Machine learning techniques for estimating mechanical properties of materials

Patent metadata

Field	Value
Publication number	US-11461519-B2
Application number	US-202017620219-A
Country	US
Kind code	B2
Filing date	Mar 6, 2020
Priority date	Jun 24, 2019
Publication date	Oct 4, 2022
Grant date	Oct 4, 2022

Abstract

Methods and apparatus for extracting one or more mechanical properties for a material based on one or more indentation parameters for the material. The method comprises receiving load-displacement data from one or more instrumented indentation tests on the material, determining, by at least one computer processor, the indentation parameters for the material based, at least in part, on the received load-displacement data, providing as input to a trained neural network, the indentation parameters for the material, determining, based on an output of the trained neural network, the one or more mechanical properties of the material, and displaying an indication of the determined one or more mechanical properties of the material to a user of the computer system.

First claim

The invention claimed is:

1. A computer system configured to train a neural network to extract one or more mechanical properties of a material from indentation parameters for the material, the computer system comprising:
at least one computer processor; and
at least one non-transitory computer readable medium encoded with instructions that, when executed by the at least one computer processor, perform a method, comprising:
providing as input to the neural network, the indentation parameters for the material;
training the neural network to map the input indentation parameters to the one or more mechanical properties of the material; and

storing the trained neural network on the at least one non-transitory computer readable medium,
wherein the neural network includes a first portion configured to receive as input low-fidelity indentation parameters, a second portion configured to receive as input high-fidelity indentation parameters, and an integration portion that integrates one or more outputs of the first portion and one or more outputs of the second portion.

2. The computer system of claim 1 , wherein training the neural network comprises training the first portion of the neural network based on the low-fidelity indentation parameters and training the second portion of the neural network based on the high-fidelity indentation parameters.

3. The computer system of claim 2 , wherein training the neural network further comprises integrating both the high-fidelity indentation parameters and the low-fidelity implementation parameters using convolution and/or recursion.

4. The computer system of claim 2 , wherein a number of datum in the high-fidelity indentation parameters used to train the first portion of the neural network is less than a number of datum in the low-fidelity indentation parameters used to train the second portion of the neural network.

5. The computer system of claim 1 , wherein the high-fidelity indentation parameters include first indentation parameters determined from one or more simulations and second indentation parameters determined from one or more indentation tests on the material.

6. The computer system of claim 1 , wherein training the neural network comprises training the neural network to reduce at least one systematic error by using data determined from one or more indentation tests on the material to determine at least some of the high-fidelity indentation parameters.

7. The computer system of claim 1 , wherein the low-fidelity indentation parameters include indentation parameters determined using one or more simulations.

8. The computer system of claim 1 , wherein
the first portion and/or the second portion of the neural network are pre-trained using a baseline training process; and
training the neural network to map the input indentation parameters to the one or more mechanical properties of the material comprises training only the second portion of the neural network with high-fidelity indentation parameters.

9. The computer system of claim 1 , wherein the neural network includes convolutional and recursive linear and/or nonlinear integration of training data with at least three levels of fidelities.

10. The computer system of claim 1 , wherein the indentation parameters include indentation parameters for multiple indenter geometries.

11. The computer system of claim 10 , wherein the multiple indenter geometries have different half-included tip angles.

12. The computer system of claim 10 , wherein the multiple indenter geometries include multiple indenter shapes.

13. The computer system of claim 1 , wherein the one or more mechanical properties include a reduced Young's modulus, a yield strength, and/or a strain hardening parameter.

14. The computer system of claim 13 , wherein the one or more mechanical properties include at least two of a reduced Young's modulus, a yield strength, and a strain hardening parameters.

15. The computer system of claim 1 , wherein the one or more mechanical properties include a plurality of points on a stress-strain curve.

16. The computer system of claim 15 , wherein the one or more mechanical properties include a plurality of strain values at different plastic strains.

17. The computer system of claim 1 , wherein the one or more indentation parameters include one or more indentation parameters extracted from a loading portion of an indentation curve, an unloading portion of the indentation curve, and/or both the loading and the unloading portion of the indentation curve.

18. The computer system of claim 1 , wherein the one or more indentation parameters include one or more of loading curvature, initial unloading slope, and plastic work ratio.

19. The computer system of claim 18 , wherein the one or more indentation parameters include the loading curvature, the initial unloading slope and the plastic work ratio.

20. The computer system of claim 1 , wherein the method further comprises:
receiving load-displacement data for the material; and
determining the one or more indentation parameters from the received load-displacement data.

21. The computer system of claim 1 , wherein the material comprises a 3D printed material.

22. The computer system of claim 1 , wherein training the neural network comprises training the neural network using training data having more than two levels of fidelities.

23. A method of training a neural network to extract one or more mechanical properties of a material from indentation parameters for the material, the method comprising:
providing as input to the neural network, the indentation parameters for the material;
training, using at least one computer processor, the neural network to map the input indentation parameters to the one or more mechanical properties of the material; and

storing the trained neural network on at least one non-transitory computer readable medium,
wherein the neural network includes a first portion configured to receive as input low-fidelity indentation parameters, a second portion configured to receive as input high-fidelity indentation parameters, and an integration portion that integrates one or more outputs of the first portion and one or more outputs of the second portion.

24. The method of claim 23 , wherein training the neural network comprises training the first portion of the neural network based on the low-fidelity indentation parameters and training the second portion of the neural network based on the high-fidelity indentation parameters.

25. The method of claim 24 , wherein training the neural network further comprises integrating both the high-fidelity indentation parameters and the low-fidelity implementation parameters using convolution and/or recursion.

26. The method of claim 24 , wherein a number of datum in the high-fidelity indentation parameters used to train the first portion of the neural network is less than a number of datum in the low-fidelity indentation parameters used to train the second portion of the neural network.

27. The method of claim 23 , wherein the high-fidelity indentation parameters include first indentation parameters determined from one or more simulations and second indentation parameters determined from one or more indentation tests on the material.

28. The method of claim 23 , wherein training the neural network comprises training the neural network to reduce at least one systema

Assignees

• Univ Nanyang Tech
• Massachusetts Inst Technology
• Univ Brown

Inventors

• Lu Lu
• Dao Ming
• Suresh Subra
• Karniadakis George

Classifications

• G06N3/048

  Activation functions · CPC title

• G06N3/045

  Combinations of networks · CPC title

• G06N3/044

  Recurrent networks, e.g. Hopfield networks · CPC title

• G06N3/08 Primary

  Learning methods · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title