GRAPH NEURAL NETWORK AND MACHINE LEARNING FOR THERMAL MANAGEMENT OF NEUROMORPHIC MATERIALS
Abstract
The increasing and continuous manipulation of digital data pervade all aspects of modern life. Computing devices have been become essential to processing and storing information. The data interchange between the processing and memory units (von Neumann architecture) sustains increasing energy demands. Energy consumption associated with data analysis constitutes a significant part of world energy demands and continues to increase. Hence it is essential to search for more energy-efficient technology. This major challenge calls for the new energy-efficient computing paradigm. Inspired by biology, the next generation of computing devices attempts to imitate the human brain s low energy consumption and dynamic learning capabilities. The scientific contribution of this project is the development of new methods for predict materials performance features, mainly the thermal management of neuromorphic materials, optimizing experimental resources to the energy-efficient device design and fabrication; the ability to simulate transient conditions in real-time identifying emergent collective behaviors; provide a new method for classification, testing and prediction combing network analysis and machine learning techniques. This project constitutes multidisciplinary research and highly collaborative work with the Dr. Ivan Schuller group at the University of California San Diego (UCSD) to explore the thermal management of neuromorphic materials and devices. We perform extensive computational simulations of heat transport in phase materials, including changes in the material s electric, thermic, structural properties. This novel approach gathers all elements supporting phase coexistence based on graphical representation encompassing critical elements of long-range correlations. The inclusion of different material properties allows to set up realistic operational conditions to explore and predict the behavior of novel geometric structures, which can be fabricated and tested directly. Predicting materials properties play a crucial role in research and technology. Methods based on machine learning offer essential tools for detecting, classifying, and predicting materials properties.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Apr 20, 2023
- Source ID
- FA95502210318
Entities
People
- Felipe Torres
Organizations
- Air Force Office of Scientific Research
- United States Air Force
- University of Chile