Graph Neural Network Truss Stress Prediction

ISEF Category: Engineering Technology: Statics and Dynamics

Subcategory: Computational Mechanics  ·  Difficulty: Advanced  ·  Setup: School Lab  ·  Time: Full Year

The Hook

A computer can predict where a bridge will crack before the bridge exists. That is the idea behind this project. You train a graph neural network on truss data, then test whether it can estimate stress in real printed structures. If the model works, you have a fast tool for design screening.

What Is It?

A truss is a framework of straight bars joined at nodes. Engineers use trusses in roofs, cranes, towers, and bridges because the shape can carry load well. In this project, you build a model that looks at the truss like a network, where each joint is a node and each bar is a connection.

A graph neural network, or GNN, is a machine learning model that learns from connected objects. Think of it like a social network app for structures. Each joint shares information with nearby joints, and the model learns patterns that relate shape, support locations, and loading to stress. Frame3DD can generate many simulated trusses with known stress values, which gives you training data before you test the model on real 3D-printed PLA trusses.

Why This Is a Good Topic

This is a strong science fair topic because you can test a clear engineering question, whether a machine learning model can predict structural stress from geometry and loading. You can measure accuracy, compare against simulation, and check where the model breaks down. The project connects to real design problems in bridges, buildings, and lightweight structures. You can learn structural mechanics, data preparation, model validation, and failure analysis in one project.

Research Questions

• How does truss geometry affect the accuracy of a graph neural network stress predictor? ?
• What is the effect of changing the number of training truss designs on prediction error? ?
• Does a graph neural network predict peak stress better than a simple baseline regression model? ?
• To what extent do support conditions change the model's error on unseen truss layouts? ?
• Which truss features, such as member length, angle, or connectivity, matter most for stress prediction? ?
• How does the model's prediction error compare between Frame3DD simulations and 3D-printed PLA test data? ?

Basic Materials

• Laptop or desktop computer with enough memory for machine learning work.
• Python installed with ML libraries such as PyTorch or TensorFlow.
• Frame3DD structural analysis software.
• 3D printer with PLA filament.
• Digital kitchen scale or hanging weight set for loading tests.
• Ruler or digital caliper for measuring printed truss dimensions.
• Safety glasses.
• Camera or smartphone for documenting deformation and failure.

Advanced Materials

• Access to a mechanical testing frame or instrumented loading rig.
• Load cell or force gauge for measuring applied load.
• Digital image correlation setup or calibrated camera system for displacement tracking.
• Finite element analysis software for comparison studies.
• High-resolution 3D scanner or coordinate measuring tool for print quality checks.
• Laboratory oven or controlled storage area for conditioning PLA samples.
• Strain gauges and data acquisition hardware for direct validation.

Software & Tools

• Python: Prepares truss data, trains the graph neural network, and evaluates prediction error.
• PyTorch Geometric: Builds graph neural network models for node and edge relationships in trusses.
• Frame3DD: Generates simulated truss stress values for training and testing data.
• NumPy and pandas: Clean structural data and organize model inputs and outputs.
• Matplotlib or Seaborn: Plot predicted stress, true stress, and error trends.

Experiment Steps

1. Define the exact stress target you want the model to predict, such as maximum member stress, nodal displacement, or failure load.
2. Choose the truss family you will study, then decide which geometry features the model will receive as inputs.
3. Generate a simulation dataset with Frame3DD, making sure your training set covers a wide range of shapes and support conditions.
4. Build a baseline model first, then compare it with the graph neural network so you can prove the GNN adds value.
5. Plan a validation set using unseen truss designs, then reserve some printed structures for final testing only.
6. Design the physical test plan so you can compare simulated predictions with real failure behavior in a fair way.

Common Pitfalls

• Training on trusses that are too similar, which lets the model memorize shapes instead of learning structure patterns.
• Mixing simulation outputs from different stress definitions, which makes the target variable inconsistent.
• Ignoring print defects in PLA, which can make real failure loads much lower than the model expects.
• Comparing simulation and experiment without matching support conditions, which creates fake error.
• Using too few truss designs, which makes the model look accurate on paper but weak on new geometries.

What Makes This Competitive

A stronger version of this project goes beyond basic prediction accuracy. You can test whether the model generalizes to truss shapes it never saw, or to print quality changes and loading differences. You can also compare several model types, then explain why the graph model helps or fails. Careful error analysis, clean validation splits, and a real-world test set make the project much stronger than a simple demo.

Project Variations

• Use Warren trusses instead of one truss family, then test whether the model transfers across geometry styles.
• Predict displacement instead of stress, then compare which target is easier for the network to learn.
• Add print-quality features, such as measured bar thickness or joint error, and test whether they improve real-world prediction.

Learn More

• MIT OpenCourseWare: Search for structural mechanics, finite element analysis, and machine learning courses that cover the math behind truss behavior.
• Frame3DD documentation: Read the software manual and example files to understand truss inputs, outputs, and load cases.
• PyTorch Geometric docs: Learn how graph neural networks represent nodes, edges, and message passing.
• NIST Engineering Statistics Handbook: Find guidance on model validation, error metrics, and uncertainty analysis.
• PubMed: Search for review articles on graph neural networks in engineering design and structural prediction.
• NASA Technical Reports Server: Search for structural health monitoring and machine learning reports that use network-based models.