Physics-Informed Machine Learning for Real-World Applications

The field of machine learning is shifting towards incorporating physical priors and inductive biases to improve model efficiency and efficacy. Researchers are exploring the use of physics-informed models to address real-world challenges, such as predicting pedestrian trajectories in hazy weather conditions, counting crowds in foggy environments, and forecasting urban microclimates. These models have shown significant improvements in performance and efficiency, reducing carbon footprints and computational resources. Notable papers in this area include:

  • A paper proposing a deep learning model that combines physical priors of atmospheric scattering with topological modeling of pedestrian relationships, achieving a 78% inference speed increase and reducing minADE/minFDE metrics by 37.2%/41.5%.
  • A paper introducing UrbanGraph, a physics-informed framework integrating heterogeneous and dynamic spatio-temporal graphs for urban microclimate prediction, improving R^2 by up to 10.8% and reducing FLOPs by 17.0%.
  • A paper presenting CarbonX, an open-source tool for computational decarbonization using Time Series Foundation Models, achieving a zero-shot forecasting Mean Absolute Percentage Error of 15.82% across 214 grids worldwide.

Sources

Hazy Pedestrian Trajectory Prediction via Physical Priors and Graph-Mamba

Trading Carbon for Physics: On the Resource Efficiency of Machine Learning for Spatio-Temporal Forecasting

Foggy Crowd Counting: Combining Physical Priors and KAN-Graph

Physics-informed learning under mixing: How physical knowledge speeds up learning

UrbanGraph: Physics-Informed Spatio-Temporal Dynamic Heterogeneous Graphs for Urban Microclimate Prediction

CarbonX: An Open-Source Tool for Computational Decarbonization Using Time Series Foundation Models

Built with on top of