A 17-Year-Old Used AI to Predict 3D Print Strength Without Breaking Parts

Mason Corey, a 17-year-old from Swedesboro, New Jersey, built a low-cost system that watches 3D-printed parts cool and then predicts how strong they will be. The catch? The system never touches the part. It uses background-oriented schlieren imaging, a technique that makes invisible air currents visible, paired with a machine-learning model trained on cooling patterns.

Seeing the Invisible

Schlieren imaging captures tiny refractions of light caused by temperature changes in the air. When a freshly printed PLA specimen cools, it creates a thermal plume. Corey's setup used a Nikon D750 camera and a pair of hairdryers to create controlled cooling conditions, then recorded the airflow around 30 printed specimens as they cooled.

The camera captured images at nine layers during printing across six different cooling scenarios. Those images became the training data. An initial analysis found a promising relationship between the thermal patterns and tensile strength, with an R-squared value of 0.808. That number suggested the approach might actually work.

The Hard Part

When the researchers applied stricter cross-validation, the correlation dropped to 0.301. The model had memorized the training set rather than learned a general rule. Corey, who co-authored the study with Widener University researchers Kyle Weber and Babak Eslami, is upfront about the limitation. The work is a proof of concept, not a production-ready tool.

The study, published in the journal Polymers, makes that boundary clear. Thirty specimens is a small dataset. The technique works well enough to tell cooling conditions apart, but it cannot yet predict the strength of an arbitrary printed part with confidence.

Why It Matters

Conventional tensile testing destroys the part. Other quality-control methods require specialized equipment. A camera-based system that infers strength from cooling behavior could, in theory, let manufacturers spot inconsistencies without sacrificing every unit. Corey's bill of materials for the imaging system came to roughly one thousand dollars, which is cheap by materials-testing standards.

The project was selected as a finalist in the 2026 Regeneron Science Talent Search. That recognition reflects the experimental design more than a finished product. Corey and his collaborators plan to expand the dataset to more than 200 specimens across multiple materials. If the correlation holds at that scale, the method could become a practical quality-control tool for filament-based manufacturing.