Reconstruct continuous physical fields from sparse sensors.
What if this were enough?
Point measurements around a body in flow.
The physical world is continuous. Our measurements are not.
Grey: never measured.
Learn from what can be measured. Infer what cannot.
Green: being inferred. Lines: reconstructed flow.
Evidence · arXiv 2602.13873
Reconstruction from 3% uniformly random measurements across Darcy flow, Helmholtz, Navier–Stokes and Poisson. Every baseline trains on complete fields. Ambient Physics never sees one.
Results in the paper ↗Wherever dense sensing is expensive, hazardous, slow, or impossible.
Complete measurements should not be a prerequisite for learning physics.
Harris Abdul Majid · Giannis Daras · Francesco Tudisco · Steven McDonagh
The University of Edinburgh · Massachusetts Institute of Technology
ICML 2026 Workshop on AI for Physics