Date of Submission

5-2026

Document Type

Dissertation

Degree Name

Doctor of Philosophy Engineering and Applied Science, Ph.D.

Department

Engineering and Applied Science Education

Advisor

Ganesh Balasubramanian, Ph.D.

Committee Member

Ronald S. Harichandran, Ph.D., P.E., F.ASCE

Committee Member

Sumith Yesudasan, Ph.D.

Committee Member

Mohamad Nassar, Ph.D.

LC Subject Headings

: Composite materials industry, Machine learning, Plastics—Extrusion, Ceramic-matrix composites, Process control, Composite materials--Mechanical properties

Abstract

Composite manufacturing increasingly requires methods that can support both process monitoring and materials design under conditions of limited sensing, nonlinear behavior, and sparse data. This work addresses that need through machine learning (ML) based studies in two complementary areas: process intelligence for polypropylene extrusion and material intelligence for SiC /SiC mini-composites. The first part focuses on industrial polypropylene extrusion, where melt temperature and melt pressure are important indicators of process stability and product quality, yet remain difficult to estimate and forecast in real time. An explainable ML workflow is first developed for melt-state prediction using industrial extrusion data. The approach combines lag alignment, Butterworth filtering, engineered features, and ensemble learning. On filtered holdout data, LightGBM achieves an RMSE of 0.0367 °C for melt temperature, while gradient boosting regression achieves an RMSE of 0.0032 MPa for melt pressure. SHAP-based analysis further identifies barrel-zone temperatures, screw speed, and throughput-related features as dominant drivers of the predictions. The study then extends from present-state estimation to short-horizon forecasting through a hybrid SARIMA + Random Forest model. By combining seasonal-linear time-series structure with nonlinear residual correction, the hybrid model outperforms the standalone forecasting models and achieves MAE/RMSE values of 0.07 /0.091 for melt temperature and 0.015/0.017 for melt pressure over a six-step forecasting horizon. The third extrusion study addresses a more demanding problem: high-frequency, long-horizon forecasting. For this purpose, a decomposition-linear residual correction network (DLRC-Net) is introduced. Across horizons from 96 to 720 steps, DLRC-Net shows strong performance for melt temperature and remains competitive for melt pressure. Ablation analysis further shows that most of the forecasting strength comes from the decomposition step and the linear backbone, while the nonlinear correction branch provides more selective, target-dependent benefit. The final part shifts from process monitoring to materials design by addressing ultimate tensile strength (UTS) prediction for SiC/SiC mini-composites using acurated literature-derived dataset. To manage missing values and target imbalance, the study combines KNN imputation, alternative rebalancing strategies, and ensemble learners including Random Forest, XGBoost, and LightGBM. The bestperforming model, XGBoost with SMOGN-based oversampling, achieves a mean absolute error of 60.1 MPa and a mean absolute percentage error of 8.8%. Featureimportance analysis further identifies key tensile, interfacial, and thermomechanical descriptors associated with composite strength. Taken together, these studies show that ML can support both how composite materials are processed and how they are designed. In extrusion, it enables more accurate melt-state prediction and improved forecasting over short and long horizons. In SiC/SiC mini-composites, it supports strength estimation and feature interpretation under sparse-data conditions. The overall contribution is a unified view of predictive composite manufacturing in which ML serves as a practical tool for process-state estimation, forecasting, and materials-property prediction.

Download

Share

COinS