(A) Context- and Physics-Aware Learning for Short- and Long-Term Mobility Forecasting

In this research line, we aim to develop advanced forecasting models that integrate contextual geospatial features and physical motion principles to improve the short- and long-term prediction of object trajectories in open environments. We aim to incorporate Newtonian dynamics into neural networks, advancing Physics-Informed Neural Networks for mobility systems. By combining physical constraints with contextual inputs such as ocean currents and weather conditions, we aim to develop systems that deliver realistic and consistent forecasts of mobility patterns. A key aspect is modeling the behavior of multiple moving entities, such as vessels, in loosely coordinated or adversarial settings where traditional models may struggle.

(B) Machine Learning-based Climate Matching for Species Risk Invasion on a Global Scale

In this research line, we aim to develop machine-learning models that match climate patterns across biogeographic regions to identify ecological analogs where invasive species may be successfully established. By aligning multidimensional climate profiles (e.g., temperature, precipitation, and seasonality), we intend to construct a scalable framework for estimating species introduction risks in novel environments through link prediction, probabilistic forecasting, and reinforcement learning. The models are trained using global climate datasets and species distribution records, with future integration of biodiversity and land-use layers.

(C) GeoAI and Underwater Acoustics for Regionalized Learning with Synthetic Data

In this research line, we aim to integrate geospatial artificial intelligence with underwater acoustic modeling to enable regionally adaptive learning frameworks for marine applications. By leveraging synthetic acoustic datasets informed by geographic, bathymetric, ecological, and vessel presence priors, we intend to develop models capable of performing classification, localization, and detection tasks without exposing sensitive signal properties or operational data. The methodology combines physics-informed generative modeling, spatial domain adaptation, and constrained optimization to support learning in data-sparse settings.