Numerical Analysis of the process for single crystal growths is high-cost computing due to the coupled heat conduction, flow, and chemical reaction calculations. Therefore, fast and highly accurate surrogate models are demanding. This study targets the high precision and stability of the surrogate models for the heat conduction calculations. We used the fully connected neural networks, and explored hyper-parameters. The activation function is a key parameter to enhance the accuracy of the surrogate models, and softplus function shows the higher precision than the standard relu function. The high accuracy is reproduced for some regression targets, and the softplus function can also generate the stable precision.
HAYASHIDA et al. (Wed,) studied this question.