Material information database including machine learning models

Abstract

An example relating to material discovery using machine learning models is disclosed. One example provides a method (200, 500) performed on a computing system. This method includes receiving a query containing one or more elemental information and material property information (202, 502), and retrieving material data (504) from a material information database (108, 204) based on the query. The material data includes structural information (310, 320, 410, 420) for each material in a set of materials that match the query, where the set includes one or more materials and, for one or more materials in the set of materials, one or more predicted material properties (324, 424) determined using one or more trained machine learning models (110). This method further includes outputting the material data (210, 520).

Description

Background Art

[0001] Background

[0001] An information database can be used to assist in searching for materials having desired properties or desired elemental compositions. However, the data exists for a small subset of the accessible material space. Further, the experimental material data for some materials may be incomplete. For example, some material properties may be missing.

Summary of the Invention

Means for Solving the Problems

[0002] Summary

[0002] Examples related to material discovery using a machine learning model are disclosed. One example provides a method performed on a computing system. The method includes receiving a query that includes one or more of elemental information and material property information, and obtaining material data from a material information database based on the query. The material data includes the structural information of each material within a set of materials that match the query, and the set includes one or more materials and one or more predicted material properties determined using one or more trained machine learning models for one or more of the materials within the set of materials. The method further includes outputting the material data.

[0003]

[0003] Another example provides a method performed on a computing system, the method including receiving a query that includes elemental information, and inputting the elemental information to a trained generative machine learning model to generate a plurality of candidate structures based on the elemental information. The method includes inputting the candidate structures to a second trained machine learning model configured to perform structural relaxation and output a relaxed candidate structure for one or more of the plurality of candidate structures, inputting the relaxed candidate structure to a third trained machine learning model configured to output predicted material properties, and further including outputting the relaxed candidate structure and the predicted material properties.

[0004]

[0004] This summary is provided to introduce a set of concepts that will be further described in the following detailed description in a simplified form. This summary is not intended to identify any important or essential features of the claims, nor to be used to limit the scope of the claims. Furthermore, the claims are not limited to any implementation that resolves any or all of the inconveniences shown anywhere in this disclosure. [Brief explanation of the drawing]

[0005] Brief explanation of the drawing [Figure 1]

[0005] An exemplary computational architecture for implementing an interactive materials discovery platform is schematically shown. [Figure 2]

[0006] A flowchart illustrating an exemplary method for processing queries is shown. [Figure 3]

[0007] The table shows exemplary material data from a materials information database, including material properties predicted using a machine learning model. [Figure 4]

[0008] The table shows exemplary material data for materials generated using a machine learning model. [Figure 5A]

[0009] This flowchart shows an exemplary method for processing material information database queries. [Figure 5B]

[0009] A flowchart shows an exemplary method for processing material information database queries. [Figure 6]

[0010] This flowchart illustrates another exemplary method for processing material information database queries by using machine learning models to generate candidate material structures and predicting the material properties of those candidate structures. [Figure 7]

[0011] A block diagram of an exemplary computing system is shown. [Modes for carrying out the invention]

[0006] Detailed explanation

[0012] Materials discovery can be viewed as a search problem in a vast chemical space. However, trial-and-error research methods are time-consuming, and material synthesis and experimentation are costly, making materials discovery difficult. Previously, quantum mechanical (QM) calculations were successful in finding new molecules or materials that met property requirements. However, QM calculations can be computationally expensive, which can limit the search space. Further complicating materials discovery is the size of the unexplored chemical space. For example, regarding molecules, estimating the accessible chemical space for organic small molecules with 30 atoms or less is 10⁻¹⁰. 20 from 10 24 While the largest databases range from 10 to 10. 12 It contains information at a sub-molecular level. Similarly, for solid materials, data exists for only a small fraction of the accessible material space. As a specific example, the Materials Project database contains information on fewer than 200,000 materials (Anubhav Jain, Shyue Ping Ong, Geoffrey Hautier, Wei Chen, William Davidson Richards, Stephen Dacek, Shreyas Cholia, Dan Gunter, David Skinner, Gerbrand Ceder, Kristin A. Persson; Commentary: The Materials Project: A materials genome approach to accelerating materials innovation. APL Mater 1 July 2013;1(1):011002). Furthermore, experimental material data may be incomplete for some materials, meaning that some material properties may be missing.

[0007]

[0013] In recent years, artificial intelligence (AI), particularly machine learning, has begun to be used to learn quantitative structure-property correlations (QSPRs) of small molecules and solid crystalline materials. Machine learning techniques can be orders of magnitude faster than QM calculations. However, machine learning models can suffer from a lack of training data. The effectiveness of machine learning for material discovery depends on the quality and quantity of training data. QM calculations can be used to generate such training data. However, a brute-force approach to exploring the material space using QM calculations may not be practical. Furthermore, such methods may not target useful regions of the material space.

[0008]

[0014] Accordingly, we disclose an example relating to an interactive materials discovery platform that includes a machine learning model and a materials information database. The materials discovery platform can integrate machine learning inference, database retrieval, quantum mechanical calculations, and machine learning training into a loop to accelerate the discovery of new materials. The discovery workflow uses data generated by user-driven discoveries to grow the materials information database and improve the machine learning models. As described in more detail below, the materials information database contains materials data for multiple materials. For each material, the materials data includes structural information and material property information. For at least some materials in the materials information database, the materials data includes predicted values ​​for one or more material properties. The material properties are predicted using the corresponding trained machine learning model. Users can query the materials information database to discover materials with selected elemental compositions and / or selected material properties.

[0009]

[0015] The interactive materials discovery platform may also include generative machine learning models for generating candidate structures for materials not enumerated in the materials information database. For example, a user query may yield few matches, or no matches at all. In response, the generative machine learning model can be used to generate candidate structures based on the elemental information of the user query. Next, a property prediction pipeline is used to predict one or more material properties of the candidate structures. The property prediction pipeline may include machine learning models for structural relaxation of the candidate structures, machine learning models for material property prediction, uncertainty estimation, and / or quantum mechanical calculations. The data for the candidate structures can then be inserted into the materials information database. Quantum mechanically calculated values ​​can be used to further train the machine learning models. In this way, the interactive machine learning models can be further improved by continuously using the materials discovery platform. By performing QM calculations in response to user-initiated queries, the machine learning models can be further trained using data from more relevant areas of the unexplored materials space.

[0010]

[0016] Before discussing these examples in detail, Figure 1 shows an exemplary computing architecture 100 for implementing an interactive materials discovery platform. As shown in Figure 1, a client computer 102 can submit queries to a remote computing system 104. The remote computing system 104 includes one or more processors 106 configured to process queries and perform various other functions. The method of processing queries will be discussed below with reference to Figure 2. In some examples, the remote computing system 104 may represent a data center. Examples of computing systems will be described in more detail below with reference to Figure 7.

[0011]

[0017] The remote computing system 104 further includes a material information database 108, one or more machine learning models 110, and a storage subsystem for storing data for training data 112. The material information database 108 contains data for multiple materials, including predicted material properties determined using the machine learning models 110. Details of the material information database are discussed below with reference to Figure 3.

[0012]

[0018] The machine learning model 110 includes multiple machine learning models for predicting material properties. For selected material properties, one or more machine learning models can be trained using training data 112 to predict the selected material properties. In this way, the machine learning model 110 is used to augment the material data in the material information database 108.

[0013]

[0019] A trained machine learning model can be configured to output predictions of material properties based on a material structure input. The material structure can be represented in any suitable way, such as a graph. In an exemplary graph, nodes represent atoms and edges represent chemical bonds. Exemplary outputs of a machine learning model may include material property data (e.g., band gap, dielectric constant) and constituent atom or ion data (e.g., properties of atoms / ions in the material structure such as ionic radius, atomic number, atomic mass, electron configuration, charge). Any suitable machine learning architecture can be used. Examples of suitable architectures for machine learning models include neural networks (NNs) and random forests. A particular example of an NN is a graph neural network (GNN). GNNs are trained to perform inference on data described by mathematical graphs. Graphs can be a suitable choice for representing crystals (e.g., a crystal unit cell), while nodes represent atoms and edges represent bonds. In some examples, the representation of the material structure may include one or more pieces of information such as space group, lattice structure, atomic positions, interatomic distances, bond angles, and symmetry information. In some examples, the material information is encoded using an encoding scheme. For example, interatomic distances can be encoded and represented by the edges of a graph. Furthermore, the type of atom can be represented, for example, by its chemical symbol or atomic number. In some cases, encoding can be used to represent two or more atoms as a single unit, such as polyatomic ions or functional groups.

[0014]

[0020] The machine learning model 110 can be trained on the training data 112 using any suitable method. In some examples, the machine learning model can be trained using supervised learning. Supervised learning involves training the machine learning model using corresponding inputs (e.g., material structure) and output values ​​(e.g., material property values). In various examples, supervised learning may include regression-type prediction problems and classification-type prediction problems. Regression techniques include predicting continuous variables such as band gap, bulk modulus, or dielectric constant. Classification tasks include predicting categorical variables such as metal vs. nonmetal, or conductivity vs. insulation. The machine learning model can include any suitable algorithm. Examples of regression learning algorithms include linear regression, regression trees, and support vector machines. Examples of classification algorithms include k-nearest neighbors, naive Bayes, and decision trees. In some examples, the machine learning model is a feedforward neural network. A feedforward NN can be trained using backpropagation to compute the gradient of the loss function.

[0015]

[0021] Each machine learning model 110 can be configured to predict material properties. Examples include predicted bandgap, predicted conductivity, predicted bulk modulus, predicted shear modulus, predicted formation energy, predicted phonon density of states peak, predicted dielectric constant, and predicted refractive index. As a specific example, a machine learning model for predicting the bandgap of a perovskite structure using crystal site feature embedding is described in Hitarth Choubisa, Mikhail Askerka, Kevin Ryczko, Oleksandr Voznyy, Kyle Mills, Isaac Tamblyn, Edward H. Sargent, Crystal Site Feature Embedding Enables Exploration of Large Chemical Spaces, Matter, Volume 3, Issue 2, 2020, Pages 433-448. In some examples, each predicted material property is predicted using a corresponding trained machine learning model. In some examples, selected material properties are predicted using multiple trained machine learning models or an ensemble of machine learning models. In some such examples, each trained machine learning model in an ensemble of trained machine learning models contains different hyperparameters. In some examples, each trained machine learning model in an ensemble of trained machine learning models is trained on the same training data. In other examples, each trained machine learning model in an ensemble of trained machine learning models is trained on different training data, for example, different subsets of the training data.

[0016]

[0022] In some examples, a machine learning model 110 can be used to determine the uncertainty of the predicted material properties. Any suitable method can be used to estimate the uncertainty. In some examples, an ensemble of trained machine learning models is used to determine the ensemble of values ​​for the predicted material properties, and the uncertainty is estimated based on the ensemble of values.

[0017]

[0023] The training data 112 used to train the machine learning model 110 includes material information of a plurality of materials. For each material in the training data 112, the material information includes structural information and material property information. The material property information can include experimentally determined values and / or QM calculated values. As will be described later, additional material data can be added to the training data 112 based on the QM calculation 114.

[0018]

[0024] Subsequently, the remote computing system 104 can further hold instructions for performing the QM calculation 114. The QM calculation 114 can utilize any suitable method. Examples include the Hartree-Fock method (HF), Moller-Plesset perturbation theory (MP2), and the DFT method. The DFT method can utilize any suitable approximation for the exchange and correlation functions such as the local density approximation (LDA), generalized gradient approximation (GGA), hybrid functions (e.g., B3LYP, PBE0), meta-GGA functional, or meta-hybrid functional (e.g., M06-L). The QM calculation can be performed using any suitable basis set such as Gaussian-type orbits and / or plane waves. Further, the QM calculation can be performed using any suitable convergence criterion. In some examples, the QM calculation includes performing convergence of the basis set. In some examples, the QM calculation includes performing k-point convergence. In some examples, the method used for the QM calculation depends on the structure of the material. For example, the QM calculation of graphite can include corrections considering van der Waals forces, whereas such corrections can be omitted in the calculation of diamond. In some examples, the method used for the QM calculation depends on the material properties being calculated. For example, the QM calculation of magnetic properties can include performing spin polarization calculations. Next, if it is determined that the material is non-magnetic, the QM calculation of the bulk modulus can include performing a non-spin polarization calculation.

[0019]

[0025] In some examples, the results of the QM calculations 114 are added to the materials information database 108. Further, in some examples, the results of the QM calculations 114 are added to the training data 112 and used to further train the machine learning model 110. This process is discussed in more detail below with respect to FIG. 2.

[0020]

[0026] Continuing, the remote computing system 104 can communicate with one or more additional computing systems 120, a cloud computing service 130, and one or more third-party material databases 140. In some examples, the remote computing system 104 can offload processing tasks to the additional computing systems 120 and the cloud computing service 130. For example, the additional computing systems 120 and the cloud computing service 130 can be configured to determine predicted material properties using a machine learning model, to train a machine learning model, or to perform QM calculations. Further, the remote computing system 104 can also obtain material data for one or more materials from the third-party material database 140. In such some examples, the material information obtained from the third-party material database 140 can be used to update the materials information database 108 or the training data 112. For example, experimental data obtained from the third-party material database 140 can be added to the training data 112 and used to further train the machine learning model 110.

[0021]

[0027] As described above, a user can query the materials information database to discover materials. FIG. 2 shows a flowchart of an exemplary method 200 for processing queries on an interactive materials discovery platform that includes a cycle for adding new material properties to the database using QM calculations. The method 200 can be performed on one or more computing systems that use, for example, distributed or cloud computing. The method 200 is an example of a method that can be implemented on the remote computing system 104 in response to a query from, for example, the client computer 102.

[0022]

[0028] Method 200 includes receiving a user query in 202. The user query includes one or more elemental information and material property information. In some examples, the query can be processed using natural language processing (NLP). For example, the user query can be input into a large language model configured to process the user query and extract elemental information and / or material property information. Elemental information relates to the elemental composition of the material. In some examples, elemental information includes a list of elements present at any appropriate concentration level in the material. In some examples, elemental information specifies a concentration range of elements in the material. In some more specific examples, elemental information includes chemical formulas. Material property information relates to one or more material properties. In some examples, material property information may include values ​​or ranges of values ​​for material properties relating to a desired material.

[0023]

[0029] An example of a user query is discussed with respect to Figure 3. Figure 3 shows exemplary material data 300 for a set of materials in a material information database (e.g., material information database 108). In the illustrated example, material data 300 includes records for materials 302, 303, 304, 305, and 306. Material data 300 further includes formula 310, structural information 320, and property information 322 for each of materials 302, 303, 304, 305, and 306. Structural information 320 includes information relating to the material structure of each material in the material information database. For example, material 302 contains iron(III) oxide (Fe2O3). The structural information 320 for Fe2O3 indicates that its structure is a crystal with 10 lattice sites in the unit cell. Structural information 320 may also include information not shown in Figure 3. For example, additional structural information may include molecular symmetry groups, bond lengths, bond angles, material space groups, unit cell lattice constants, unit cell lattice vectors, coordinates of each site within the unit cell, elemental information of each atom occupying a site within the unit cell, and other information. Generally, each material listed in the material information database contains a different structure. However, in some cases, different materials may contain relatively small structural variations.

[0024]

[0030] The property information 322 includes information relating to one or more material properties. Examples of property information include numerical values, spectral data, and categorical data (e.g., metallic or nonmetallic). As shown in Figure 3, the property information 322 includes band gap values ​​324 for each of the materials 302, 303, 304, 305, and 306. The property information 322 also includes information relating to the determination of the property values. For example, the band gap of material 302 (Fe2O3) is 2.2 eV, which was determined experimentally. Furthermore, the band gap of material 303 (C) is 4.1 eV, as determined by DFT / PBE (DFT with Perdew-Burke-Ernzerhof (PBE) exchange-correlation function). The material information database may also include molecular information. For example, material 306 contains H2O. Since material 306 is a molecule, the band gap does not apply. However, in examples involving molecular crystals, the material information may include band gap information.

[0025]

[0031] For at least some materials in the materials information database, the material data includes predicted characteristic values ​​determined by a machine learning model. In the example shown in Figure 3, material 304 (Al2O3) includes a band gap of 5.7 eV predicted using a trained machine learning model. For the ML-predicted characteristic, material data 300 includes an uncertainty of 326. In the illustrated example, the uncertainty of the ML-predicted band gap for material 304 is 0.1 eV.

[0026]

[0032] Returning to Figure 2, based on the user query, method 200 includes retrieving material data from the material information database in 204. In 206, method 200 determines whether the material information database contains records for one or more materials that match the query. If records matching the query exist in the material information database, method 200 may include retrieving the corresponding material data from the material information database.

[0027]

[0033] In some cases, the material information database contains partial material data for selected materials that match the user query. However, the material data for selected materials may lack material property information relevant to the user query. Therefore, in 207, method 200 optionally includes determining whether the material data for selected materials contains material property values ​​relevant to the user query. If the material data contains the desired material property information, method 200 can proceed to 208. However, if the material data lacks information about the desired material properties, the missing data can be generated using a machine learning model, as described later in 224.

[0028]

[0034] As an example for illustrative purposes, a user query may include elemental information indicating a material containing iron (Fe) and oxygen (O). Referring again to Figure 3, material data 300 contains information about a material that matches the elemental information. Therefore, in response to the query, material data for material 302 (Fe2O3) can be obtained. As a second example for illustrative purposes, a user query may include material property information indicating a band gap in the range of 5.5 to 6.0 eV. Since material data 300 contains information about a material that matches the specified band gap, material data for material 304 (Al2O3) can be obtained.

[0029]

[0035] Returning to Figure 2, if it is determined that the material information database contains information on a material that matches the query, in step 208, method 200 optionally includes determining whether the uncertainty of the predicted material properties is within the uncertainty threshold. In some examples, the uncertainty threshold is a threshold selected by the user. In such some examples, the user query may include a threshold instruction, e.g., a threshold or a threshold percentage. If the uncertainty of the predicted material properties is within the threshold, method 200 includes outputting the material data in step 210. In some examples where no uncertainty threshold is provided, step 208 may be omitted.

[0030]

[0036] Returning to step 206, if it is determined that the material information database does not contain any records of a material that match the user query, method 200 includes generating one or more candidate structures using one or more trained generative machine learning models in 220. In some examples, method 200 may proceed to step 220 if the number of material records matching the query is below a threshold. In further examples, method 200 may include using a trained generative machine learning model to generate one or more candidate structures based on the user's request. The trained generative machine learning model is configured to generate candidate structures based on elemental information. In some examples, the trained generative machine learning model uses a skeletal structure to generate candidate structures. The skeletal structure contains structural information of one or more different crystal lattices, each containing lattice sites of atoms. In some examples, the skeletal structure includes charges associated with the lattice sites of the skeletal structure. As an example, a perovskite skeletal structure can be used to generate candidate structures of a material having the formula AMO3, where A and M are different metals. In other examples, any other suitable method for generating candidate structures can be used.

[0031]

[0037] Figure 4 shows exemplary material data 400 of candidate structures generated in response to a user query containing elemental information including combinations of elements A, Mx, My, and oxygen (O), where A represents an alkali metal and Mx and My represent different transition metals. In this example for explanation, no record is found in the material information database for a material matching the elemental information. Therefore, a trained generative machine learning model is used to generate candidate structures based on the elemental information. As a result, the trained generative machine learning model outputs candidate structures for each of the generated materials 402, 403, 404, and 405. Material data 400 also includes equations 410 and structural information 420 corresponding to the candidate structures of the generated materials 402, 403, 404, and 405.

[0032]

[0038] Returning to Figure 2, after generating one or more candidate structures, method 200 includes relaxing one or more candidate structures using a trained machine learning model in 222. The trained machine learning model used in 222 may include any suitable machine learning model as described above. In some examples, the trained machine learning model includes a GNN with ML-trained interatomic potentials. An example of a trained machine learning model configured to perform structural relaxation is described in Chen, C., Ong, SP, “A universal graph deep learning interatomic potential for the periodic table,” Nature Computational Science 2, 718-728 (2022). In other examples, any other suitable method may be used to perform structural relaxation on the candidate structures. Examples include energy minimization methods and QM calculations using interatomic potentials.

[0033]

[0039] After structural relaxation, Method 200 includes determining one or more predicted material properties for the relaxed candidate structure using one or more trained machine learning models. In some examples, the predicted material properties correspond to material property information in a user query. For example, if the user query indicates a band gap as a selected material property, the predicted band gap for the relaxed candidate structure can be determined using a trained machine learning model configured to predict the band gap. In some examples, Method 200 further includes determining the uncertainty of the predicted material properties. An example of determining the uncertainty of the predicted material properties is described above.

[0034]

[0040] Referring again to Figure 4, the material data 400 also shows the predicted material properties of the generated materials 402, 403, 404, and 405. To obtain the predicted material properties, structural information (e.g., graphs) of candidate structures corresponding to the generated materials 402, 403, 404, and 405 is input into a trained machine learning model configured to perform structural relaxation. Each relaxed candidate structure is then input into a trained machine learning model configured to output a predicted bandgap. Exemplary bandgap values ​​are shown as material property information 422, including the bandgap values ​​424 and corresponding uncertainties 426 for each generated material 402, 403, 404, and 405. The bandgap values ​​424 are for illustrative purposes only and do not represent actual results from the trained machine learning model.

[0035]

[0041] As described above, the material information database may, in some examples, contain partial data on a material. If the material information database contains structural information for a material that matches the user query, but lacks information about the selected material properties for that material, method 200 can proceed to 207 through 224. In this example, the structural information for the selected material can be input into a corresponding trained machine learning model configured to output predicted values ​​for the selected material properties.

[0036]

[0042] After determining the predicted material properties of the candidate structure, method 200 includes determining in 208 whether the uncertainty of the predicted material properties is within a threshold range. If "yes", method 200 includes outputting the material data of the candidate structure in 226. In some examples, the material data of the candidate structure is inserted into the material information database, as indicated by arrow 228. For example, material data 400 can be added to the material information database. In this way, material data for new materials generated in response to user queries can be added to the material information database.

[0037]

[0043] Returning to step 210, if it is determined that the uncertainty of the predicted material properties is not within the threshold range, method 200 optionally includes in step 230 performing a QM calculation to determine the QM calculated value of the selected material properties. Any suitable QM method as described above can be used. After performing the QM calculation, method 200 includes in step 232 outputting the QM calculated value of the selected material properties. The QM calculated value of the selected material properties can be added to the material information database, as indicated by arrow 228. In this way, the material information database can be enhanced with the QM calculated value of material properties. In some examples, the method includes forming a material record containing the QM calculated value and outputting the material record to a user account in the material information database.

[0038]

[0044] Furthermore, in some scenarios, users may require more certain values ​​for the generated materials than those predicted in previous machine learning predictions. Therefore, users can input a request to perform a QM calculation on the generated materials. Once the QM calculation is performed, the calculated material properties can be stored in the user's account on the interactive material discovery platform. In this way, users of the material information database can obtain the QM calculation values ​​upon request. Alternatively, the results of the QM calculation requested by the user can be output to the material information database.

[0039]

[0045] In some examples, method 200 includes outputting QM calculations of material properties in 234 to be used as training data for training a corresponding machine learning model. For example, the QM calculations of a band gap can be used to further train a machine learning model configured to predict the band gap of a material. In this way, the machine learning model can be improved in response to user queries.

[0040]

[0046] Figures 5A and 5B show a flowchart of an example of a method 500 that processes queries, including retrieving material data from a material information database containing predicted material properties. Method 500 can be performed on a computing system such as a remote computing system 104 configured to process user queries from a client computer.

[0041]

[0047] Referring to Figure 5A, Method 500 includes receiving a query in 502 that includes one or more elemental information and / or material property information. In some examples, Method 500 includes processing the query using a large language model. For example, the query may be input to the large language model to extract elemental information and / or material property information. Method 500 includes retrieving material data from a material information database based on the query in 504. The material data includes structural information for each material in a set of materials that match the query, and the set includes one or more materials. The material data further includes, for one or more materials in the set of materials, one or more predicted material properties determined using one or more trained machine learning models. Each machine learning model may be configured to determine a predicted value for the corresponding material property. In some examples, two or more machine learning models or an ensemble of machine learning models may be used to determine the predicted values ​​for the corresponding material properties. In some examples, in 506, the predicted material properties include one of the following: predicted bandgap, predicted conductivity, predicted bulk modulus, predicted shear modulus, predicted formation energy, predicted phonon density of states peak, predicted dielectric constant, or predicted refractive index. In some examples, in 508, one or more trained machine learning models include one or more graph neural networks.

[0042]

[0048] In some examples, method 500 includes determining structural information in 509 by inputting candidate structures into a trained machine learning model configured to perform structural relaxation on materials matching the query and output relaxed candidate structures. As discussed above, candidate structures can be generated using a trained generative machine learning model.

[0043]

[0049] In some examples, in 510, the material data includes uncertainty in the predicted material properties. In such examples, in 512, the uncertainty in the predicted material properties is determined by obtaining an ensemble of predicted values ​​for the predicted material properties from an ensemble of individual machine learning models, and estimating the uncertainty based on the ensemble of predicted values. In other examples, the uncertainty can be calculated using any other suitable method.

[0044]

[0050] Continuing to Figure 5B, method 500 optionally includes comparing the uncertainty of the predicted material properties of the selected material with an uncertainty threshold in 514. In some examples, method 500 includes determining in 516 that the uncertainty of the predicted material properties is greater than the uncertainty threshold. In response, method 500 includes performing a QM calculation on the selected material to determine the calculated QM values ​​for the selected material properties of the selected material.

[0045]

[0051] Method 500 further includes outputting material data in 520. In some examples, once the QM calculation is performed in 516, Method 500 includes outputting the QM calculated values ​​for the selected material properties in 522.

[0046]

[0052] As described above, in some examples, each predicted material property is predicted using a corresponding trained machine learning model. In some such examples, method 500 further includes in 524 further training the corresponding trained machine learning model for the selected material property using the QM calculations of the selected material property.

[0047]

[0053] Figure 6 shows a flowchart of another exemplary method 600 for processing queries, which includes generating candidate material structures using a machine learning model and predicting the material properties of the candidate structures. Method 600 can be performed on any suitable computing system, such as a remote computing system 104 configured to process user queries from a client computer.

[0048]

[0054] Method 600 includes receiving a query containing elemental information in 602. In some examples, Method 600 includes processing the query using a large language model. Method 600 further includes inputting elemental information into a trained generative machine learning model in 604 to generate multiple candidate structures based on the elemental information. In some examples, Method 600 includes determining in 606 that the material information database does not contain information for any material that matches the query before inputting elemental information into the trained generative machine learning model.

[0049]

[0055] Method 600 further includes, in 608, inputting the candidate structures into a second trained machine learning model configured to perform structural relaxation on one or more candidate structures generated in 604 and output the relaxed candidate structures. In some examples, Method 600 includes, in 610, performing a QM-based structural relaxation calculation on the relaxed candidate structures. Any suitable QM method, such as those listed above, can be used.

[0050]

[0056] Next, method 600 further includes in 612 inputting a relaxed candidate structure into a third trained machine learning model configured to output predicted material properties. In some examples, in 614, the third trained machine learning model includes an ensemble of trained machine learning models, and the method includes obtaining an ensemble of predicted values ​​for predicted material properties from the ensemble of machine learning models, and estimating uncertainty based on the ensemble of predicted values. In some examples, method 600 includes in 616 performing a QM calculation for a selected material to determine the QM calculated values ​​for a selected material property.

[0051]

[0057] Method 600 further includes outputting relaxed candidate structures and predicted material properties in 618. In some examples, Method 600 includes updating a material information database in 620 to include relaxed candidate structures and predicted material properties of the relaxed candidate structures. In examples where the QM calculation is performed in 616, Method 600 includes outputting the QM calculation values ​​for the selected material properties in 622. In some such examples, Method 600 includes outputting the QM calculation values ​​to a user account of an interactive material discovery platform in 624. In some examples, each predicted material property is predicted using a corresponding trained machine learning model. In some such examples, Method 600 includes further training the corresponding trained machine learning model for the selected predicted material properties using the QM calculation values ​​for the selected material properties in 626.

[0052]

[0058] In some embodiments, the methods and processes described herein can be linked to a computing system of one or more computing devices. Specifically, such methods and processes can be implemented as computer application programs or services, application programming interfaces (APIs), libraries, and / or other computer program products.

[0053]

[0059] Figure 7 schematically shows a simplified representation of a computing system 700 configured to provide all of the computing functions described herein. The computing system 700 may take the form of one or more personal computers, server computers, and data centers, for example. Remote computing system 104 is an example of a computing system 700.

[0054]

[0060] The computing system 700 includes a logical subsystem 702 and a storage subsystem 704. The computing system 700 may optionally include a display subsystem 706, an input subsystem 708, a communication subsystem 710, and / or other subsystems not shown in Figure 7.

[0055]

[0061] The logical subsystem 702 includes one or more physical devices configured to execute instructions. For example, the logical subsystem 702 may be configured to execute instructions that are part of one or more applications, services, programs, routines, libraries, objects, components, data structures, or other logical constructs. Such instructions may be implemented to perform a task, implement a data type, transform the state of one or more components, achieve a technical effect, or reach a desired result.

[0056]

[0062] The logical subsystem 702 may include one or more hardware processors configured to execute software instructions. In addition, or separately, the logical subsystem 702 may include one or more hardware or firmware devices configured to execute hardware or firmware instructions. The processors of the logical subsystem 702 may be single-core or multi-core, and the instructions executed thereon may be configured with respect to sequential, parallel, and / or distributed processing. Individual components of the logical subsystem 702 may be optionally distributed between two or more separate devices located separately and / or configured for cooperative processing. Aspects of the logical subsystem 702 may be virtualized and executed by remotely accessible networked computing devices configured within a cloud computing configuration.

[0057]

[0063] The storage subsystem 704 includes one or more physical devices configured to temporarily and / or permanently hold computer information, such as data and instructions, that can be executed by the logical subsystem 702. If the storage subsystem 704 includes two or more devices, those devices may be located together and / or separately. The storage subsystem 704 may include volatile, non-volatile, dynamic, static, read / write, read-only, random access, sequential access, location addressable, file addressable, and / or content addressable devices. The storage subsystem 704 may include removable devices and / or embedded devices. When the logical subsystem 702 executes an instruction, the state of the storage subsystem 704 may be changed, for example, to hold different data.

[0058]

[0064] The storage subsystem 704 may include removable devices and / or embedded devices. The storage subsystem 704 may, in particular, include optical memory (e.g., CD, DVD, HD-DVD, Blu-ray disc, etc.), semiconductor memory (e.g., RAM, EPROM, EEPROM, etc.), and / or magnetic memory. The storage subsystem 704 may include volatile, non-volatile, dynamic, static, read / write, read-only, random access, sequential access, position-addressable, file-addressable, and / or content-addressable devices.

[0059]

[0065] Aspects of the logic subsystem 702 and the memory subsystem 704 can be integrated into one or more hardware logic components. Such hardware logic components may include, for example, program-specific integrated circuits and application-specific integrated circuits (PASIC / ASIC), program-specific standard products and application-specific standard products (PSSP / ASSP), systems on a chip (SOC), and composite programmable logic devices (CPLD).

[0060]

[0066] It will be understood that the memory subsystem 704 includes one or more physical devices. However, aspects of the instructions described herein may instead be propagated by a communication medium (e.g., electromagnetic signals, optical signals, etc.) that is not held by a physical device for a finite duration.

[0061]

[0067] The logical subsystem 702 and the storage subsystem 704 may cooperate to instantiate one or more logical machines. As used herein, the term “machine” is used to collectively refer to any combination of hardware, firmware, software, instructions, and / or other arbitrary components that cooperate to provide the functionality of a computer. A machine may be instantiated by a single computing device, or it may include two or more subcomponents instantiated by two or more different computing devices. In some implementations, a machine includes local components (e.g., software applications executed by the computer’s processor) that cooperate with remote components (e.g., cloud computing services provided by a network of server computers). The software and / or other instructions that give a particular machine its functionality may optionally be stored as one or more unexecuted modules on one or more suitable storage devices.

[0062]

[0068] The terms “module” and “program” can be used to describe aspects of a computing system 700 implemented to perform a specific function. In some cases, a module, program, or engine may be instantiated by a logical subsystem 702 that executes instructions held by a storage subsystem 704. It will be understood that different modules and / or programs may be instantiated from the same application, service, code block, object, library, routine, API, function, etc. Similarly, the same module and / or program may be instantiated by different applications, services, code block, object, routine, API, function, etc. The terms “module” and “program” may encompass individual or grouped entities such as executable files, data files, libraries, drivers, scripts, database records, etc.

[0063]

[0069] As used herein, “service” is understood to mean an application program that can run across multiple user sessions. A service may be available to one or more system components, programs, and / or other services. In some implementations, a service may run on one or more server computing devices.

[0064]

[0070] If included, the display subsystem 706 may be used to present a visual representation of the data held by the storage subsystem 704. This visual representation may take the form of a graphical user interface (GUI). Since the methods and processes described herein modify the data held by the storage subsystem 704 and thus change the state of the storage subsystem 704, the state of the display subsystem 706 can also be changed to visually represent the changes in the underlying data. The display subsystem 706 may include one or more display devices utilizing virtually any kind of technology. Such display devices can be combined with the logical subsystem 702 and the storage subsystem 704 within a shared enclosure, or they can be peripheral display devices.

[0065]

[0071] If included, the input subsystem 708 may include or interface with one or more input devices such as a keyboard, mouse, or touchscreen. In some examples, the input subsystem 708 may include or interface with selected natural user input (NUI) components. Such components may be integrated or peripheral devices, and the conversion and / or processing of input actions may be handled onboard or offboard. Examples of NUI components may include microphones for conversation and / or speech recognition, and infrared, color, stereoscopic, and / or depth cameras for machine vision and / or gesture recognition. In some examples, the interface may include a large language model that can be used, for example, when processing user queries.

[0066]

[0072] If included, the communication subsystem 710 may be configured to connect the computing system 700 to one or more other computing devices in a communicative manner. The communication subsystem 710 may include wired and / or wireless communication devices compliant with one or more different communication protocols. In a non-limiting example, the communication subsystem 710 may be configured for communication over a wireless telephone network or a wired or wireless local area network or wide area network. In some examples, the communication subsystem 710 may enable the computing system 700 to send and receive messages with other devices over a network such as the Internet.

[0067]

[0073] Another example provides a method performed on a computing system. This method includes receiving a query containing one or more elemental and material property information. The method further includes retrieving material data from a material information database based on the query, the material data including structural information for each material in a set of materials matching the query, the set including one or more materials, and for one or more materials in the set of materials, one or more predicted material properties determined using one or more trained machine learning models. The method further includes outputting the material data. In some such examples, the method further includes determining whether the uncertainty of the predicted material properties of the selected material is greater than an uncertainty threshold, and in response, performing quantum mechanics (QM) calculations for the selected material to determine the QM calculations for the selected material properties of the selected material, and outputting the QM calculations for the selected material properties. In some such examples, each predicted material property is alternatively or additionally predicted using a corresponding trained machine learning model, and the method further includes using the QM calculations for the selected material properties to further train the corresponding trained machine learning models for the selected predicted material properties. In some such examples, the method alternatively or additionally includes determining structural information by inputting candidate structures into a trained machine learning model configured to perform structural relaxation and output relaxed candidate structures for materials matching the query. In some such examples, the predicted material properties alternatively or additionally include one of the following: predicted bandgap, predicted conductivity, predicted bulk modulus, predicted shear modulus, predicted formation energy, predicted phonon density of states peak, dielectric constant, or refractive index. In some such examples, one or more trained machine learning models alternatively or additionally include one or more graph neural networks. In some such examples, the material data alternatively or additionally includes the uncertainty of one of the predicted material properties among one or more predicted material properties. In some such examples, the uncertainty of the predicted material properties is alternatively or additionally determined by obtaining an ensemble of predicted values ​​for the predicted material properties from an ensemble of individual machine learning models, and estimating the uncertainty based on the ensemble of predicted values.In some such examples, receiving a query may, in addition, involve feeding the query into a large language model configured to extract one or more elemental or material property information.

[0068]

[0074] Another example provides a method performed on a computing system, which includes receiving a query containing elemental information and inputting the elemental information into a trained generative machine learning model to generate a plurality of candidate structures based on the elemental information. The method further includes inputting the candidate structures into a second trained machine learning model configured to perform structural relaxation on one or more of the plurality of candidate structures and output the relaxed candidate structures, inputting the relaxed candidate structures into a third trained machine learning model configured to output predicted material properties, and outputting the relaxed candidate structures and predicted material properties. In some such examples, the method further includes determining that, before inputting the elemental information into the trained generative machine learning model, the material information database does not contain information for any material that matches the query. In some such examples, the method alternatively or additionally includes updating the material information database to include the relaxed candidate structures and predicted material properties for the relaxed candidate structures. In some such examples, the third trained machine learning model optionally or additionally includes an ensemble of trained machine learning models, and the method includes obtaining an ensemble of predicted values ​​for predicted material properties from the ensemble of machine learning models, and estimating uncertainty based on the ensemble of predicted values. In some such examples, the method optionally or additionally includes performing quantum mechanical (QM) calculations of the selected material to determine the QM calculations for the selected material properties, and outputting the QM calculations for the selected material properties. In some such examples, each predicted material property is optionally or additionally predicted using the corresponding trained machine learning model, and the method further includes using the QM calculations for the selected material properties to further train the corresponding trained machine learning model for the selected predicted material properties.

[0069]

[0075] Another example provides a computing system including a logical subsystem and a memory subsystem that holds instructions executable by the logical subsystem to receive queries containing elemental information and to input the elemental information into a trained generative machine learning model to generate multiple candidate structures based on the elemental information. The instructions can further execute to input a candidate structure into a second trained machine learning model configured to perform structural relaxation on one or more of the multiple candidate structures and output the relaxed candidate structure, input the relaxed candidate structure into a third trained machine learning model configured to output predicted material properties, and output the relaxed candidate structure and predicted material properties. In some such examples, the instructions can further execute to determine that the material information database does not contain information for any material matching the query before inputting the elemental information into the trained generative machine learning model. In some such examples, the instructions can further execute, alternatively or additionally, to update the material information database to include the relaxed candidate structure and predicted material properties of the candidate structure. In some such examples, the instructions can further execute, alternatively or additionally, to input a query into a large language model configured to extract elemental information. In some such examples, the command may alternatively or additionally perform the following actions: receive user input requesting to perform QM calculations on relaxed candidate structures to determine QM calculated values ​​for selected material properties; form a material record containing structural information based on the relaxed candidate structures and the QM calculated values ​​for selected material properties; and output the material record for storage in the user's account.

[0070]

[0076] The configurations and / or methods described herein are essentially illustrative, and numerous modifications are possible; therefore, it should be understood that these particular embodiments or examples should not be considered in an restrictive sense. The specific routines or methods described herein may represent one or more of any number of processing strategies. Thus, the various actions illustrated and / or described may be performed in the order illustrated and / or described, in other orders, simultaneously, or omitted. Similarly, the order of the processes described above can be changed.

[0071]

[0077] The contents of this disclosure include all novel and non-trivial combinations and subcombinations of the various processes, systems, and configurations disclosed herein, as well as other features, functions, actions, and / or characteristics, and any and all equivalents thereof.

Claims

1. A method (200, 500) performed on a computing system, Receiving queries that include one or more elemental information and material property information (202, 502), Based on the aforementioned query, material data is obtained from the material information database (108, 204) (504), wherein the material data Structural information (310, 320, 410, 420) for each material in the set of materials that match the query, wherein the set includes one or more materials, and For one or more materials within the set of materials, one or more predicted material properties (324, 424) determined using one or more trained machine learning models (110). Including obtaining (504), and Output the aforementioned material data (210, 520). Methods including (200, 500).

2. Determining that the uncertainty of the predicted material properties of the selected material is greater than the uncertainty threshold. In response, quantum mechanical (QM) calculations are performed on the selected material to determine the QM calculated values ​​of the selected material properties of the selected material, and Output the QM calculated value of the selected material properties. The method according to claim 1, further comprising:

3. The method according to claim 2, further comprising predicting each predicted material property using a corresponding trained machine learning model, and further training the corresponding trained machine learning model for the selected predicted material property using the QM calculation value for the selected material property.

4. The method according to claim 1, further comprising determining the structural information by performing structural relaxation on materials that match the query and inputting candidate structures into a trained machine learning model configured to output relaxed candidate structures.

5. The method according to claim 1, wherein the predicted material properties include one of the following: predicted band gap, predicted conductivity, predicted bulk modulus, predicted shear modulus, predicted formation energy, predicted phonon density of states peak, dielectric constant, or refractive index.

6. The method according to claim 1, wherein the one or more trained machine learning models include one or more graph neural networks.

7. The method according to claim 1, wherein the material data includes the uncertainty of one of the one or more predicted material properties.

8. The method according to claim 7, comprising obtaining an ensemble of predicted values ​​for the predicted material properties from an ensemble of individual machine learning models, and determining the uncertainty of the predicted material properties by estimating the uncertainty based on the ensemble of predicted values.

9. The method according to claim 1, wherein receiving the query includes inputting the query into a large-scale language model configured to extract one or more of the elemental information or material property information.

10. A method (200, 600) performed on a computing system, Receiving queries containing elemental information (202, 602), In order to generate multiple candidate structures based on the elemental information, the elemental information is input into a trained generative machine learning model (110) (220, 604), and With respect to one or more of the aforementioned candidate structures, Inputting the candidate structure into a second trained machine learning model (110) configured to perform structural relaxation and output the relaxed candidate structure (222, 608), Inputting the relaxed candidate structures into a third trained machine learning model (110) configured to output predicted material properties (224, 612), and Output the relaxed candidate structure and the predicted material properties (210, 618) Methods including (200, 600).

11. The method according to claim 10, further comprising determining that, before inputting the elemental information into the trained generative machine learning model, the material information database does not contain information for any material that matches the query.

12. The method according to claim 11, further comprising updating the material information database to include the relaxed candidate structure and the predicted material properties of the relaxed candidate structure.

13. The third trained machine learning model includes an ensemble of trained machine learning models, and the method is Obtaining an ensemble of predicted values ​​for the predicted material properties from the ensemble of the aforementioned machine learning models, and Estimating uncertainty based on the ensemble of the aforementioned predicted values. The method according to claim 10, including the method described in claim 10.

14. Perform quantum mechanical (QM) calculations on the selected material to determine the QM calculated values ​​of the selected material properties, and Output the QM calculated value of the selected material properties. The method according to claim 10, further comprising:

15. The method according to claim 14, further comprising predicting each predicted material property using a corresponding trained machine learning model, and further training the corresponding trained machine learning model for the selected predicted material property using the QM calculation value for the selected material property.

16. Logical subsystem (702) and, Receiving queries containing elemental information (202, 602), In order to generate multiple candidate structures based on the elemental information, the elemental information is input into a trained generative machine learning model (110) (220, 604), and With respect to one or more of the aforementioned candidate structures, Inputting the candidate structure into a second trained machine learning model (110) configured to perform structural relaxation and output the relaxed candidate structure (222, 608), Inputting the relaxed candidate structures into a third trained machine learning model (110) configured to output predicted material properties (224, 612), and Output the relaxed candidate structure and the predicted material properties (210, 618) To perform this, the logical subsystem (702) has a storage subsystem (704) that holds instructions that can be executed by the logical subsystem (702) and A computing system including (104,700).

17. The computing system according to claim 16, wherein, before inputting the elemental information into the trained generative machine learning model, the instruction can be further executed to determine that the material information database does not contain information for any material that matches the query.

18. The calculation system according to claim 17, wherein the instruction is further executable to update the material information database to include the relaxed candidate structure and the predicted material properties of the candidate structure.

19. The computational system according to claim 16, wherein the instructions are further executable to input the query into a large-scale language model configured to extract the elemental information.

20. The system receives input from a user requesting that a QM calculation be performed on the relaxed candidate structure to determine the QM calculated values ​​for the selected material properties. A material record including structural information is formed based on the relaxed candidate structure and the QM calculated values ​​of the selected material properties. Output the material record to store it in the user's account. The calculation system according to claim 16, wherein the instruction is further executable for the purpose of the calculation.

Images