Public Abstract

DE-SC0021253: FAIR Data and Interpretable AI Framework for Architectured Metamaterials

Award Status: Expired

Institution: California Institute of Technology, Pasadena, CA
UEI: U2JMKHNS5TG4
DUNS: 009584210

Most Recent Award Date: 07/29/2022
Number of Support Periods: 3
PM: Lentz, Margaret

Current Budget Period: 09/23/2022 - 09/22/2023
Current Project Period: 09/23/2020 - 09/22/2023
PI: Daraio, Chiara

Supplement Budget Period: N/A

Public Abstract

FAIR Data and Interpretable AI Framework for Architectured Metamaterials L. Catherine Brinson, Duke University (PI) Cynthia Rudin, Duke University (Co-PI) Chiara Daraio, California Institute of Technology (Co-PI) Since the inception of the “materials genome initiative” (MGI) in 2011, rapidly developing data-driven methods are increasingly employed to complement the traditional materials science tools of experimentation, theory, and computation. While much of this early progress has been executed by individual research groups, large-scale adoption of Artificial Intelligence (AI) for materials research has been limited by a lack of sufficient datasets and frameworks. To leverage the full power of AI models to advance fundamental scientific understanding for the discovery and design of new materials, materials’ data resources should be openly shared and standardized and new AI methodologies developed targeting this data. This project aims to fill this need using as an exemplar the domain of mechanical metamaterials: architected, composite materials that derive superior mechanical performance, such as combined light weight and high strength, controllable fracture, extreme energy/vibration absorption, or even negative effective mass or stiffness, due to their tessellated geometrical structures. These unique properties make metamaterials enticing for applications ranging from batteries, energy harvesting, seismic protection, vibration/sound insulation, frequency filtering, wave guiding, heat management, RF telecommunication, and ultrasonic imaging. To overcome both trial-and-error fabrication and ad-hoc testing of new metamaterial geometries, as well as the limitations of topology optimization, we will pursue two objectives: 1) We will use physics-based simulations of metamaterial properties based on structure to create large Findable, Accessible, Interoperable, and Reusable (FAIR) scientific datasets covering a vast space of geometric configurations. These data have properties that are completely different from typical AI/ML datasets involving computer vision or natural language processing. We will use well-known repositories to engage a broad range of researchers around these data, thereby driving new innovations in AI/ML. 2) We will develop a new AI framework and interpretable machine learning (ML) algorithms to enable robust discoveries with scientific data. While the vast design space of all available metamaterial configurations cannot be exhaustively explored through experiments or even simulations, our new AI framework – Learning Refined Compositional Rules (LRCR) – provides two significant advances: 1) LRCR can rapidly find subtle yet important patterns that relate to desired properties and 2) LRCR leverages results at coarse scales to build new rules for patterns at finer and finer scales. A key aspect of this framework is the implementation of interpretable machine learning algorithms to identify classes of building blocks for the creation of novel material configurations with desired properties, while also informing practicality constraints on the ability to reliably 3D-print the new materials. Furthermore, we will make a simulation tool publicly available for other AI/ML researchers and data scientists to add to our repository, explore their own hypotheses, and verify predictions, thereby increasing the robustness of our FAIR dataset and predictions from our AI framework over time, with the help of a broad scientific community.

FAIR Data and Interpretable AI Framework for Architectured Metamaterials

L. Catherine Brinson, Duke University (PI)

Cynthia Rudin, Duke University (Co-PI)

Chiara Daraio, California Institute of Technology (Co-PI)

Since the inception of the “materials genome initiative” (MGI) in 2011, rapidly developing data-driven methods are increasingly employed to complement the traditional materials science tools of experimentation, theory, and computation. While much of this early progress has been executed by individual research groups, large-scale adoption of Artificial Intelligence (AI) for materials research has been limited by a lack of sufficient datasets and frameworks. To leverage the full power of AI models to advance fundamental scientific understanding for the discovery and design of new materials, materials’ data resources should be openly shared and standardized and new AI methodologies developed targeting this data. This project aims to fill this need using as an exemplar the domain of mechanical metamaterials: architected, composite materials that derive superior mechanical performance, such as combined light weight and high strength, controllable fracture, extreme energy/vibration absorption, or even negative effective mass or stiffness, due to their tessellated geometrical structures. These unique properties make metamaterials enticing for applications ranging from batteries, energy harvesting, seismic protection, vibration/sound insulation, frequency filtering, wave guiding, heat management, RF telecommunication, and ultrasonic imaging.

To overcome both trial-and-error fabrication and ad-hoc testing of new metamaterial geometries, as well as the limitations of topology optimization, we will pursue two objectives: 1) We will use physics-based simulations of metamaterial properties based on structure to create large Findable, Accessible, Interoperable, and Reusable (FAIR) scientific datasets covering a vast space of geometric configurations. These data have properties that are completely different from typical AI/ML datasets involving computer vision or natural language processing. We will use well-known repositories to engage a broad range of researchers around these data, thereby driving new innovations in AI/ML. 2) We will develop a new AI framework and interpretable machine learning (ML) algorithms to enable robust discoveries with scientific data. While the vast design space of all available metamaterial configurations cannot be exhaustively explored through experiments or even simulations, our new AI framework – Learning Refined Compositional Rules (LRCR) – provides two significant advances: 1) LRCR can rapidly find subtle yet important patterns that relate to desired properties and 2) LRCR leverages results at coarse scales to build new rules for patterns at finer and finer scales. A key aspect of this framework is the implementation of interpretable machine learning algorithms to identify classes of building blocks for the creation of novel material configurations with desired properties, while also informing practicality constraints on the ability to reliably 3D-print the new materials. Furthermore, we will make a simulation tool publicly available for other AI/ML researchers and data scientists to add to our repository, explore their own hypotheses, and verify predictions, thereby increasing the robustness of our FAIR dataset and predictions from our AI framework over time, with the help of a broad scientific community.