Approximation of Time-Consuming Simulation Based on Generative Adversarial Network
Ryohei Orihara, Ryota Narasaki, Yuma Yoshinaga, Yasuhiro Morioka, Yoshiyuki Kokojima · 2018
A simulation model often has parameters to be calibrated through evaluation against data obtained by experiments. The process could be prohibitively expensive if the execution of the model requires huge computational resources and/or the parameter search space is vast. For technology CAD (TCAD) topography simulation, used in semiconductor manufacturing, this is exactly the case. We present a method to replace the simulation with approximate computation based on Generative Adversarial Network. The method receives pairs of a simulation parameter set and a simulation result image as training examples and learns a model to mimic the simulation. Unlike most GANbased models, its learning is carried out in a supervised manner. We evaluate its performance by cross-validation and comparison with a human expert. Despite the small size of the available dataset, our method successfully generates an image close to the ground truth in 71% of the cases and qualitatively reproduces the educated guesses of the human expert on the behavior of the simulator.