A Neural-Ordinary-Differential-Equations Based Generic Approach for Process Modeling in DTCO: A Case Study in Chemical-Mechanical Planarization and Copper Plating

Process modeling is a cornerstone for design-technology co-optimization, especially in achieving high-volume manufacturing. Data-driven process models are susceptible to overfitting due to a lack of comprehensive understanding of the underlying process mechanisms. Semi-physical process models often fall short in accurately predicting advanced nodes, primarily due to oversimplified process mechanisms. To address these challenges, we propose a generic process modeling approach based on neural ordinary differential equations. This scheme integrates ODEs to enable time-evolution prediction of process topography while incorporating neural networks to depict process mechanisms in a data-driven manner. We apply the generic modeling scheme to chemical mechanical polishing and copper electroplating processes, considering they serve as quintessential representations of process modeling challenges. Through silicon data on 28/32/40 nm nodes, the accuracy of our proposed models outperforms traditional semi-physical models and can effectively compete with data-driven approaches. These findings underscore the considerable advantages of our proposed modeling scheme in terms of accuracy enhancement for existing models, automated model construction, more accurate modeling of complex mechanisms than the phenomenological modeling strategy and reducing the risk of overfitting. Moreover, the successful application of our method to diverse processes underscores its generality and potential for extension to other manufacturing processes.