D. Hill, Y. Gong, S. Rakheja, A. Tunga, A. Janotti, R. Hu, M. Grupen
SRI International,
United States
Summary:
Developing the next generation of electronic devices typically relies on inaccurate or computationally expensive models, or requires slow development through many fabrication iterations to gather empirical data. Our approach brings together innovations in ML-enhanced materials simulation and new device simulation approaches to dramatically speed up computational time while maintaining high accuracy. This enables accurate prediction of novel device performance based on first principles properties with minimal computational resources. Current high accuracy atomistic simulation technique like hybrid DFT can generate accurate electronic structure information for small numbers of atoms but is so computationally expensive that it is intractable for realistic sized domains of 10s of thousands of atoms or more. Alternative faster approaches sacrifice accuracy. Similarly, transport/TCAD methods with high accuracy like NEGF and Monte Carlo, are also too computationally expensive to compute at scale. Tractable approaches rely on approximations and empirical fitting, which hampers their ability to accurately predict device behavior outside of the fitting set. We utilize a machine learning process for learning atomistic properties to enable accurate implementation of DFTB with >10,000x speed up over conventional high-fidelity simulation approaches. We then utilize that DFT data as the basis for transport simulations that describe device-level performance. We utilize a new transport methodology that treats regions of the electronic structure as independent Fermi gases that can exchange energy, momentum, and particles with each other and the lattice, enabling rapid non-quasistatic transport computation with near-quantum transport accuracy and >100x speed up over conventional approaches.