Singing Sculpture Sound Pattern Classifier

ISEF Category: Technology Enhances the Arts

Subcategory: Engineering Effects  ·  Difficulty: Advanced  ·  Setup: University Lab  ·  Time: Full Year

The Hook

A metal plate can act like a visual sound fingerprint. Tap the right frequency, and sand or beads jump into sharp lines and rings called Chladni patterns. Now imagine turning those patterns into a classifier for rain, traffic, and voice. That blends art, sensing, and machine learning in one project.

What Is It?

This project uses a piezo actuator to drive a metal plate at specific sound frequencies. When the plate vibrates, the material on top gathers into patterns that mark the plate’s vibration modes. Those mode shapes can change with the sound that drives them, almost like a drum skin that draws its own map.

You then treat each pattern like an image input for a convolutional neural network, or CNN, which is a model that learns visual features. In simple terms, the CNN looks for repeated shapes, edges, and textures that separate one class from another. Your job is to test whether patterns created from environmental sounds can help the model tell rain, traffic, and voice apart.

Why This Is a Good Topic

This is a strong science fair topic because you can change one input at a time, collect repeatable images, and compare how well a model classifies each sound class. It connects to real problems in acoustic sensing, smart environments, and non-contact sound recognition. You can learn vibration physics, imaging, data labeling, and model evaluation without needing a clinical or wet biology lab.

Research Questions

• How does the driving frequency affect the clarity of Chladni mode boundaries for each sound class?
• What is the effect of plate shape or mounting style on CNN classification accuracy?
• Does adding a controlled texture layer, such as sand versus fine salt, improve pattern separation?
• To what extent can pattern images from rain, traffic, and voice be classified better than raw audio features alone?
• Which image preprocessing method produces the most stable CNN performance across repeated trials?
• How does the signal amplitude affect the reproducibility of mode-shape images within the same sound class?

Basic Materials

• Piezo actuator or vibration exciter.
• Metal plates of different shapes and thicknesses.
• Audio source with recorded environmental sounds.
• Function generator or audio playback interface.
• Amplifier matched to the actuator.
• Fine sand, salt, or lycopodium powder.
• Smartphone or digital camera with manual settings.
• Tripod or fixed camera mount.
• Ruler or calipers.
• Notebook for trial logging.

Advanced Materials

• Laser vibrometer or accelerometer.
• High-speed camera or DSLR with manual exposure control.
• DAQ interface for collecting actuator and response data.
• Precision function generator.
• Interchangeable plate materials, such as aluminum, steel, and brass.
• Vibration isolation table or sturdy isolation base.
• Reference microphone with calibrated response.
• Computer with GPU access for model training.

Software & Tools

• Python: Processes images, organizes labels, and trains the classification model.
• ImageJ: Measures pattern contrast, area coverage, and edge sharpness in each trial image.
• Audacity: Cleans and segments environmental sound recordings before they drive the actuator.
• Google Colab: Runs CNN training and testing in a free cloud notebook.
• Fiji: Helps batch-process image stacks and standardize brightness or thresholding.

Experiment Steps

1. Define the sound classes you want to compare and decide how many repeated trials each class needs.
2. Choose one plate geometry and one imaging setup so your data stay comparable from run to run.
3. Plan how you will convert each sound sample into a vibration input, then into a labeled image dataset.
4. Build a standard curve or reference set so you can compare pattern strength, clarity, or symmetry across samples.
5. Decide which CNN input format and preprocessing steps you will test first, such as cropping, grayscale conversion, or thresholding.
6. Set evaluation rules before training, including train-test splits, accuracy, confusion matrix metrics, and repeatability checks.

Common Pitfalls

• Letting room noise or stray bumps trigger the plate, which mixes unwanted vibration into your class labels.
• Changing camera distance or lighting between trials, which makes the image data look different for the wrong reasons.
• Using one plate shape for every sound class without enough repetition, which hides whether the model learned the class or just the setup.
• Skipping calibration of the actuator output, which means a louder recording may reflect input strength, not mode shape differences.
• Training the CNN on too few images, which can create high accuracy on paper but weak performance on new trials.

What Makes This Competitive

A stronger version of this project goes beyond a simple photo classifier. You can compare multiple plate geometries, test whether the model still works under added noise, and separate the effect of frequency from the effect of sound class. A competitive entry also reports confusion matrices, cross-validation, and failure cases, not just accuracy. If you can show which design choices make the pattern-based classifier more reliable, you will have a much stronger engineering story.

Project Variations

• Test whether circular, square, and rectangular plates produce different classification accuracy for the same sound classes.
• Compare sand, salt, and powder as the visible medium for Chladni pattern capture.
• Train one model on raw pattern images and another on edge-enhanced images to see which input works better.

Learn More

• MIT OpenCourseWare: Search for vibration, waves, and machine learning courses that explain the physics and modeling behind this project.
• PubMed: Search review articles on acoustic sensing, vibration-based classification, and sound recognition methods.
• NASA Earth Observatory: Browse background on environmental sound and remote sensing-style classification thinking.
• NOAA Climate.gov: Use background articles to understand rain and weather sound patterns as real environmental signals.
• ImageJ documentation: Find guides for image thresholding, segmentation, and measurement on the Fiji or ImageJ help pages.