2026 Greater San Diego Science and Engineering Fair

Optimizing Sustainable Materials
Senior Sophia Monsowitz investigated bacterial cellulose as a good, eco-friendly alternative to petroleum-based plastics. Her research focused on how various drying methods impact the tensile strength, mass loss, and drying time of kombucha-derived cellulose particles. The following drying methods were compared: incubator drying, desiccation drying, air-drying and solvent-exchange air-drying. She hypothesized that aggressive drying would cause structural inconsistencies, whereas controlled methods would yield more stable performance. After testing each method under identical fermentation conditions, her results showed that mass loss was highest in incubator-dried samples, followed by desiccation and air-drying. Notably, SCOBY pellicles treated with ethanol exchange exhibited the highest tensile strength.
Though Monsowitz noted that a simplified setup led to some pre-test sample fracturing, her findings were clear: faster drying methods better preserve the integrity of the fiber network and hydrogen bonding. Her study concludes that the choice of drying method fundamentally dictates the rate and pattern of moisture loss, directly influencing the material’s durability.

Predicting Bee Colony Collapse
Senior Charles Brum developed a sophisticated machine learning approach to predict and locate honeybee colony collapses. Utilizing a dataset of over 55,000 rows spanning from 2015 to 2024, Brum’s model analyzed 28 distinct input parameters. The neural network architecture featured two hidden layers (256 and 128 neurons, respectively) and utilized Mini-Batch Gradient Descent to optimize the reward and loss systems.
To eliminate potential software bias, Brum implemented his architecture across two different frameworks: PyTorch and TensorFlow. After standardizing the data using Python, he compared the performance of both systems. His findings revealed that TensorFlow was more effective at handling higher predicted loss scenarios, maintaining low mean average errors for predictions in the 15–20% and 20–25% loss ranges.
While Brum noted that error rates fluctuated by state based on local independent variables, his research ultimately suggests that TensorFlow provides a more efficient framework for modeling complex ecological data like colony collapse.