Théo Onillon

Visiting Student Researcher

– Developed an end-to-end ML pipeline for city-scale human mobility prediction, inferring hourly origin-destination (OD) flows from heterogeneous spatial datasets across the Bay Area, Boston, and Los Angeles. – Designed a Graph Attention Network (GAT) with embedded physical constraints (gravity laws, distance decay) for robust spatiotemporal forecasting of OD flows. – Explored purpose-decomposed OD flow prediction using the TimeGeo activity-based dataset, yielding promising preliminary results on temporal dynamics of work, home and shopping trips. – Developed reproducible experiment workflows, ablation studies and evaluation pipelines; contributed to publication-grade research deliverables under Prof. Marta C. González and Prof. Maria Laura Delle Monache.

– Contributed to the eMob-Twin project (city-scale electromobility digital twin) under Dr. Carlos Canudas de Wit.

– Developed hybrid physics-informed ML models for vehicle mobility forecasting, enforcing mass-conservation constraints to ensure physical consistency. – Benchmarked multiple modeling approaches and designed evaluation pipelines preserving dynamical system laws.

One-month technical internship focused on data preparation for satellite-based flood detection models.

Performed manual annotation of satellite imagery to build ground-truth datasets for machine learning models.

Developed Python scripts to automate repetitive preprocessing tasks and speed up the labeling process.