Artificial intelligence method for material property optimization and new material generation

Abstract

A computer-implemented, machine learning method for material property optimization. A dataset for a material is obtained that includes a molecule shape description, a molecular weight, and a target material property for the material. The material is encoded to a computing format including a three dimensional (3D) graph representation of a main molecule for the material, a 3D graph representation of a side chain, the molecular weight, and a degree of substitution of the side chain. The 3D graph representations are encoded by a neural network that includes one or more graph attention layers each into a latent representation usable for the material property optimization. The method has applications including, but not limited to, use cases in medicine / healthcare (e.g., for AI assisted drug design), molecular or material design and development, and chemical engineering.

Description

ARTIFICIAL INTELLIGENCE METHOD FOR MATERIAL PROPERTY OPTIMIZATIONAND NEW MATERIAL GENERATIONCROSS-REFERENCE TO PRIOR APPLICATION

[0001] Priority is claimed to U.S. Provisional Application Serial No. 63 / 542,787 filed on October 6, 2023, the entire contents of which is hereby incorporated by reference herein.FIELD

[0002] The present invention relates to Artificial Intelligence (Al) and machine learning (ML), and in particular to a method, system, data structure, computer program product and computer-readable medium for material property optimization and new material generation. BACKGROUND

[0003] Material property enhancement is a critical endeavor in materials science and engineering, driven by the ever-expanding demands of modem technology and industry. The development of materials with superior mechanical, thermal, electrical, and optical properties requires to overcome complex technical challenges. A fundamental obstacle lies in the vast search space of possible material compositions and structures. Moreover, the experimental evaluation of these materials is often a time-consuming and resource-intensive process. Consequently, there is a growing effort with computational and predictive methods to accelerate the discovery and optimization of materials.

[0004] U.S. Patent Application Publication No. 2020 / 0257933 Al describes a framework that applies machine learning to the identification of alloys or composite with desired properties of interest. The framework includes a selection algorithm for the application of gold standard regression techniques and a physics-based model for the prediction of tensile strength that also provides additional prior information. Gomez-Bombarelli, R., et al., “Automatic chemical design using a data-driven continuous representation of molecules,” arXiv. 1610.02415 (2016) describe a method for material generation, in which a model takes the molecule in accordance with simplified molecular-input line-entry system (SMILES) notation and trains a variational autoencoder with an additional property prediction branch. The encoder and decoder are recurrent neural networks. Wu, S., et al., “Machine-leaming-assisted discovery of polymers with high thermal conductivity using a molecular design algorithm,” npj Computational Materials 5, 66 (2019) discuss molecular design algorithm utilized machine learning models trained on a restricted dataset of polymeric properties and implemented with a transfer learning approach. It generated hypothetical polymer structures based on SMILES representations and incorporated Bayesian methods to enhance the selection of potential polymer candidates. Zhu, M.X., et al., “Machine-leaming-driven discovery of polymers molecular structures with high thermalconductivity,” International Journal of Heat and Mass Transfer, vol. 162(32): 120381 (2020) discuss an approach involving "fingerprinting" polymers into numerical representations, referred to as an extended connectivity fingerprint (ECFP), which recognize its structural features. ECFP is used for building quantitative structure -property relationship model analyzing key correlations between the crystallization temperature (Tc) and polymer structure. Each of the foregoing references is hereby incorporated by reference herein.SUMMARY

[0005] In an embodiment, the present invention provides a computer-implemented, machine learning method for material property optimization. A dataset for a material is obtained. The dataset includes a molecule shape description for the material, a molecular weight for the material, and a target material property for the material. The material is encoded to a computing format including a three dimensional (3D) graph representation of a main molecule for the material, a 3D graph representation of a side chain for the main molecule, the molecular weight, and a degree of substitution of the side chain. At least the 3D graph representation of the main molecule and the 3D graph representation of the side chain are encoded by a neural network that includes one or more graph attention layers each into a latent representation usable for the material property optimization. The method has applications including, but not limited to, use cases in medicine (e.g., digital medicine, personalized healthcare, Al-assisted drug or vaccine development, diagnosis or treatment, disease prediction, etc.), molecular or material design and development, and chemical engineering.BRIEF DESCRIPTION OF THE DRAWINGS

[0006] Embodiments of the present invention will be described in even greater detail below based on the exemplary figures. The present invention is not limited to the exemplary embodiments. All features described and / or illustrated herein can be used alone or combined in different combinations in embodiments of the present invention. The features and advantages of various embodiments of the present invention will become apparent by reading the following detailed description with reference to the attached drawings which illustrate the following:

[0007] FIG. 1 schematically illustrates a computer-implemented method and computer system for guided material property optimization and new material generation according to an embodiment of the present invention;

[0008] FIG. 2 schematically illustrates a method and system for molecule property prediction according to an embodiment of the present invention;

[0009] FIG. 3 schematically illustrates a molecule encoder and a method for encoding molecules according to an embodiment of the present invention;

[0010] FIG. 4 illustrates a flow diagram of a method for guided material property optimization and new material generation according to an embodiment of the present invention; and

[0011] FIG. 5 is a block diagram of an exemplary processing system, which can be configured to perform any and all operations disclosed herein.DETAILED DESCRIPTION

[0012] Embodiments of the present invention provide a computer-implemented method and computer system for guided material property optimization and new material generation that introduce a new mechanism to describe materials in a computable format, a two-phase optimization process, and a new model with an inventive graph attention layer that allows to encode 3D molecules.

[0013] In an embodiment, the present invention provides a computer-guided method for material property optimization and new material generation which is achieved by: 1) modifying an existing material such that the property of interest is optimized, or 2) by sampling a random material. In contrast to existing technology, embodiments of the present invention provide an approach for representing materials in a computational format that surpasses the existing methods by encoding 3D molecule structures together with other features, by the design of a deep neural network with a new layer type that is referred to herein as “molecule graph attention”, and a two-step optimization process based on the proposed material encoding and the model. During the first phase, it takes an existing material as a starting point (or a randomly sampled material) and starts to modify it such that the property of interest is optimized (increased, decreased or reaching a target value). Then, in the second phase, the resulting predicted material is modified such that it satisfies the physical and chemical constrains of a real material. Embodiments of the present invention provide improvements over existing technology by allowing for the utilization of more complete information by the usage of 3D molecule structures, and the possibility of optimizing existing material, which facilitate the creation of new materials that are similar to existing ones and therefore are easier to introduce in the manufacturing process.

[0014] In an embodiment, the present invention provides a system to generate new materials that enhance a selected material property (e.g., glass transition temperature (Tg), melting temperature (Tm), elongation, etc.) starting from an existing material. The system implements three steps, each of which provides for enhanced computer functionality over existing technology. The first step is a two-phase material optimization process, the second step is the material representation, and the third step is application of the molecule attention layer.

[0015] In a first aspect, the present invention provides a computer-implemented, machine learning method for material property optimization. A dataset for a material is obtained. The dataset includes a molecule shape description for the material, a molecular weight for the material, and a target material property for the material. The material is encoded to a computing format including a three dimensional (3D) graph representation of a main molecule for the material, a 3D graph representation of a side chain for the main molecule, the molecular weight, and a degree of substitution of the side chain. At least the 3D graph representation of the main molecule and the 3D graph representation of the side chain are encoded by a neural network that includes one or more graph attention layers each into a latent representation usable for the material property optimization.

[0016] In a second aspect, the present invention provides the method according to the first aspect, further comprising training a model on top of a latent space including the latent representations, and using the trained model for the material property optimization to generate a new material.

[0017] In a third aspect, the present invention provides the method according to the first aspect or second aspect, further comprising optimizing the target material property to generate a virtual intermediate material based on the latent representation.

[0018] In a fourth aspect, the present invention provides the method according to any of the first to third aspects, further comprising fixing the virtual intermediate material based on constraints that correspond to physical properties and chemical properties of the material to generate a refined material that satisfies the physical properties and the chemical properties of the material, the refined material corresponding to a new material derived from the material.

[0019] In a fifth aspect, the present invention provides the method according to any of the first to fourth aspects, wherein fixing the virtual intermediate material occurs iteratively by computing a loss between a current virtual intermediate material and a set of the physical properties and the chemical properties until the loss is 0 or the current virtual intermediate material satisfies the physical properties and the chemical properties of the material.

[0020] In a sixth aspect, the present invention provides the method according to any of the first to fifth aspects, further comprising optimizing, iteratively, the virtual intermediate material based on a gradient of a derivation associated with the virtual intermediate material and a loss objective, wherein the optimization of the virtual intermediate material occurs iteratively based on a comparison of a predicted material property and the target material property, or for a predefined number of iterations.

[0021] In a seventh aspect, the present invention provides the method according to any of the first to sixth aspects, wherein the computing format further includes two or more 3D graphrepresentations of side chains for the main molecule and two or more degrees of substitution of the side chains.

[0022] In an eighth aspect, the present invention provides the method according to any of the first to seventh aspects, wherein the neural network includes one or more molecule encoders and one or more numerical encoders configured to, in parallel, encode the 3D graph representation of the main molecule, the 3D graph representation of the side chain, the molecular weight, and the degree of substitution into the latent representations.

[0023] In a ninth aspect, the present invention provides the method according to any of the first to eighth aspects, wherein the 3D graph representation of the main molecule and the 3D graph representation of the side chain comprise atom types, 3D atom positions, and an adjacency matrix.

[0024] In a tenth aspect, the present invention provides the method according to any of the first to ninth aspects, wherein the one or more molecule encoders are configured to convert the atom types to the latent representations using an embedding layer.

[0025] In an eleventh aspect, the present invention provides the method according to any of the first to tenth aspects, wherein the one or more numerical encoders are configured to convert the 3D atom positions to the latent representations using a dense layer.

[0026] In a twelfth aspect, the present invention provides the method according to any of the first to eleventh aspects, wherein the one or more molecule encoders are further configured to combine, by concatenation, the latent representations of the atom types and the latent representations of the 3D atom positons to generate a concatenated latent representation.

[0027] In a thirteenth aspect, the present invention provides the method according to any of the first to twelfth aspects, wherein the neural network is a differentiable deep neural network that includes the one or more graph attention layers, which are configured to combine the concatenated latent representation, a set of latent prototypes, and the adjacency matrix.

[0028] In a fourteenth aspect, the present invention provides a computer system for material property optimization comprising one or more processors which, alone or in combination, are configured to perform a machine learning method for material property optimization according to any of the first to thirteenth aspects.

[0029] In a fifteenth aspect, the present invention provides a tangible, non-transitory computer-readable medium for material property optimization which, upon being executed by one or more hardware processors, provide for execution of a machine learning method according to any of the first to thirteenth aspects.Two-Phase Material Optimization Process:

[0030] To optimize a selected property of a given material, embodiments of the present invention provide a two-phase refinement pipeline. FIG. 1 is a high-level diagram illustrating the workflow 100 of a two-phase refinement pipeline according to an embodiment of the present invention. The process starts with the selection of a material of interest (material to optimize) 102, and the target property that is to be optimized as the starting point. An example of a practical material is cellulose triacetate and Tg as the selected property. The optimization process is carried out in two phases as follows.Phase 1 : Property Optimization

[0031] In this phase, the material is virtually modified such that the selected property is enhanced. At first, the intermediate virtual material 104, is equal to the selected material 102 to optimize. Then, the intermediate virtual material 104 is sent to the molecule property prediction module 106 according to an embodiment of the present invention, which will perform an estimation of the selected property for the current virtual material 104.

[0032] The virtual material 104 is then modified based on the resulting gradient of the derivation 108 with respect to the intermediate virtual material 104 and the loss objective. The intermediate virtual material 104 is defined as x, where x is the combination of the 3D graph of the main chain or the material 104, the 3D graph of the first side chain, the 3D graph of the second side chain, the degree of substitution of the first side chain, the degree of substitution of the second side chain, and the molecular weight (see boxes a in FIGS. 2 and 3). The loss objective could be any derivable function that can measure the distance between the current estimation of the material property, and the desired property. The following is an example of a suitable objective function:where y is the predicted material property, and t is the target material property. The gradients g are computed as follows:9 = Vz(£(y|t)).The updates are then computed as follows:where I is a learning rate that controls the scale of the updates. Finally, the updated intermediate material x is computed as follows:X = X + u.

[0033] The procedure that updates the intermediate virtual material 104 is repeated for a predefined number of iterations or until the estimated property is close enough to the target value (i.e., lower than e value):|y - t| < e.

[0034] During this process, the material 104 can achieve impossible shapes and configurations. Therefore, to achieve a realistic candidate, the second phase performs a material rectification process that converts the material 104 back to a plausible candidate.Phase 2: Material Rectification

[0035] The intermediate virtual material is an output of the first phase. This estimated material (intermediate virtual material 104) should have estimated properties that are very close to the desired target, however, it might violate certain physical and chemical properties. Therefore, the second phase is focused on fixing these artifacts by further refining the intermediate virtual material 104. The effect is achieved by computing the loss between the current virtual material, and a set of predefined rules and constraints 110 that any material should fulfill. Examples of these constraints are:For the polysaccharide case, the sum of the substation degree must be lower than 3 : Lsc= max(SCl + SC2 - 3, 0)2There are no negative values v (where v can be any input value):The adjacency matrix cannot contain isolated elements. Therefore, the summation of each row needs to be greater than 2 (this include the element of the main diagonal).

[0036] The virtual material is iteratively modified following the same gradient strategy (108) that was applied in the first phase, but replacing £(y|t) by the summation of all the individual constraint losses that were just introduced. The iterations should stop once the loss is 0, or the molecule has fulfilled all the physical and chemical constraints thereby resulting in a refined material 112.Molecule Property Prediction:

[0037] To perform the material property enhancing process described above, the materials are converted into a computable format that can be fed to a derivable machine learning model. The processing method to express the material into a computable format and the designed machine learning model each represent further improvements over existing technology according to embodiments of the present invention.Material Representation:

[0038] FIG. 2 schematically illustrates a method and system for molecule property prediction according to an embodiment of the present invention. FIG. 2 includes workflow 200 for molecule property prediction according to an embodiment of the present invention. In orderto perform material property predictions with any machine learning model, embodiments of the present invention provide an inventive method to describe the physical concept of the material in a computable format. In particular, embodiments of the present invention provide to describe a given material as the combination of the 3D graph representation of the main chain molecule 202, the 3D graph representation of the side chains 204, and numerical properties such as the molecular weight 206, and the degree of substitution for the side chains 208 (see box a in FIG. 2):Main chain 202: The main chain contains the 3D molecular graph that represents the core of the molecule. This feature is provided for all the materials.Side chains 204 (1 and 2): The side chains are molecules that are typically stochastically attached to the main chain 202. Depending on the material, there could be none, or multiple side chains, therefore they are not required in all embodiments. In an exemplary embodiment, the present invention considers two possible side chains 204.Numerical features 206 and 208: These are typically numerical features that gather information about the whole molecule. In this set of features, the following features are considered:Molecular weight 206: This feature contains information about the size of the entire molecule.Substitution degrees 208: This feature shows the substitution rate of the side chain 204 in the main chain 202. In some materials, the attachment of the side chains 204 into the main chain 202 occurs stochastically. This feature acts as a parameter by which the ratio or attachment probability of each side chain 204 is modeled.Material Modeling:

[0039] To perform the material property prediction, the material representation discussed above is modeled with the model according to an embodiment of the present invention. Box b in FIG. 2 contains a block diagram that illustrates the information flow (workflow 200) of the model according to an embodiment of the present invention. The model encodes all the inputs in parallel. Each input is encoded by independent modules that project the inputs into latent representations. The latent representations are then merged by the feature merger 210 which converts them into a fixed length latent representation. The resulting representation is now fed to the property prediction module 212, and the estimated property for the material is produced 214. In embodiments the property prediction module 212 is a set of layers which function to map the merged features from the feature merger 210 into the predicted property 214. The layers my include a set of dense (or other layers), batch normalization, dropout, etc. The information of therepresentations is projected into latent spaces until you obtain an output layer in which at least one of the outputs is the predicted value. In detail:Molecule encoder 216: FIG. 3 illustrates a flow diagram of this module 216. The numerical encoders 218 are converted into latent representations through an embedding layer. The 3D graph molecule (box a in FIG. 3) is composed of the atom types 300, the 3D atom positions 302, and the adjacency matrix 304. The atom types 300 are converted into latent representations though an embedding layer 306. The 3D atom positions 302 are also converted into latent representations with a simple dense layer 308. Both representations are now combined with a simple concatenation operation 310. This representation is now used in the molecule graph attention layer 312 (see box b in FIG. 3). This layer 312 takes inputs as a set of latent prototypes 314 that are learned during the training process, the concatenated latent representations produced in the previous step (from concatenation 310), and the adjacency matrix 304. In embodiments the latent prototypes 314 are learned together with all the model parameters (weights). The prototypes 314 may be regularized with a KL-Divergence or a Chamfer distance between them and the resulting vector output from 308. Then the information is combined as follows:V dim (1) s' = softmax S), v' = s'@v, I’ = v’ + q, where p, I, and A are the latent prototype 314, the latent representation coming from the atom types 300 and the positions 302, and the adjacency matrix 304 respectively, and @ stands for the dot product operation. After I’ is computed, they are fed into a dense layer, a batch normalization layer, and a dropout layer. These layers may be a modified version of an attempting layer which is typically used in transformers (LLMs). The addition of the adjacency matrix 304 is used in the computation of the intermediate tensors. The adjacency matrix 304 can help mask the responses between the connected nodes. The meaning of q is query, k stands for key, v is called value, s are scores, and 1 stands for latents (which is the final representations). The variables q, k, and v may measure similarities between two sets, and merge them to get a new representation.The output of the graph attention layer 312 is fed to a masked dense layer 316. The masked dense layer 316 may equate all the latent representations to zero in which there is noconnection in the graph. With this operation we further reinforce the presence of the graph representation by forcing the resulting latent I’ to be masked by the adjacency matrix by:Finally, the output is fed to an additional molecule graph attention layer 318 (as discussed above). The molecule graph attention layer 318 may provide the model more freedom to fit the data.

[0040] Embodiments of the present invention thus provide for general improvements to computers in machine learning systems to provide for improved property optimization of 3D molecules and structures. Moreover, embodiments of the present invention can be practically applied to use cases to effect further improvements in a number of technical fields including, but not limited to, medical (e.g., digital medicine, personalized healthcare, Al-assisted drug or vaccine development, diagnosis or treatment, disease prediction, etc.), molecular or material design and development, chemical engineering, etc. Depending on the application different encoding methods may need to be adapted to the features of the medical data.

[0041] In an exemplary embodiment, the present invention can be applied to a use case for property prediction for polymers. With the ability to anticipate material behavior and properties accurately, researchers can make informed decisions which materials to prioritize for synthesis, reducing the need for trial and error and saving expert, physical and computational resources, and providing for more accurate results. The data source for this use case includes: 1) a database of annotated target property values and structural information of samples, and 2) corresponding 3D mol files obtained from structural information. Application of the method according to an embodiment of the present invention provides that:Property data and structural data for polysaccharide esters are prepared.Each molecule is divided into its main-chain and side-chain components, which are separately input into the system in the form of 3D mol files. The 3D mol file format, includes information about atom information, adjacency information, and 3D spatial coordinates.To retain additional information that may be lost during the division of the molecule, such as degree of substitution (DS) for the side chain(s) or molecular weight, the encoder module of the autoencoder processes the individual inputs for the main chain and side chain(s), capturing the distinct features of each component. Additionally, the supplementary structurespecific data is fed into the network to ensure that critical structural information is retained.Following the encoding process, the latent space captures the merged features of both the main chain and side chain(s), as well as the supplementary structural details. This latent space, which represents the consolidated structural information, serves as the foundation for the prediction of polymer properties. A regression model trained on top of this latent spaceefficiently learns the intricate relationships between these features and the corresponding properties, resulting in a robust and accurate prediction mechanism.

[0042] As a result of application of the method, a highly accurate prediction model for polymer properties is obtained as output. The integration of 3D mol files and the comprehensive structural information elevates the precision of the prediction process, providing improved accuracy and ensuring better predictive outcomes. As technicity, by ingeniously combining 3D molecule data, and incorporating the supplementary structural data within the framework, this embodiment establishes a mechanism that can help a researcher make a more informed and accurate decision to prioritize material synthesis, reducing the need for trial and error, and saving expert, physical and computational resources. For example, more accurate predictions can be provided for materials with optimized target properties. Laboratory equipment could be adjusted in an automated or semi-automated manner to synthesize predicted materials.

[0043] In an exemplary embodiment, the present invention is applied for adaptive chemical structure engineering via latent space manipulation. By employing the approach according to an embodiment of the present invention, researchers gain insights into the specific chemical structural modifications required to achieve desired material properties. This enables the identification of critical partial structures that significantly influence the targeted property. The data source for this use case includes: 1) a database of annotated target property values and structural information of samples, and 2) corresponding 3D mol files obtained from structural information. Application of the method according to an embodiment of the present invention provides that the latent space, representing the amalgamation of critical structural attributes from the main-chain and side-chain components, acts as a springboard for property manipulation. By the method, controlled perturbations are introduced to the latent space, leading to corresponding modifications in the chemical structure. The output of the method includes modified chemical structures that align with the desired properties and insights for identification of critical partial structures that significantly influence the targeted property. As technicity, the approach enables the direct engineering of structures by interpreting their latent space representations. This provides researchers information about the chemical structural adjustments needed to attain desired material properties. Laboratory equipment could be adjusted in an automated or semiautomated manner based on the predictions.

[0044] In an embodiment, the present invention provides a computer-implemented method for guided material optimization and new material generation, the method comprising the steps of1) Creating (or acquiring) a dataset of materials that contain the molecule shape description- which includes the 3D graph representations of the main molecule and the side chains, the molecular weight or the molecule size, and the target material property.2) Encoding the material to a computing format that includes: a. The 3D graph representation of the main molecule. b. The 3D graph representations of the side chains. c. The molecular weight. d. The degree of substitution of the side chains.3) Encoding the 3D molecule graphs using a fit neural network model (e.g., a trained neural network) that includes molecule graph attention layers.4) In some embodiments, implementing a two-phase optimization process that: a. First, changes the material to optimize the target material property. During this process virtual intermediate materials are created. b. Second, fixes the intermediate virtual material such that it respects the physical and chemical properties of a real material.

[0045] Embodiments of the present invention provide for the following improvements and technical advantages over existing technology:1) Representing materials based on their 3D structural graph of the main chain, the side chains, the molecular weight, and the degree of substitution of the side chains. In particular, embodiments of the present invention provide for the division of the molecule into the six components.2) Providing a molecule graph attention layer that is used to build a deep neural network model that can effectively encode materials and is differentiable.3) Providing a two-phase material optimization procedure that comprises: a. A material property optimization phase that produces virtual intermediate materials based on a gradients method. b. A material rectification phase that fixes possible artifacts of the resulting virtual intermediate material such that it satisfies the physical and chemical constrains of a real material.

[0046] Thus, embodiments of the present invention enhance computer functionality and also provide improvements to the fields of molecular structure and / or material design and development including:1) Introducing a computer-guided method for material generation through tailored modifications of existing materials or by exploring random material samples.2) Providing to represent materials in a computational format, incorporating 3D molecule structures and additional features using a specialized deep neural network architecture comprising the molecule graph attention layer.3) Optimizing materials in a two-step process, initially enhancing the property of interest and then refining the resulting material to adhere to physical and chemical constraints, ensuring real -world applicability.

[0047] These improvements allow to enhance the utilization of comprehensive information, particularly leveraging 3D molecule structures, allowing for easier introduction of new materials into the manufacturing process compared to existing technology. Also, the division of the molecule into main chains and side chains in the approach according to embodiments of the present invention takes into account modifications that are practically achievable and relevant to real-world chemical processes and synthesis.

[0048] FIG. 4 illustrates a flow diagram of a method for guided material property optimization and new material generation according to an embodiment of the present invention. Flow diagram 400 includes obtaining a dataset for a material including a target material property at step 402. In embodiments the dataset includes a molecule shape description for the material, a molecular weight for the material, and the target material property for the material. The flow diagram 400 includes at step 404 encoding the material to a computing format including a 3D graph representation of a main molecule for the material and a 3D graph representation of a side chain for the material, the molecular weight, and a degree of substitution of the side chain. The flow 400 includes at step 406 encoding at least the 3D graph representations into a latent representation. The encoding may be executed by a neural network that includes one or more graph attention layers for encoding the 3D graph representation of the main molecule and the 3D graph representation of the side chain into latent representations that are usable for the material property optimization.

[0049] Referring to FIG. 5, a processing system 500 can include one or more processors 502, memory 504, one or more input / output devices 506, one or more sensors 508, one or more user interfaces 510, and one or more actuators 512. Processing system 500 can be representative of each computing system disclosed herein.

[0050] Processors 502 can include one or more distinct processors, each having one or more cores. Each of the distinct processors can have the same or different structure. Processors 502 can include one or more central processing units (CPUs), one or more graphics processing units (GPUs), circuitry (e.g., application specific integrated circuits (ASICs)), digital signal processors (DSPs), and the like. Processors 502 can be mounted to a common substrate or to multiple different substrates.

[0051] Processors 502 are configured to perform a certain function, method, or operation (e.g., are configured to provide for performance of a function, method, or operation) at least when one of the one or more of the distinct processors is capable of performing operations embodying the function, method, or operation. Processors 502 can perform operations embodying the function, method, or operation by, for example, executing code (e.g., interpreting scripts) stored on memory 504 and / or trafficking data through one or more ASICs. Processors 502, and thus processing system 500, can be configured to perform, automatically, any and all functions, methods, and operations disclosed herein. Therefore, processing system 500 can be configured to implement any of (e.g., all of) the protocols, devices, mechanisms, systems, and methods described herein.

[0052] For example, when the present disclosure states that a method or device performs task “X” (or that task “X” is performed), such a statement should be understood to disclose that processing system 500 can be configured to perform task “X”. Processing system 500 is configured to perform a function, method, or operation at least when processors 502 are configured to do the same.

[0053] Memory 504 can include volatile memory, non-volatile memory, and any other medium capable of storing data. Each of the volatile memory, non-volatile memory, and any other type of memory can include multiple different memory devices, located at multiple distinct locations and each having a different structure. Memory 504 can include remotely hosted (e.g., cloud) storage.

[0054] Examples of memory 504 include a non-transitory computer-readable media such as RAM, ROM, flash memory, EEPROM, any kind of optical storage disk such as a DVD, a Blu- Ray® disc, magnetic storage, holographic storage, a HDD, a SSD, any medium that can be used to store program code in the form of instructions or data structures, and the like. Any and all of the methods, functions, and operations described herein can be fully embodied in the form of tangible and / or non-transitory machine-readable code (e.g., interpretable scripts) saved in memory 504.

[0055] Input-output devices 506 can include any component for trafficking data such as ports, antennas (i.e., transceivers), printed conductive paths, and the like. Input-output devices 506 can enable wired communication via USB®, DisplayPort®, HDMI®, Ethernet, and the like. Input-output devices 506 can enable electronic, optical, magnetic, and holographic, communication with suitable memory 506. Input-output devices 506 can enable wireless communication via WiFi®, Bluetooth®, cellular (e.g., LTE®, CDMA®, GSM®, WiMax®, NFC®), GPS, and the like. Input-output devices 506 can include wired and / or wireless communication pathways.

[0056] Sensors 508 can capture physical measurements of environment and report the same to processors 502. User interface 510 can include displays, physical buttons, speakers, microphones, keyboards, and the like. Actuators 512 can enable processors 502 to control mechanical forces.

[0057] Processing system 500 can be distributed. For example, some components of processing system 500 can reside in a remote hosted network service (e.g., a cloud computing environment) while other components of processing system 500 can reside in a local computing system. Processing system 500 can have a modular design where certain modules include a plurality of the feature s / functions shown in FIG. 5. For example, I / O modules can include volatile memory and one or more processors. As another example, individual processor modules can include read-only-memory and / or local caches.

[0058] While subject matter of the present disclosure has been illustrated and described in detail in the drawings and foregoing description, such illustration and description are to be considered illustrative or exemplary and not restrictive. Any statement made herein characterizing the invention is also to be considered illustrative or exemplary and not restrictive as the invention is defined by the claims. It will be understood that changes and modifications may be made, by those of ordinary skill in the art, within the scope of the following claims, which may include any combination of features from different embodiments described above.

[0059] The terms used in the claims should be construed to have the broadest reasonable interpretation consistent with the foregoing description. For example, the use of the article “a” or “the” in introducing an element should not be interpreted as being exclusive of a plurality of elements. Likewise, the recitation of “or” should be interpreted as being inclusive, such that the recitation of “A or B” is not exclusive of “A and B,” unless it is clear from the context or the foregoing description that only one of A and B is intended. Further, the recitation of “at least one of A, B and C” should be interpreted as one or more of a group of elements consisting of A, B and C, and should not be interpreted as requiring at least one of each of the listed elements A, B and C, regardless of whether A, B and C are related as categories or otherwise. Moreover, the recitation of “A, B and / or C” or “at least one of A, B or C” should be interpreted as including any singular entity from the listed elements, e.g., A, any subset from the listed elements, e.g., A and B, or the entire list of elements A, B and C.

Claims

CLAIMSWhat is claimed is:

1. A computer-implemented method for material property optimization, the method comprising: obtaining a dataset for a material, the dataset including a molecule shape description for the material, a molecular weight for the material, and a target material property for the material; encoding the material to a computing format including a three dimensional (3D) graph representation of a main molecule for the material, a 3D graph representation of a side chain for the main molecule, the molecular weight, and a degree of substitution of the side chain; and encoding at least the 3D graph representation of the main molecule and the 3D graph representation of the side chain, by a neural network that includes one or more graph attention layers, each into a latent representation usable for the material property optimization.

2. The computer-implemented method of claim 1, further comprising training a model on top of a latent space including the latent representations, and using the trained model for the material property optimization to generate a new material.

3. The computer-implemented method of claim 1, further comprising optimizing the target material property to generate a virtual intermediate material based on the latent representation.

4. The computer-implemented method of claim 3, further comprising fixing the virtual intermediate material based on constraints that correspond to physical properties and chemical properties of the material to generate a refined material that satisfies the physical properties and the chemical properties of the material, the refined material corresponding to a new material derived from the material.

5. The computer-implemented method according to claim 4, wherein fixing the virtual intermediate material occurs iteratively by computing a loss between a current virtual intermediate material and a set of the physical properties and the chemical properties until the loss is 0 or the current virtual intermediate material satisfies the physical properties and the chemical properties of the material.

6. The computer-implemented method according to claim 4 or claim 5, further comprising optimizing, iteratively, the virtual intermediate material based on a gradient of a derivation associated with the virtual intermediate material and a loss objective, wherein the optimization of the virtual intermediate material occurs iteratively based on a comparison of a predicted material property and the target material property, or for a predefined number of iterations.

7. The computer-implemented method according to any of the preceding claims, wherein the computing format further includes two or more 3D graph representations of side chains for the main molecule and two or more degrees of substitution of the side chains.

8. The computer-implemented method according to any of the preceding claims, wherein the neural network includes one or more molecule encoders and one or more numerical encoders configured to, in parallel, encode the 3D graph representation of the main molecule, the 3D graph representation of the side chain, the molecular weight, and the degree of substitution into the latent representations.

9. The computer-implemented method according to claim 8, wherein the 3D graph representation of the main molecule and the 3D graph representation of the side chain comprise atom types, 3D atom positions, and an adjacency matrix.

10. The computer-implemented method according to claim 9, wherein the one or more molecule encoders are configured to convert the atom types to the latent representations using an embedding layer.

11. The computer-implemented method according to claim 9 or claim 10, wherein the one or more numerical encoders are configured to convert the 3D atom positions to the latent representations using a dense layer.

12. The computer-implemented method according to any of claims 9 to 11, wherein the one or more molecule encoders are further configured to combine, by concatenation, the latent representations of the atom types and the latent representations of the 3D atom positons to generate a concatenated latent representation.

13. The computer-implemented method according to any of the preceding claims, wherein the neural network is a differentiable deep neural network that includes the one or more graph attention layers, which are configured to combine the concatenated latent representation, a set of latent prototypes, and the adjacency matrix.

14. A computer system comprising one or more hardware processors which, alone or in combination, are configured to provide for execution of a method for material property optimization comprising the following steps: obtaining a dataset for a material, the dataset including a molecule shape description for the material, a molecular weight for the material, and a target material property for the material; encoding the material to a computing format including a three dimensional (3D) graph representation of a main molecule for the material, a 3D graph representation of a side chain for the main molecule, the molecular weight, and a degree of substitution of the side chain; and encoding at least the 3D graph representation of the main molecule and the 3D graph representation of the side chain, by a neural network that includes one or more graph attention layers, each into a latent representation usable for the material property optimization.

15. A tangible, non-transitory computer-readable medium having instructions thereon, which, upon being executed by one or more processors provide for execution of a method for material property optimization comprising the following steps: obtaining a dataset for a material, the dataset including a molecule shape description for the material, a molecular weight for the material, and a target material property for the material; encoding the material to a computing format including a three dimensional (3D) graph representation of a main molecule for the material, a 3D graph representation of a side chain for the main molecule, the molecular weight, and a degree of substitution of the side chain; and encoding at least the 3D graph representation of the main molecule and the 3D graph representation of the side chain, by a neural network that includes one or more graph attention layers, each into a latent representation usable for the material property optimization.

Images