A Unified Mechanics Theory-Informed AI LLM for Structural Analysis and Design of Bridges
Artificial Intelligence for Designing Bridges.
Project description
The state-of-the-art structural analysis and design of bridges are based on Newton's universal laws, which do not account for aging and degradation of structures. Therefore, the design is valid only for the first day of the structure. As structural stiffness degrades, sagging deflections increase, and eventually the structure fails due to fatigue. Therefore, engineers conduct accelerated laboratory experiments to predict the structure's life expectancy. Because it is not possible to predict the degradation behavior over time using Newtonian mechanics. Recently developed Unified Mechanics Theory unifies the laws of Newton and the second law of thermodynamics at an ab initio level. As a result, it is possible to design with the knowledge of how the bridge will perform over the years and when it should be taken out of service.
To achieve the project's goals, we will need to train and fine-tune MechGPT, [based on the Meta Large Language Model (LLM)]. MechGPT was trained in the atomistic modeling of materials failure. We need to build on its existing knowledge. First, we need to train MechGPT for Unified Mechanics Theory. Then we need to train MechGPT for structural analysis methods. Each training session requires testing with prompts to ensure accuracy. If it is hallucinating, we need to fine-tune it.
Additionally, we must ensure it does not forget prior knowledge. This process typically involves asking thousands of questions, verifying accuracy, and providing feedback if the answer is incorrect. The final product will perform a structural analysis of a bridge and predict its life expectancy, with minimal input from an engineer. This will be the first application of Unified Mechanics Theory in Structural Engineering in the literature.
Project outcome
I. Foundation Weeks (1-4): Student to conduct a traditional structural analysis and traditional life expectancy estimate for a target bridge using classical methods (Newtonian mechanics). Provide access to baseline structural models, relevant design codes, and introductory materials on structural degradation.
II. Theory & Training Weeks (5-10) Assign the student to study introductory materials on AI LLM, UMT, and the concept of merging Newton's laws with the second law of thermodynamics. Explain the architecture of MechGPT and its existing knowledge base. Identify and categorize the specific structural analysis methods to train the LLM. Conduct weekly tutorial sessions on UMT concepts. Explain the high-level data flow and parameters in the LLM fine-tuning process.
III. Testing, Validation, and Final Product Weeks (11-30). Develop thousands of structured prompts covering classical structural analysis to fine-tune MechGPT. Verify the accuracy of MechGPT outputs against the known solutions. Systematically provide feedback for fine-tuning based on identified "hallucinations" or inaccuracies.
IV. Review and approve prompt sets with the student. Discuss strategies for troubleshooting LLM "hallucinations." Provide constructive feedback on the report section. The student will be trained to contribute to the final report section documenting the validation process manual.
Weekly mentoring meetings with my students are 2 hours long. I will review weekly progress, troubleshoot technical issues, discuss the theory behind the week's work, and plan next steps. I will also provide informal/on-demand support daily (via chat/email for quick questions, resource requests, and immediate feedback on small tasks
Students will also be required to provide a progress Presentation, a mid-project presentation, and a final presentation. Students will be asked to present findings, challenges, and contributions to the project team to practice technical communication skills.
Performance will be assessed based on: Technical deliverables, conceptual understanding, professionalism, communication, and quality of the work.
Project details
| Length of commitment | Longer than a semester (about 6-9 months) |
| Start time | Anytime |
| In-person, remote, or hybrid? | In-Person |
| Level of collaboration | Small group project (2-3 students) |
| Benefits | Stipend |
| Who is eligible | Juniors and Seniors |
Core partners
- Prof. Cemal Basaran
- Mr. Brandon McDonald, PhD student
- Ms. Jahnavi Kollu, MS student
Project mentor
Cemal Basaran
Professor
Civil, Structural and Environmental Engineering
Phone: (716) 645-4375
Email: cjb@buffalo.edu
Start the project
- Email the project mentor using the contact information above to express your interest and get approval to work on the project. (Here are helpful tips on how to contact a project mentor.)
- After you receive approval from the mentor to start this project, click the button to start the digital badge. (Learn more about ELN's digital badge options.)
Preparation activities
Once you begin the digital badge series, you will have access to all the necessary activities and instructions. Your mentor has indicated they would like you to also complete the specific preparation activities below. Please reference this when you get to Step 2 of the Preparation Phase.
Students MUST complete the Intro to CCR course in UB Learns before gaining access to the course allocations. They will receive a certification of completion for the course, which they can share with the professor or TA. More info on that is here:
https://docs.ccr.buffalo.edu/en/latest/getting-access/#intro-to-ccr-course-requirement
Keywords
engineering, computer science, artificial intelligence, civil engineering, large language model