IP Library › Granted Patent US 12,190,236
Granted Patent B2
US 12,190,236 · App. 17/240,554 · Granted Jan 7, 2025

Predicting properties of materials from physical material structures

Inventors: Annette Ada Nkechinyere Obika (London, GB); Tian Xie (Cambridge, MA); Victor Constant Bapst (London, GB); Alexander Lloyd Gaunt (Ely, GB); James Kirkpatrick (London, GB)
Assignee: DeepMind Technologies Limited
G06N3/08G06N3/04
View Patent ↗
Loading inventors, assignments & file history…
Monitor This Case
Get email alerts when status or documents change.
Order Certified Copies
Most orders are placed with the USPTO same day — all within 24 business hours.
Order via The Patent Place →
Pre-filled with this patent's details
Quick Facts
Patent No.
US 12,190,236
App. No.
17/240,554
Granted
Jan 7, 2025
Kind
B2
Abstract

Methods, computer systems, and apparatus, including computer programs encoded on computer storage media, for predicting one or more properties of a material. One of the methods includes maintaining data specifying a set of known materials each having a respective known physical structure; receiving data specifying a new material; identifying a plurality of known materials in the set of known materials that are similar to the new material; determining a predicted embedding of the new material from at least respective embeddings corresponding to each of the similar known materials; and processing the predicted embedding of the new material using an experimental prediction neural network to predict one or more properties of the new material.

Claims (59)

1. A method performed by one or more computers, the method comprising:

maintaining data specifying a set of known materials each having a respective known physical structure;

receiving data specifying a new material;

identifying a plurality of similar known materials in the set of known materials that are similar to the new material;

determining a predicted embedding of the new material from at least respective embeddings corresponding to each of the plurality of similar known materials, each respective embedding representing the respective known physical structure of a corresponding similar known material in the plurality of similar known materials; and

processing the predicted embedding of the new material using an experimental prediction neural network to predict one or more properties of the new material;

wherein the experimental prediction neural network has been trained as a component of a material property prediction neural network comprising (i) a graph neural network configured to receive a graph representation of a physical structure of a material and to generate, from the graph representation, an embedding of the material, (ii) the experimental prediction neural network configured to receive an experimental embedding derived from the embedding of the material and predict one or more properties of new material that would be measured in a physical experiment on the material, and (iii) a simulation prediction neural network configured to receive the embedding of the material and predict one or more properties of the new material that would be measured in a simulation on the material, wherein the material property prediction neural network has been trained by:

performing first training of the graph neural network and the simulation prediction neural network on simulation training data that comprises, for each of a plurality of first materials (i) data identifying a physical structure of the first material and (ii) simulated values for each of the one or more properties that were measured in a simulation of the first material; and

performing second training of at least the graph neural network and the experimental prediction neural network on experimental training data that comprises, for each of a plurality of second materials, experimental values for each of the one or more properties that were measured in a physical experiment on the second material.

2. The method of claim 1 , wherein, for each of the one or more properties, the experimental prediction neural network is configured to regress a predicted value of the one or more properties.

3. The method of claim 2 , wherein the one or more properties include an energy of formation of the new material.

4. The method of claim 1 , further comprising:

determining that the new material does not have a stable correspondence to any of the known materials in the set of known materials, wherein:

identifying the plurality of similar known materials, determining the predicted embedding, and processing the predicted embedding are performed in response to determining that the new material does not have a stable correspondence to any of the known materials in the set of known materials.

5. The method of claim 4 , further comprising:

receiving data specifying a second new material;

determining that the second new material has a stable correspondence to a second known material in the set of known materials;

determining a predicted embedding of the second new material from an embedding of the second known material that represents the known physical structure of the second known material; and

processing an experimental embedding derived from the predicted embedding of the second new material using the experimental prediction neural network to predict one or more properties of the second new material.

6. The method of claim 1 , wherein identifying a plurality of known materials in the set of known materials that are similar to the new material comprises:

identifying the plurality of similar known materials by generating a representation of the new material as a weighted combination of the plurality of similar known materials that assigns a respective weight to each of the plurality of similar known materials.

7. The method of claim 6 , wherein identifying a plurality of known materials in the set of known materials that are similar to the new material comprises:

identifying the plurality of similar known materials and generating the representation through phase diagram decomposition.

8. The method of claim 6 , wherein determining a predicted embedding of the new material from at least respective embeddings corresponding to each of the similar known materials comprises:

for each of the plurality of similar known materials, generating an experimental embedding of the similar known material from the embedding of the similar known material; and

generating the predicted embedding of the new material by performing an interpolation between the experimental embeddings using the weights assigned to the similar known materials in the weighted combination.

9. The method of claim 8 , wherein generating the experimental embedding of the similar known material comprises:

processing an input comprising the embedding using a compositional neural network to generate the experimental embedding.

10. The method of claim 9 , wherein the input further comprises, for each of one or more of the one or more properties, a simulated value of the property for the similar known material.

11. The method of claim 1 , wherein the data specifying the new material is a chemical formula of the new material.

12. The method of claim 1 , further comprising:

generating the respective embeddings of each of the plurality of similar known materials, comprising, for each of the similar known materials:

generating a graph representation of the known physical structure of the similar known material; and

processing the graph representation of the known physical structure using a graph neural network to generate the embedding of the similar known material.

13. The method of claim 1 , wherein the known physical structure is an atomistic structure of the known material.

14. The method of claim 1 , wherein the experimental prediction neural network is a feedforward neural network that is configured to process the predicted embedding to regress predicted values of the one or more properties of the new material.

15. The method of claim 1 , wherein

the first training and second training have been performed jointly.

16. The method of claim 15 , wherein the first training and second training have been trained in a joint training that comprises adjusting respective weights assigned to the first training and the second training as the joint training progresses.

17. A system comprising:

one or more computers; and

one or more storage devices storing instructions that when executed by the one or more computers, cause the one or more computers to perform:

maintaining data specifying a set of known materials each having a respective known physical structure;

receiving data specifying a new material;

identifying a plurality of similar known materials in the set of known materials that are similar to the new material;

determining a predicted embedding of the new material from at least respective embeddings corresponding to each of the plurality of similar known materials, each respective embedding representing the respective known physical structure of a corresponding similar known material in the plurality of similar known materials; and

processing the predicted embedding of the new material using an experimental prediction neural network to predict one or more properties of the new material;

wherein the experimental prediction neural network has been trained as a component of a material property prediction neural network comprising (i) a graph neural network configured to receive a graph representation of a physical structure of a material and to generate, from the graph representation, an embedding of the material, (ii) the experimental prediction neural network configured to receive an experimental embedding derived from the embedding of the material and predict one or more properties of new material that would be measured in a physical experiment on the material, and (iii) a simulation prediction neural network configured to receive the embedding of the material and predict one or more properties of the new material that would be measured in a simulation on the material, wherein the material property prediction neural network has been trained by:

performing first training of the graph neural network and the simulation prediction neural network on simulation training data that comprises, for each of a plurality of first materials (i) data identifying a physical structure of the first material and (ii) simulated values for each of the one or more properties that were measured in a simulation of the first material; and

performing second training of at least the graph neural network and the experimental prediction neural network on experimental training data that comprises, for each of a plurality of second materials, experimental values for each of the one or more properties that were measured in a physical experiment on the second material.

18. One or more non-transitory computer-readable storage media storing instructions that, when executed by one or more computers, cause the one or more computers to perform:

maintaining data specifying a set of known materials each having a respective known physical structure;

receiving data specifying a new material;

identifying a plurality of similar known materials in the set of known materials that are similar to the new material;

determining a predicted embedding of the new material from at least respective embeddings corresponding to each of the plurality of similar known materials, each respective embedding representing the respective known physical structure of a corresponding similar known material in the plurality of similar known materials; and

processing the predicted embedding of the new material using an experimental prediction neural network to predict one or more properties of the new material;

wherein the experimental prediction neural network has been trained as a component of a material property prediction neural network comprising (i) a graph neural network configured to receive a graph representation of a physical structure of a material and to generate, from the graph representation, an embedding of the material, (ii) the experimental prediction neural network configured to receive an experimental embedding derived from the embedding of the material and predict one or more properties of new material that would be measured in a physical experiment on the material, and (iii) a simulation prediction neural network configured to receive the embedding of the material and predict one or more properties of the new material that would be measured in a simulation on the material, wherein the material property prediction neural network has been trained by:

performing first training of the graph neural network and the simulation prediction neural network on simulation training data that comprises, for each of a plurality of first materials (i) data identifying a physical structure of the first material and (ii) simulated values for each of the one or more properties that were measured in a simulation of the first material; and

performing second training of at least the graph neural network and the experimental prediction neural network on experimental training data that comprises, for each of a plurality of second materials, experimental values for each of the one or more properties that were measured in a physical experiment on the second material.

Assignments (2)
ASSIGNMENT OF ASSIGNOR'S INTEREST Recorded Apr 29, 2025
From: DEEPMIND TECHNOLOGIES LIMITED
To: GDM HOLDING LLC
Reel/Frame 071109/0414 →
ASSIGNMENT OF ASSIGNOR'S INTEREST Recorded Jun 7, 2021
From: OBIKA, ANNETTE ADA NKECHINYERE; XIE, TIAN; BAPST, VICTOR CONSTANT; GAUNT, ALEXANDER LLOYD; KIRKPATRICK, JAMES
To: DEEPMIND TECHNOLOGIES LIMITED
Reel/Frame 056456/0182 →
Continuity (2)
Provisional Application 63015328 · Apr 24, 2020
Related Publication 20210334655A1 · Oct 28, 2021
References Cited (40)
US 10830747B2 · Agrawal · 2020 [cited by examiner]
US 11587646B2 · Colby · 2023 [cited by examiner]
US 11901045B2 · Katsuki · 2024 [cited by examiner]
US 20220044769A1 · Ramprasad · 2022 [cited by examiner]
Bartel et al., “A critical examination of compound stability predictions from machine learned formation energies,” npj Computational Materials, 6(97): 11 pages. [cited by applicant]
Battaglia et al., “Relational inductive biases, deep learning, and graph networks,” CoRR, Jun. 2018, arxiv.org/abs/1806.01261, 40 pages. [cited by applicant]
Butler et al., “Machine learning for molecular and materials science,” Nature, Jul. 2018, 559:547-555. [cited by applicant]
Carleo et al., “Machine learning and the physical sciences,” Rev. Mod. Phys., Dec. 2019, 91:045002. [cited by applicant]
Chen et al., “Graph networks as a universal machine learning framework for molecules and crystals,” Chem. Mater., Apr. 2019, 31(9):3564-3572. [cited by applicant]
Cohen et al., “Challenges for density functional theory,” Chemical Reviews, Dec. 2011, 112(1):289-320. [cited by applicant]
Cohen et al., “Insights into current limitations of density functional theory,” Science, Aug. 2008, 321(5890):792-794. [cited by applicant]
Curtarolo et al., “AFLOW: An automatic framework for high-throughput materials discovery,” Computational Materials Science, Jun. 2012, 58:218-226. [cited by applicant]
Dunn et al., “Benchmarking materials property prediction methods: the matbench test set and automatminer reference algorithm,” npj Computational Materials, Sep. 2020, 6(138):1-10. [cited by applicant]
Duvenaud et al., “Convolutional networks on graphs for learning molecular fingerprints,” CoRR, Sep. 2015, arxiv.org/abs/1509.09292, 9 pages. [cited by applicant]
Faber et al., “Prediction errors of molecular machine learning models lower than hybrid DFT error,” J. Chem. Theory Comput., Sep. 2017, 13(11):5255-5264. [cited by applicant]
Ghiringhelli et al., “Big data of materials science: critical role of the descriptor,” Phys. Rev. Lett., Mar. 2015, 114:105503. [cited by applicant]
Hellenbrandt et al., “The inorganic crystal structure database (icsd)—present and future,” Crystallography Reviews, 2004, 10(1):17-22. [cited by applicant]
Isayev et al., “Universal fragment descriptors for predicting properties of inorganic crystals,” Nature Communications, Jun. 2017, 8(15679):1-12. [cited by applicant]
Jain et al., “Commentary: The materials project: A materials genome approach to accelerating materials innovation,” APL Materials, 1:011002. [cited by applicant]
Jain et al., “Formation enthalpies by mixing gga and gga+ u calculations,” Physical Review B, Jul. 2011, 84:045115. [cited by applicant]
Jha et al., “Deep learning the chemistry of materials from only elemental composition,” Scientific Reports, 8(17593):13 pages. [cited by applicant]
Jha et al., “Enhancing materials property prediction by leveraging computational and experimental data using deep transfer learning,” Nature Communications, Nov. 2019, 10(5316): 12 pages. [cited by applicant]
Kim et al., “Experimental formation enthalpies for intermetallic phases and other inorganic compounds,” Scientific Data, Oct. 2017, 4(170162):11 pages. [cited by applicant]
Kirklin et al., “The open quantum materials database (oqmd): assessing the accuracy of dft formation energies,” npj Computational Materials, Dec. 2015, 1(15010):15 pages. [cited by applicant]
Lany, “Semiconductor thermochemistry in density functional calculations,” Physical Review B, 78:245207. [cited by applicant]
Ong et al., “Python materials genomics (pymatgen): A robust, open-source python library for materials analysis,” Computational Materials Science, Feb. 2013, 68:314-319. [cited by applicant]
Ouyang et al., “Sisso: A compressed-sensing method for identifying the best low-dimensional descriptor in an immensity of offered candidates.” Physical Review Materials, Aug. 2018, 2:083802. [cited by applicant]
Saal et al., “Materials design and discovery with high-throughput density functional theory: The open quantum materials database (oqmd),” JOM, Sep. 2013, 65:1501-1509. [cited by applicant]
Scarselli et al., “The graph neural network model,” IEEE Transactions on Neural Networks, Jan. 2009, 20(1):61-80. [cited by applicant]
Schmidt et al., “Recent advances and applications of machine learning in solid-state materials science,” npj Computational Materials, Aug. 2019, 5(83):36 pages. [cited by applicant]
Schutt et al., “Schnet—a deep learning architecture for molecules and materials,” The Journal of Chemical Physics, Mar. 2018, 148:241722. [cited by applicant]
Senior et al., “Improved protein structure prediction using potentials from deep learning,” Nature, Jan. 2020, 577:706-710. [cited by applicant]
Silver et al., “A general reinforcement learning algorithm that masters chess, shogi, and Go through self-play,” Science, Dec. 2018, 362(6419):1140-1144. [cited by applicant]
Szegedy et al., “Inception-v4, inception-resnet and the impact of residual connections on learning,” Thirty-First AAAI Conference on Artificial Intelligence, Feb. 2017, 31(1):4278-4284. [cited by applicant]
Van Den Oord et al., “Pixel recurrent neural networks,” Proceedings of the 33rd International Conference on Machine Learning, 2016, 48:1747-1756. [cited by applicant]
Vaswani et al., “Attention is all you need,” CoRR, Jun. 2017, arxiv.org/abs/1706.03762, 15 pages. [cited by applicant]
Xie et al., “Crystal graph convolutional neural networks for an accurate and interpretable prediction of material properties,” Phys. Rev. Lett., Apr. 2018, 120(145301):15 pages. [cited by applicant]
Yang et al., “Analyzing learned molecular representations for property prediction,” J. Chem. Inf. Model, Jul. 2019, 59(8):3370-3388. [cited by applicant]
Zaheer et al., “Deep sets,” CoRR, Mar. 2017, arxiv.org/abs/1703.06114, 29 pages. [cited by applicant]
Zhang et al., “A strategy to apply machine learning to small datasets in materials science,” npj Computational Materials, May 2018, 4(25):1-8. [cited by applicant]