Speaker
Description
Chemical Vapor Deposition (CVD) optimization is critical for advancing thin-film quality and process efficiency in semiconductor and optoelectronic applications, yet traditional methods like Computational Fluid Dynamics (CFD) simulations and empirical tuning are often computationally intensive or lack adaptability. To address this challenge, this study presents a data-driven machine learning (ML) framework that optimizes two key CVD performance metrics—deposition rate and film uniformity—by integrating XGBoost-based predictions with polynomial surrogate modeling for computationally efficient, constraint-aware process improvement. Building on prior work where XGBoost was identified as the most accurate predictor of CVD outcomes compared to other ML models and CFD benchmarks, this research advances the field by constructing surrogate equations from XGBoost-predicted outputs over a test dataset, enabling rapid optimization via the Sequential Least Squares Programming (SLSQP) algorithm. Three optimization scenarios were systematically evaluated: (i) maximizing deposition rate under a uniformity constraint, ensuring high throughput without compromising film quality; (ii) minimizing non-uniformity while meeting a minimum deposition threshold, critical for precision applications; and (iii) maximizing the deposition-to-uniformity ratio, a balanced metric for industrial scalability. The optimized process parameters—susceptor temperature and inlet gas velocity—were derived for each case, revealing distinct operational regimes: high-temperature, low-velocity conditions maximized deposition rate (achieving a 12% improvement over baseline), while moderate-temperature, high-velocity setups minimized non-uniformity (to <5% variation), and an intermediate trade-off regime optimized the deposition-to-uniformity ratio (yielding a 15% gain). Comparative analysis with traditional gradient-based optimization techniques demonstrated that the proposed ML-driven approach not only matched accuracy but also reduced computational time by 40 times, highlighting its suitability for real-time adaptive control in industrial settings. The success of this framework lies in its ability to replace costly CFD iterations with lightweight surrogate models while maintaining fidelity, a breakthrough for intelligent manufacturing and digital twin implementations. Beyond CVD, the methodology’s flexibility combining ML with constrained optimization offers broad applicability in other thin-film deposition techniques and complex material synthesis processes. Future directions include embedding this framework with in-situ sensor feedback for closed-loop control and extending it to multi-objective scenarios incorporating additional film properties (e.g., stress, stoichiometry). By bridging the gap between data-driven modeling and industrial process optimization, this work contributes a scalable, efficient paradigm for next-generation manufacturing, aligning with Industry 4.0 advancements in automation and predictive analytics.