Ana Santos Lopes

Ana Santos Lopes

Biography

Ana Santos Lopes is a rising junior at Bowdoin College pursuing a combined major in
Computer Science and Mathematics. Originally from Rio Grande do Norte (Brazil), she is
passionate about combining technology and neuroscience to address real-world challenges,
expand access to educational opportunities, and improve healthcare systems. During her first
research experience, Ana collaborated with Brazilian researchers to develop tools that integrated
virtual reality, EEG signals, and assistive technologies to improve rehabilitation for patients
with spinal cord injuries. After this experience, she interned at EPFL in Switzerland, where she
worked on EMG analysis and processing, applying computational methods to extract patterns
from physiological data. Currently, Ana is interested in understanding how LLMs can help people
learn more effectively, focusing on improving education while expanding opportunities for
underprivileged communities. Beyond her academic and research pursuits, Ana enjoys writing,
reading, and having fun with tech development. She hopes to build a career in Human-Centered AI.

Machine Learning Classification of Spices From Sensor Data After Odor
Source Removal
Ana Santos Lopes1 and Paul Liang2
1Department of Computer Science and Mathematics, Bowdoin College
2Department of Media Arts and Sciences, Massachusetts Institute of Technology

The ability to smell is important in how humans perceive and interact with their environment,
as odor-associated chemical signals can persist after their source has been removed, with
potential applications in forensic science and allergy detection. Despite its importance, the
detectability of these signatures after source removal remains underexplored. This study
investigates whether machine learning (ML) models can classify spices using sensor data
collected after source removal. To address this question, sensor data were collected from
four ground spices (black pepper, nutmeg, paprika, and rosemary) along with environmental
controls using a box-shaped sensor array. Each experiment included an initial 2 minute
ambient recording, a 5 minute recording with the substance inside the box, and a post-removal
recording lasting either 5 minutes or up to 8 hours. The data collection comprised 104 shortruns
and 19 long runs. ML classifiers were trained on short-runs and evaluated in two settings:
short-to-short and short-to-long. An Extra Trees classifier achieved the highest overall accuracy
in the short-to-short evaluation, reaching 97.5%. The short-to-long examined classification
performance as sensor signals evolved after source removal. These findings demonstrate
that classification remains possible for a period after source removal, supporting further
investigation of odor signature persistence.