Symposium X—MRS/The Kavli Foundation Frontiers of Materials

Symposium X_buonassisi_800 wideTonio Buonassisi, Massachusetts Institute of Technology

Finding a Path from “Promise” to “Performance”—Toward Realizing Novel Materials Predicted via Generative AI

Written by Sophia Chen

It’s all the rage: artificial intelligence—from generative models to automated laboratories—will help researchers find novel and useful materials faster. But Tonio Buonassisi, a mechanical engineer at Massachusetts Institute of Technology, knows that to get to that place, the field has its work cut out for it.

“This is not your normal tech-vangelist talk where I'm going to go up and show you how great AI is and how autonomous labs are going to rock the world,” said Buonassisi during Thursday’s Symposium X. “Yes, I do believe that. And I also believe that the path to getting there is a hard one.”

In Thursday’s Symposium X, titled “Finding a Path from ‘Promise’ to ‘Performance’—Toward Realizing Novel Materials Predicted via Generative AI,” Buonassisi outlined his perspectives on the state of automation and automated materials discovery, as well as the role of AI in these processes.

Buonassisi focused on pragmatic challenges. The adoption of AI will involve growing pains, he says. He described the trajectory as curve, where the field will hit what feels like unproductive bottlenecks: “You're putting in a lot of investment and not getting out much return. So this talk is about a lot of that part of the curve, and creating community around getting us over that part.”

Currently, the materials community has already adopted automation in a variety of settings. Scientists at the University of Liverpool, for example, use co-bots, or robots designed to work in close proximity to humans. Researchers at Boston University have also developed a robot arm that coordinates with 3D printers to manipulate materials samples.

Working at Singapore’s Agency for Science, Technology, and Research, Buonassisi started a program known as Accelerated Materials Development for Manufacturing, where they adopted a method known as “islands of automation.” To do this, instead of aiming to create a fully autonomous materials discovery process, they focused on accelerating “specific parts of the process and […] strategically targeted those points in the workflow that require more time, or that suffer from irreproducibility due to human factors,” he said.

The adoption of AI, which largely comprises models known as neural network, presents many challenges for the field, Buonassisi described in the talk. For example, the field needs to better identify which problems are most appropriate for autonomous or automated systems. On top of that, they need to increase the odds that experiments can actually synthesize and test an AI-predicted material. He listed several practical questions in the use of AI: for example, how should researchers choose a neural network architecture? How do they select, obtain, and validate training data? How should researchers represent materials to the neural network? How do they diagnose issues when things are going wrong? How should they select candidate materials that the neural network generates? How should they quantify the uniqueness or the novelty of the AI-generated materials? “These are […] examples of some of the things that, in my opinion, would yield truly high-impact papers, if you're a graduate student, postdoc, or early-stage career professor,” said Buonassisi. Using AI to discover unusual materials presents a particular challenge, he said. Experts have concluded that they would probably not have been able to predict the cuprates, high-temperature superconductors, given today’s AI tools and data prior to 1987 when researchers discovered the material.

The field should also critically examine the social structures that drive materials research, Buonassisi said. Individuals face different incentives than those of academia as an institution, and that can present a tension. While Buonassisi says this is a systemic problem and does not promise any solutions, he proposes a different academic structure in which the total number of students and postdocs are reduced. As they become more senior, they begin to manage “undergraduate researchers and staff who can augment their capabilities, who can complement some of their weaknesses in computation automation, in domain expertise,” he said.

Symposium X—MRS/The Kavli Foundation Frontiers of Materials features lectures aimed at a broad audience to provide meeting attendees with an overview of leading-edge topics.


Symposium SB11: Bio-based and Biomimetic Polymers in Soft Robotics

John A Rogers, Northwestern University

Flexible Electronic Skin for Sensing and Haptic Actuation

Written by Kazi Zihan Hossain

John Rogers and his group from Northwestern University have been pioneering flexible electronics that can be attached to human skin for various medical purposes. Traditional rigid electronics can be made flexible through geometrical manipulation such as thinning, incorporating wavy structures, or materials innovation. Instead of limiting the technology’s development to sensing body activity, these epidermal electronics also possess therapeutic potential. Skin-like wireless electronics have been used to monitor the vital health signs of babies in developing countries where adequate resources were not available. Soft electronics attached to the throat can be used to monitor the swallowing and vocal activity of stroke patients instead of bulky and cumbersome wired instruments, which is crucial for personalized rehabilitation. Haptic feedback integrated into wearables can assist prosthetic control and sensory substitution for people with nerve problems. The latest efforts have been made to attach flexible electronics to different positions of the human body and expand the haptic feedback to incorporate thermal feedback and multimodal systems to contribute to better healthcare systems.


Symposium MT03: Machine Learning Methods, Data and Automation for Sustainable Electronics

Victor Fung, Georgia Institute of Technology

Physics-Informed Pre-training of Graph Neural Networks for Materials Property Predictions

Written by Kazi Zihan Hossain

Materials scientists have used machine learning models and artificial intelligence (AI) to predict the properties of different materials. Traditional approaches require individual models and re-training for every different system to predict different properties. A model trained to predict a specific property can not be used to predict another property of the materials. To overcome this issue and develop a single model that can be used to predict different properties of materials, Victor Fung from the Georgia Institute of Technology has presented a Graph Neural Network (GNN) approach. In this technique, the composition and structure of a material can be encoded into a graph so that the properties of the materials can be related to the structure. Traditionally, GNN approaches work well when enough data is available to train the model, which may not be feasible from a material science perspective. Therefore, Fung and colleagues used the transfer learning technique to pre-train the model from a large dataset and then used the model to tune the properties of the materials. Different models, such as Crystal Graph Convolutional Neural Networks (CGCNN) and TorchMD-Net, were used to predict the properties of the materials, where the latter performed well under denoising scenarios.