Neural Network Training for Library-Based CD Metrology

The patent performs preliminary determination of the optimal number of diffraction orders before executing the full RCWA simulation. By pre-calculating which diffraction orders contribute significantly to the solution based on convergence criteria, the system avoids unnecessary computations of higher-order terms that would increase computation time without improving accuracy. This preliminary action resolves the contradiction by establishing the appropriate level of detail beforehand.

The patent implements a strategy where diffraction orders are included progressively until a convergence threshold is met, rather than including all possible orders. This partial action approach includes only the necessary number of diffraction orders required to achieve the desired accuracy, excluding excessive higher-order terms that would waste computational resources. The system performs partial computation with sufficient orders rather than exhaustive computation with all possible orders.

The patent performs preliminary determination of the optimal number of neurons and iteration count before executing full neural network training. By pre-assessing the complexity of the spectral library and the required accuracy level, the system calculates the sufficient number of neurons and iterations needed, avoiding both under-training and over-training. This preliminary configuration resolves the contradiction by establishing appropriate training parameters beforehand.

The patent implements feedback mechanisms during neural network training where the system monitors convergence metrics and adjusts the number of iterations and neurons dynamically. If the training achieves the desired accuracy earlier than expected, the system terminates training early, avoiding unnecessary iterations. This feedback control ensures that computational resources are used efficiently while maintaining the required precision for CD metrology.