Date of Award

2026-08-01

Degree Name

Master of Science

Department

Manufacturing Engineering

Advisor(s)

Amit Lopes

Abstract

The computational bottleneck of structural feasibility screening frequently hinders the transition from a digital 3D model to a physically manufactured component. Traditionally, engineers have relied on either overly rigid heuristic constraint checks or computationally exhaustive finite element simulations. To address this inefficiency, this thesis proposes and validates a hybrid machine-learning framework to predict the structural manufacturability of 3D mechanical designs. Moving beyond the conventional reliance on isolated scalar parameters, the proposed methodology extracts and integrates both scalar manufacturing constraints (e.g., tolerance, minimum feature thickness) and spatial geometric descriptors (e.g., bounding volume, aspect ratio) directly from STL mesh data. Utilizing a Random Forest ensemble classifier, the hybrid framework achieved over 93% internal validation accuracy. Mathematical feature importance analysis confirmed the thesis hypothesis, proving that physical feasibility is a highly non-linear, emergent property driven by the interaction between scalar and geometric domains. To ensure industrial robustness, the algorithm was evaluated on an independent dataset (Dataset B) using an out-of-sample generalization test. Under this simulated domain shift, the framework maintained a robust baseline predictive accuracy of 84.4% and an exceptional Receiver Operating Characteristic Area Under the Curve (ROC-AUC) of 0.897. Furthermore, to transition the framework from a theoretical classifier into a practical manufacturing tool, the model's ensemble probabilities were statistically calibrated using isotonic regression. This enabled the deployment of a threshold-based "Review-Zone" triage policy. When applied to unseen data, the calibrated system autonomously approved or rejected 67.8% of the workflow volume, achieving a highly elevated functional accuracy of 92.8%, and intelligently isolated only the most ambiguous 32.2% of designs for manual engineering review. This study presents a novel integration of scalar design parameters and 3D geometric features for early-stage feasibility prediction, addressing a key limitation in existing approaches that rely on either parameter-based or geometry-based analysis alone. The proposed hybrid framework demonstrates superior predictive performance and provides a scalable, data-driven tool for intelligent design decision-making in modern manufacturing systems. Ultimately, this research demonstrates that hybrid feature integration and probabilistic calibration effectively bridge the gap between theoretical data science and practical, risk-aware design validation in smart manufacturing pipelines.

Language

en

Provenance

Received from ProQuest

File Size

89 p.

File Format

application/pdf

Rights Holder

Md Mohsin Uddin Fahim

Download

Share

COinS