Surrogate Model-Assisted Machine Learning Optimization of Chemical Vapor Deposition for Enhanced Film Deposition Rate and Uniformity

Research Hypothesis:
This research hypothesizes that a machine learning–assisted optimization framework can significantly enhance the efficiency and accuracy of optimizing key performance indicators—film deposition rate and thickness uniformity—in a Chemical Vapor Deposition (CVD) system. By leveraging XGBoost-predicted outputs and constructing surrogate models, it is possible to identify optimal process conditions with substantially reduced computational cost compared to direct CFD-based optimization.

What the Data Shows:
The dataset contains the predicted results from a pre-trained XGBoost regression model applied to a set of test cases derived from CFD simulations of a vertical CVD reactor. Each data point includes the input parameters (susceptor temperature and inlet gas velocity) and the corresponding model-predicted deposition rate and uniformity. These predictions were used to develop second-order polynomial surrogate models that accurately represent the relationship between the process inputs and outputs. The dataset supports three optimization scenarios: (i) maximizing deposition rate under a uniformity constraint, (ii) minimizing non-uniformity under a minimum deposition rate constraint, and (iii) maximizing the deposition-to-uniformity ratio.

Notable Findings:

XGBoost provided the most accurate predictions among tested ML models when compared with CFD simulation results, demonstrating high fidelity in capturing nonlinear behaviors of the CVD process.

The polynomial surrogate models achieved high R² scores and enabled rapid optimization using the Sequential Least Squares Programming (SLSQP) algorithm.

Optimal process windows were identified with improved uniformity and deposition performance, outperforming traditional optimization approaches in both accuracy and computational speed.

Description and Interpretation:
The dataset includes input parameters (temperature and velocity), XGBoost-predicted outputs (deposition rate and uniformity), and results from surrogate model-based optimization for each scenario. The data is structured in a CSV/Excel file format, with clearly labeled columns for inputs, outputs, and optimization metrics. The surrogate models were derived using polynomial regression and validated against the original XGBoost predictions.

This dataset can be interpreted to understand how changes in reactor operating conditions affect deposition characteristics and how machine learning can streamline optimization in CVD applications. It is especially useful for researchers developing intelligent control systems, digital twin frameworks, or data-driven design tools for thin-film fabrication. The surrogate models included enable rapid response surface analysis and can be integrated into real-time optimization pipelines for smart manufacturing.