turns ON, the potential at node NM[1,1] becomes V PR -V W[2,1] , node NM The potential at ref[1] is V PR This is the result.

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

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

[0142] Next, at times T04-T05, the potential of wiring WL[2] becomes low. The transistors in memory cell MC[2,1] and memory cell MCref[2] Tr11 is turned off, and the potentials of nodes NM[2,1] and NMref[2] are It is retained.

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

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

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

[0146] Wiring BL[1] is supplied with current from the current source circuit CS. Also, wiring BL[1] The current flowing through is discharged to the current mirror circuit CM and memory cells MC[1,1] and [2,1]. It is released. Also, current flows from wiring BL[1] to offset circuit OFST. Current source circuit The current supplied from circuit CS to wiring BL[1] is I C,0 , offset from wiring BL[1] The current flowing through circuit OFST is I α,0 Therefore, the following equation holds true.

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

[0148] [Sum of Products operation on the first and second data] Next, at times T05-T06, the potential of wiring RW[1] is higher than the reference potential. X[1] It becomes a large potential. At this time, the memory cell MC[1,1] and the memory cell MCref 1], the potential V X[1] is supplied to each capacitive element C11, and the potential of the gate of the transistor Tr12 rises due to capacitive coupling. Note that the potential V is the potential corresponding to the second data supplied to the memory cell MC[1 x[1] ,1] and the memory cell MCref[1]. .

[0149] The change amount of the potential of the gate of the transistor Tr12 is a value obtained by multiplying the change amount of the potential of the wiring RW by the capacitive coupling coefficient determined by the configuration of the memory cell. The capacitive coupling coefficient is calculated based on the capacitance of the capacitive element C1 1, the gate capacitance of the transistor Tr12, and parasitic capacitance, etc. Hereinafter, for the sake of convenience, it will be described assuming that the change amount of the potential of the wiring RW and the change amount of the potential of the gate of the transistor Tr12 are the same, that is, the capacitive coupling coefficient is 1. In actuality, the potential V should be determined considering the capacitive coupling coefficient. x

[0150] When the potential V X[1 ] is supplied to the capacitive elements C11 of the memory cell MC[1] and the memory cell MCref[1], the potentials of the nodes NM[1] and NMref[1] rise to V X [1] respectively.

[0151] Here, at time T05 - T06, the current I flowing from the wiring BL[1] to the transistor Tr12 of the memory cell MC[1,1] MC[1,1],1 can be expressed by the following formula .

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

[0153] That is, potential V in wiring RW[1] X[1] By supplying this, from wiring BL[1] The current flowing through transistor Tr12 of memory cell MC[1,1] is ΔI MC[1,1] =I MC[1,1],1 -I MC[1,1],0 It increases.

[0154] Also, at times T05-T06, the wiring BLref to the memory cell MCref[1] Current I flowing through transistor Tr12 MCref[1],1 This can be expressed by the following formula: ru.

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

[0156] That is, potential V in wiring RW[1] X[1] By supplying the wiring from BLref The current flowing through transistor Tr12 of memory cell MCref[1] is ΔI MCref[ 1] =I MCref[1],1 -I MCref[1],0 It increases.

[0157] Furthermore, we consider the current flowing through wiring BL[1] and wiring BLref. In this circuit, current I is supplied from the current source circuit CS. Cref It is supplied. Also, it flows through the wiring BLref The current is discharged to the current mirror circuit CM and memory cells MCref[1], [2]. The current discharged from wiring BLref to the current mirror circuit CM is I CM,1 So, next The following equation holds true.

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

[0159] Wiring BL[1] receives current I from the current source circuit CS. C It is supplied. Also, wiring BL[1] The current flowing through is discharged to the current mirror circuit CM and memory cells MC[1,1] and [2,1]. It is released. Furthermore, current flows from wiring BL[1] to the offset circuit OFST. Wiring The current flowing from BL[1] to the offset circuit OFST is I α,1 Then the following equation holds: 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 The difference (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 The potential V W[1,1] and V X[1] The value will be proportional to the product of the two numbers. .

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

[0165] Next, at times T07-T08, the potential of wiring RW[1] is higher than the reference potential. X[1] The potential becomes larger, and the potential of wiring RW[2] is higher than the reference potential. X[2] The potential becomes large. This allows memory cell MC[1,1] and memory cell MCref[1] to be respectively The capacitance element C11 has a potential V X[1] A supply is provided, and via capacitive coupling, node NM[1,1] and The potentials of the nodes NMref[1] are V X[1] It rises. Also, the memory cell MC [2,1] and the capacitive element C11 of the memory cell MCref[2] have a potential V X[ 2] Power is supplied, and capacitive coupling connects the power of nodes NM[2,1] and NMref[2]. Each position is V X[2] It rises.

[0166] Here, at time T07-T08, the wiring BL[1] to memory cell MC[2,1] Current I flowing through transistor Tr12 MC[2,1],1 It can be expressed by the following formula: .

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

[0168] That is, potential V in wiring RW[2] X[2] By supplying this, from wiring BL[1] The current flowing through transistor Tr12 of memory cell MC[2,1] is ΔI MC[2,1] =I MC[2,1],1 -I MC[2,1],0 It increases.

[0169] Also, at times T05-T06, the wiring BLref to the memory cell MCref[2] Current I flowing through transistor Tr12 MCref[2],1 This can be expressed by the following formula: ru.

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

[0171] That is, potential V in wiring RW[2] X[2] By supplying the wiring from BLref The current flowing through transistor Tr12 of memory cell MCref[2] is ΔI MCref[ 2] =I MCref[2],1 -I MCref[2],0 It increases.

[0172] Furthermore, we consider the current flowing through wiring BL[1] and wiring BLref. In this circuit, current I is supplied from the current source circuit CS. Cref It is supplied. Also, it flows through the wiring BLref The current is discharged to the current mirror circuit CM and memory cells MCref[1], [2]. The current discharged from wiring BLref to the current mirror circuit CM is I CM,2 So, next The following equation holds true.

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

[0174] Wiring BL[1] receives current I from the current source circuit CS. C It is supplied. Also, wiring BL[1] The current flowing through is discharged to the current mirror circuit CM and memory cells MC[1,1] and [2,1]. It is released. Furthermore, current flows from wiring BL[1] to the offset circuit OFST. Wiring The current flowing from BL[1] to the offset circuit OFST is I α,2 Then the following equation holds: stand.

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

[0176] Then, from equations (E1) to (E8) and equations (E12) to (E15), the current I α,0 and current I α,2 The difference (differential 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 The potential V W[1,1] and potential V X[1] The product of and the potential V W [2,1] and potential V X[2] The value will be determined by the product of and the sum of .

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

[0180] As shown in equations (E9) and (E16), the difference input to the offset circuit OFST Minute current ΔI α This is the potential V corresponding to the first data (weight). X And the second data (input data) Potential V corresponding to (T) W The value corresponds to the sum of the products of these two factors. In other words, the differential current. ΔI α By measuring with an offset circuit OFST, the first data and the second data The result of a sum-of-products operation can be obtained.

[0181] Note that the above specifically refers to memory cells MC[1,1], [2,1] and memory cells MCref[ We focused on [1] and [2], but the number of memory cells MC and memory cells MCref can be set arbitrarily. This is possible. The number of rows m of the memory cell MC and memory cell MCref can be set to any 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] Furthermore, by increasing the number of rows n of memory cell MC and memory cell MCref, parallel processing can be achieved. The number of multiply-accumulate operations performed can be increased.

[0184] As described above, by using the semiconductor device MAC, the product of the first data and the second data can be calculated. It is possible to perform summation operations. Note that the memory cell MC and memory cell MCref are shown in Figure 1. By using the configuration shown in 1, a multiply-accumulate circuit can be constructed with a small number of transistors. This makes it possible to reduce the circuit size of the semiconductor device MAC.

[0185] When using a semiconductor device MAC for computation in a neural network, memory cell MC The number of rows m corresponds to the number of input data supplied to one neuron, and the number of rows of memory cells MC. The number n can be represented as the number of neurons. For example, the hidden layer HL shown in Figure 9(A) Let's consider the case where a multiply-accumulate operation is performed using a semiconductor device MAC. In this case, a memory cell The number of rows m in the MC is the number of input data supplied from the input layer IL (the number of neurons in the input layer IL). Set the number of columns n in the memory cell MC to the number of neurons in the hidden layer HL. It is possible.

[0186] Furthermore, the structure of the neural network to which the semiconductor device MAC is applied is not particularly limited. For example, semiconductor devices like MACs use convolutional neural networks (CNNs) and recurrent neural networks. RNN (Restricted Network), Autoencoder, Boltzmann Machine (Restricted Boltzmann Machine) It can also be used for things like (including the synth).

[0187] As described above, by using the semiconductor device MAC, the sum-of-products performance of the neural network can be performed. It can perform calculations. Furthermore, the cell array CA contains memory cells MC and memory cells as shown in Figure 11. By using Ricell MCref, improvements in calculation accuracy, reduction in power consumption, or circuit size can be achieved. This allows us to provide an integrated circuit (IC) that can reduce the size of the components. [Examples]

[0188] In this example, we will explain in detail an example of predicting the physical properties of an organic compound. The T1 level was selected as a physical property value to be predicted in relation to the molecular structure. The T1 level value corresponds to the short-wavelength emission peak in the phosphorescence spectrum obtained by low-temperature PL measurement. These values ​​were obtained from wavelength. There are a total of 420 data points, 380 for training and 380 for testing. We used 40 values ​​to evaluate the validity of the predictive model.

[0189] For formulating molecular structures mathematically, we use R, an open-source cheminformatics toolkit. I used RDKit. RDKit allows you to input the molecular structure from SMILES notation into a fingerprint. The fingerprinting method can be used to convert it into mathematical formula data. cular and atom pair types were used.

[0190] The input values ​​for predicting physical properties are mathematical formulas expressed only in Circular type, Ato The formulas used were those expressed using the Pair type alone, and also formulas that combined both. For the lar type, the radius was specified as 4, and for the Atom Pair type, the path length was specified as 30. The bit length of the fingerprint was set to 2048. Note that the radius of the Circular type and A The path length of a tom pair is defined as the length of the path from a starting element (which is set to 0) to the length of the path connected to that element. This is the number of elements counted.

[0191] Note that when expressed as Circular type alone, out of 420 types of organic compounds, the formula is There were two pairs that were identical. On the other hand, there were Atom Pair type alone, or Circular When r-type and Atom Pair-type are linked and expressed together, mathematical formulas are used between different organic compounds. We confirmed that they were all different and not identical.

[0192] The machine learning method used was a neural network. The programming language used was Python. For hon, Chainer was used as the machine learning framework. Neural network The structure of the workpiece consists of two hidden layers. The number of neurons in each layer is 2048 (C) in the input layer. (Number of bits for ircular type alone or Atom Pair type alone) or 4096 (Ci (Number of bits when rcular type and Atom Pair type are concatenated), first hidden layer and the second The two hidden layers were set to 500, and the output layer to 1. The ReLU function was used as the activation function for the hidden layers. Ta.

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

[0194] Based on the above results, the fingerprints of Circular and Atom Pair types are When formulas expressed by law are used in combination, the result is greater than when each is used individually. The mean squared error of the test data decreased, and the prediction accuracy of the T1 level improved.

[0195] From the above, different substructures are generated for each fingerprint type, and these substructures Because information about presence or absence can be used to supplement information related to the entire molecular structure, different types of finger patterns The method of describing molecular structure using multiple Lint methods is effective for predicting physical properties using machine learning. This can be understood.

[0196] Furthermore, in cases where different compounds have the same notation using one fingerprinting method, By concatenating other fingerprints, the resulting formulas can be different. This makes it easier. By using only one type of fingerprint template, compounds with the same notation are eliminated. Rather than increasing the number of bits, it is better to combine two or more types of fingerprints. However, the generated mathematical formulas are unlikely to be identical, and the differences between compounds are expressed using the smallest possible number of bits. This is desirable because it allows for a significant reduction in the computational load in machine learning. [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 Zista, Tr22: Transistor, Tr23: Transistor, 20: Information terminal, 21: Input Power unit, 22: Calculation unit, 25: Output unit, 30: Data server

Claims

1. A physical property prediction system comprising a memory circuit, an offset circuit electrically connected to the memory circuit via a first wiring, and an activation function circuit electrically connected to the offset circuit via a second wiring, The offset circuit comprises a first to third transistor, a capacitive element, and a resistive element. Either the source or the drain of the first transistor is electrically connected to the gate of the second transistor. Either the source or the drain of the first transistor is electrically connected to the first terminal of the capacitive element. The source or drain of the first transistor, the other of which is electrically connected to the wiring that supplies the first potential, The gate of the first transistor is electrically connected to wiring that supplies a potential to control the conduction state of the first transistor. Either the source or the drain of the second transistor is electrically connected to the first power line. The source or drain of the second transistor is electrically connected to the source or drain of the third transistor. The source or drain of the second transistor, the other of which is electrically connected to the second wiring, The source or drain of the third transistor, the other of which is electrically connected to the second power line, The gate of the third transistor is electrically connected to the wiring that supplies the second potential. The second terminal of the capacitive element is electrically connected to the first wiring. The second terminal of the capacitive element is electrically connected to the first terminal of the resistive element. The second terminal of the resistive element is electrically connected to the wiring that supplies the third potential. The aforementioned physical property prediction system includes a first step of generating a first mathematical formula using the Circular type of the fingerprint method for the molecular structure of an organic compound, The second step involves generating a second mathematical formula using the Atom Pair type of fingerprinting method for the molecular structure of the aforementioned organic compound, A third step is to generate a third formula by concatenating the first formula and the second formula, The fourth step involves learning the correlation between the third mathematical formula and the physical properties, The process includes a fifth step of predicting the desired physical properties from the molecular structure of the target substance based on the results of the learning described above. A material property prediction system in which the first and second steps are performed simultaneously.

2. In claim 1, The memory circuit has memory elements, The memory element comprises a transistor and a capacitive element. The transistor is a physical property prediction system having an oxide semiconductor in the channel formation region.

Images