High-throughput computational screening of materials with target thermal conductivity has the capability to transform many industries such as thermoelectricity generation and high performance micro-/nano electronic devices, which produce a significant amount of excess heat during operation and searching for materials with high thermal conductivity is extremely important for the disruptive development of such micro-/nano-electronics in order to prolong their working life and increase reliability However, this potential has not been implemented due to the huge computational resources needed by current first-principles based thermal conductivity calculations and challenges of theoretical models because of the highly complex and nonlinear relationships from atomic structures of materials to the thermal transport properties. Deep learning has transformed an increasing number of fields where big data are available such as image and speech recognition, and medical image analysis. However, the materials science has remained largely untapped by deep learning despite its high economical potential. This two-year EAGER project aims to develop novel deep neural network techniques to achieve fast and accurate computational prediction of thermal conductivity for high-throughput thermal material discovery. The development of a reliable, fast, and accurate deep learning models is a necessity towards experimental validation and realistic application of high-throughput thermal materials screening. Simultaneously the program will aim to enhance diversity by engaging minority and underrepresented students to participate in STEM research. The participants will also develop understanding of both atomistic simulations of thermal transport and big data analytics; hence contributing to workforce development.
Deep learning algorithm has been well developed in computer science, while direct thermal conductivity prediction from atomic structures has been made available in materials science. However, intuitive combination of this progress is not an easy task, since the scientific data (thermal conductivity of materials) cannot be quickly expanded to the level required by deep learning. To this end, this project will use heterogeneous multi-resolution thermal conductivity data and scarce data for training efficient and accurate deep learning models, which has never been realized for AlphaGO-like deep learning models before. In this exploratory stage, the focus will be on (1) developing graph and spatial 3D convolutional neural networks (CNNs) for thermal conductivity modeling by exploiting their automated hierarchical feature learning and non-linear mapping learning, and (2) developing multi-resolution data based deep neural network models for thermal conductivity prediction. (3) Experimental and DFT-based computational validation of predicted materials with extremely high or low thermal conductivity. A robust, reliable, and high accuracy deep learning model for thermal conductivity prediction will facilitate development of advanced functional materials in industry, such as energy conversion, storage, and thermal management.
This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.
