The rapid expansion of artificial intelligence (AI) data centers around the world is intensifying a vexing problem: what to do with the vast amounts of heat generated by the powerful chips that drive AI. And the challenge is not limited to the technology industry. Across transportation, manufacturing, consumer electronics, and other sectors, vast amounts of thermal energy are generated and wasted every day, making heat management an increasingly important challenge for modern energy systems.
“Heat is often treated as a byproduct, but it’s really an energy resource,” said Sobin Alosious, a postdoctoral researcher in Lucy Family Institute for Data & Society‘s Artificial Intelligence (AI) & Materials Initiative at the University of Notre Dame. “By understanding how materials store, transfer, and release heat, we can design better materials that improve energy efficiency across a wide range of technologies.”
Now, researchers in the Initiative are using AI to identify materials capable of doing exactly that.
Through work led by Tengfei Luo, Dorini Family Professor of Aerospace and Mechanical Engineering, Meng Jiang, Associate Professor of Computer Science and Engineering, and Alosious, the Initiative is creating machine learning (ML) frameworks to accelerate the discovery of advanced materials. Machine learning enables computers to learn patterns from existing data and apply those patterns to make predictions about new materials. The researchers combine experimental results with simulations and physics-based estimates to identify promising materials more efficiently.
A recent publication from the team demonstrates a transfer learning approach, where ML models learn from other relevant datasets to predict polymer heat capacity. Polymers, a class of materials composed of long chains of repeating molecular building blocks and commonly used in technologies such as lithium-ion batteries and computer chip packaging, are critical for thermal applications. Heat capacity plays an important role in thermal energy storage, waste heat recovery, electronics cooling, and battery thermal management, but accurately measuring it can be costly, and computational predictions often require tradeoffs between speed and accuracy.
To overcome these limitations, the researchers developed an ML framework that combines experimental measurements, which are scarce, with large datasets generated through molecular dynamics simulations. Molecular dynamics simulations are much cheaper to generate than experimental measurements because they use computer models instead of laboratory testing, but the resulting data are less precise because the simulations rely on simplified representations of real-world molecular behavior. By extracting useful information from lower-cost data sources before refining it with smaller collections of high-quality experimental data, the model improves prediction accuracy while reducing reliance on expensive experiments.
The results demonstrate the power of integrating multiple datasets of varying accuracy and cost to build highly accurate surrogate models in materials science. By leveraging information across multiple sources, the framework helps overcome one of the most persistent limitations in polymer informatics: the scarcity of experimentally measured property data.
The study also reflects a broader strategy within the AI & Materials Initiative to accelerate materials discovery through data-driven tools. To generate the large datasets needed for ML, the team developed ADEPT, an automated molecular simulation engine that rapidly converts polymer structures into atomistic models and predicts material properties.
Building on these capabilities, the researchers created PolyGraphMT, a graph neural network framework that integrates experimental and computational data of varying quality to identify structure–property relationships across a wide range of polymer characteristics, including thermal, mechanical, transport, electronic, and optical properties. The resulting framework has been used to screen approximately 13,000 known polymers as well as nearly one million virtual polymers, generating tens of millions of property predictions.
To make these capabilities accessible to researchers, the team developed an integrated polymer discovery platform that combines property prediction, AI-assisted screening, literature support, and automated simulation preparation in a single workflow. Researchers can define application-specific design goals and constraints, and the platform rapidly identifies high-confidence candidate polymers from a library of more than one million materials for further simulation and experimental evaluation.
Together, ADEPT, PolyGraphMT, and the polymer discovery platform form an integrated discovery pipeline that links automated simulation, machine learning, candidate screening, and higher-fidelity validation to continuously improve the design of advanced materials.
“The AI & Materials Initiative is creating a new paradigm for materials discovery,” said Luo, who also leads the University’s MÖNSTER Lab (MOlecular/Nano-Scale Transport & Energy Research Laboratory). “We are developing platforms that continuously generate data, learn from that data, and use those insights to guide the new discovery.”
By integrating physics-based modeling, data generation, and AI into a single workflow, the Initiative aims to transform materials discovery from a slow, sequential process into a scalable engine for innovation.
“The real opportunity is enabling scientists to explore materials spaces that would be impossible to investigate experimentally alone,” said Nitesh Chawla, the Frank M. Freimann Professor of Computer Science and Engineering and Lucy Family Director for Data & AI Academic Strategy, who leads Notre Dame’s Data, AI, and Computing Initiative. “Through efforts that bring together data, artificial intelligence, and computing, we are building a foundation for discovery at scale, where AI helps identify the most promising directions and reduces the time from idea to application.”
To learn more about the AI & Materials Initiative, please visit the Lucy Family Institute for Data & Society website.
Contact:
Christine Grashorn, Program Director, Engagement and Strategic Storytelling
Lucy Family Institute for Data & Society / University of Notre Dame
cgrashor@nd.edu / 574.631.4856
lucyinstitute.nd.edu / @lucy_institute
About the Lucy Family Institute for Data & Society
Guided by Notre Dame’s Mission, the Lucy Family Institute adventurously collaborates on advancing data-driven and artificial intelligence (AI) convergence research, translational solutions, and education to ethically address society’s vexing problems. As an innovative nexus of academia, industry, and the public, the Institute also fosters data science and AI access to strengthen diverse and inclusive capacity building within communities.