A thesis submitted in partial fulfillment of the requirements for the degree of Master of Engineering in Structural Engineering

Thesis (M. Eng.) - Asian Institute of Technology, 2025

This thesis presents the development of an AI-assisted framework for the preliminary design and optimization of prestressed concrete (PSC) I-girder bridge superstructures using Artificial Neural Networks (ANN) and a Large Language Model (LLM)-based design assistant. The framework addresses the challenges of time-consuming manual processes and high computational costs associated with traditional bridge design methods. A synthetic dataset comprising 7,560 parametric bridge designs was generated using CSi Bridge Express, incorporating configuration by varying span lengths, number of lanes, girder types and spacing, slab thicknesses, and number of interior diaphragms, all in compliance with AASHTO LRFD 2020 specifications.Multiple ANN models are trained to predict structural design checks, reinforcement quantities, internal forces, prestressing losses, and Bill of Quantities (BOQs). These models achieved high performance, with classification accuracies exceeding 99% and regression R² scores above 98% for most outputs. The trained models are integrated into a desktop-based design tool, built using the WPF framework and connected to a Flask backend. A Gemini-based LLM agent interprets user prompts and autonomously routes the requests to appropriate ANN tools based on intent.Three intelligent tools are developed: Tool-1 for complete bridge design generation, Tool-2 for individual span design exploration, and Tool-3 for parametric variation based design evaluation. The output bridge designs are ranked based on material cost and validated by comparison with CSi Bridge Express design results. Most predictions fell within a ±10% error range, confirming the framework{u2019}s accuracy.The developed system enables rapid and cost-effective design decision-making, significantly enhancing productivity in early bridge planning stages. While the framework demonstrated high accuracy for deck and diaphragm predictions, refinement is recommended for girder shear reinforcement and long-term prestress losses. This research highlights the transformative potential of AI in structural engineering, offering a scalable and intelligent solution for the preliminary design of PSC I-girder bridges.