IEEE Electron Device Letters · 2023 · 20 citations · 13 references
ComputingComputer EngineeringComputer ScienceComputational PhysicsMachine LearningSimulation-based DesignDeep LearningDevice ModelingDevice SimulationGraph-based LearningMesh GenerationNeural NetworksDeep Learning AccelerationDeep Learning AcceleratorsGraph Neural NetworksMesh Refinement TechniquesNumerical Device SimulationPhysics-based Device ModelingRedundant ComputationUnstructured Mesh Generation
There is an increasing number of studies to accelerate the TCAD simulation with deep learning models. Such studies rely on performing a procedure that interpolates an unstructured mesh into a structured mesh. This procedure, however, incurs intrinsic errors and redundant computation. To avoid this unnecessary procedure, this letter proposes a new method that can treat unstructured mesh itself to mimic TCAD device simulation. The method is to convert the unstructured mesh into a graph and then, directly applies a novel graph neural network (MHAT-GNN). In 45nm process, the proposed method outperforms pre-existing methods in terms of accuracy and efficiency.
13