A simulation physics-guided neural network for predicting semiconductor structure with few experimental data

Prediction of semiconductor Critical Dimensions (CDs) from ellipsometry requires the machine learning model. However, proper training of a machine learning model is challenging because the measurement process of typical experimental CD data, which is mostly carried out with transmission electron microscopy (TEM), is a time-and cost-consuming process. To obtain a robust machine learning model with few experimental data, we propose a physics-guided neural network (PGNN) architecture. PGNN extracts spectrum-CD physics from simulation data and constructs physics-guided loss function for guiding the model optimization. The proposed algorithm has superior performance compared with other baseline algorithms and can be properly trained only with small experimental CD data, including label noise.