Neural ODE Granular Flow Models

ISEF Category: Engineering Technology: Statics and Dynamics

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

The Hook

Grains do not move like water. They can jam, arch, and suddenly pour faster when a clog breaks. That weird behavior makes hopper flow a real test for simulation. If you can predict it from video, you are doing the kind of modeling engineers use for silos, bins, and conveyor systems.

What Is It?

Granular flow is how lots of small solid pieces, like cereal, sand, or beads, move together. A hopper is a container with a hole at the bottom. When the grains drain out, they can flow smoothly, or they can clog and burst through in fits.

A neural ODE, short for neural ordinary differential equation, is a machine learning model that learns how a system changes over time. Think of it like a smart rulebook for motion. You give it video data from the flowing grains, and it learns the hidden parameters that describe the flow. In this kind of project, those learned parameters feed into a continuum or DEM-like simulator. A continuum model treats the grains like a bulk material, while DEM, or discrete element method, tracks many individual particles one by one.

The cool part is validation. You do not just fit one hopper and stop. You test whether the learned model still predicts flow in a different hopper shape. That checks whether your model learned real physics, or just memorized one setup.

Why This Is a Good Topic

This is a strong science fair topic because you can measure it from video, compare predictions against real drainage, and change one design factor at a time, like hopper angle, outlet size, or grain type. It connects to shipping, storage, agriculture, and factory handling, where clogging wastes time and money. You can learn computer vision, parameter fitting, simulation, and model validation, which are all core research skills.

Research Questions

• How does hopper outlet size change the neural ODE model's predicted drainage rate for cereal or sand?
• What is the effect of grain shape on the fitted flow parameters and clog frequency?
• Does a neural ODE trained on one hopper geometry predict drainage in a different geometry with acceptable error?
• To what extent does packing density before release change the model's ability to capture early flow behavior?
• Which video features, such as surface level, outlet width, or front shape, improve parameter estimation most?
• How does moisture level in sand or cereal clumping change the agreement between simulation and observed flow?
• What is the effect of training set size on the stability of the learned rheology parameters?

Basic Materials

• Clear plastic hoppers or funnel-shaped containers
• Sand, rice, cereal, or bead samples
• Smartphone or digital camera with tripod
• Meter stick or ruler in frame
• Printed scale markers for video calibration
• Consistent lighting setup
• Tray or tarp for collecting spilled material
• Notebook for logging geometry, material type, and trial order
• Computer with spreadsheet software
• Python installed with basic data analysis libraries.

Advanced Materials

• High-speed camera or high-frame-rate smartphone
• Transparent hopper set with interchangeable outlet geometries
• Precision scale for mass flow measurements
• Particle size analyzer or sieve set
• ImageJ or similar tracking software
• Python environment with machine learning and optimization libraries
• Access to a DEM package or continuum simulation software
• Motion tracking markers or calibration grid
• Environmental controls for humidity testing
• University lab access for repeatable flow rigs.

Software & Tools

• Python: Fits the neural ODE, processes video-derived measurements, and compares model predictions to experiments.
• ImageJ: Measures grain interface height, outlet width, and flow-front movement from video frames.
• OpenCV: Extracts frames and tracks motion features from hopper-drainage videos.
• Jupyter Notebook: Keeps your analysis, plots, and model tests in one reproducible place.
• MATLAB: Supports numerical fitting and simulation if your school already has access.

Experiment Steps

1. Define the flow feature you want to predict, such as drainage rate, surface height, or clog probability.
2. Choose one hopper geometry as your training system and a different geometry as your validation system.
3. Design a video workflow that gives you the same view, scale, and lighting for every trial.
4. Decide which input measurements your model will learn from, such as grain type, fill level, and outlet size.
5. Build a plan for converting video into numeric data before you fit the neural ODE.
6. Set a validation rule that tests whether the simulator predicts a new hopper better than a simple baseline model.

Common Pitfalls

• Changing camera angle between trials, which breaks the pixel-to-distance calibration.
• Using grains with mixed sizes or broken pieces, which changes flow behavior from run to run.
• Training the model on only one hopper shape, which makes the results look accurate but fail on new geometry.
• Ignoring clogs and restart events, which removes the most informative part of granular flow.
• Comparing simulation output to raw video without extracting the same physical quantity from both sides, which makes the validation unfair.

What Makes This Competitive

A stronger project does more than fit one dataset. It compares at least two hopper geometries, tests a clear baseline model, and reports error with a real validation metric. You can also raise the level by separating geometry effects from material effects, then checking which learned parameters stay stable across trials. If your model predicts a new setup well, and you can explain why, your project starts to look like real engineering research.

Project Variations

• Use rice, beads, or sand to compare how particle shape changes the learned flow law.
• Test whether a neural ODE trained on circular hoppers can predict flow in wedge-shaped or slot hoppers.
• Compare a video-only model against a model that also uses mass-flow data to see which predicts clogging better.

Learn More

• MIT OpenCourseWare: Search for courses on computational modeling, fluid mechanics, and numerical methods to build your simulation background.
• PubMed: Search review articles on granular flow, jamming, and hopper discharge for peer-reviewed overviews.
• NASA Technical Reports Server: Search for granular media, dust flow, and particle transport studies with useful modeling ideas.
• USGS Publications Warehouse: Search for sediment transport and granular behavior reports that connect particle flow to real systems.
• OpenStax University Physics: Use the mechanics sections for a clear refresher on forces, motion, and energy before modeling.