Multi-task learning surrogate model for rapid prediction of negative tone development metrics

Abstract

Negative tone development (NTD) lithography process is widely adopted in advanced technology nodes to meet the demand for higher integrated circuit density. However, traditional rigorous models to simulate the NTD resist shrinkage are computationally intensive. To address this challenge, we propose a high-speed, physics-aware multi-task learning (MTL) surrogate model for NTD simulation. Our approach uses 8 NTD parameters to predict the geometric metrics of 3 dimensional photoresist profiles, including the critical dimension, sidewall angle, and height losses for both flat and protruding regions. The proposed architecture uses a dynamic homoscedastic uncertainty weighting loss function to balance the gradients across those physically distinct targets. The model achieves exceptional prediction accuracy, demonstrating a coefficient of determination (R2) exceeding 0.98 for all spatial dimensions, while accelerating the simulations by over 105 times compared to the rigorous model. Furthermore, we employ shapley additive explanations (SHAP) to decode the surrogate model, quantitatively verifying its physical interpretability. Feature sensitivity analysis aligns fundamentally with the established knowledge in lithography domain, revealing the dominant influences of NTD parameters on the photoresist profile.