GeoPFN supplementary animations: sequential sensor placement and continuous sensing path

Overview

These two animations show GeoPFN reconstructing six synthetic geological fields (layers, channel, lenses, fault, channel network and weathered bedrock) as noisy sensor readings accumulate. GeoPFN recomputes the predictive mean and both uncertainty maps after every new reading, in less than a second on a GPU, illustrating rapid map updating during sequential sensing. They supplement Section 5.4 and Figure 13 of the GeoPFN preprint.

• Video 1, sequential sensor placement: starting from four sensors, each new sensor is placed where the current structural standard deviation is highest, up to 256 sensors.
• Video 2, continuous sensing path: a single moving sensor crosses the domain and then moves to the neighbouring cell with the highest structural standard deviation, for 2,048 moves.

In both videos, the structural uncertainty concentrates along unresolved contacts, channel margins, lens boundaries and the fault zone, so new readings accumulate near these features. The measurement uncertainty varies smoothly with the sensor layout and does not follow geological features.

Keywords

uncertainty quantification, adaptive sampling, site characterisation

Related works

• Preprint: https://doi.org/10.31223/X5JV68
• GeoPFN code and trained model used: https://doi.org/10.5281/zenodo.23984593
• Geoprior 2-D dataset (the synthetic geological prior of the fields): https://doi.org/10.5281/zenodo.23975277

How to cite

When using these files, please cite our preprint:

Mom, S., Otake, Y. and Wu, S. (2026). GeoPFN: subsecond geological field reconstruction with constrained uncertainty maps. EarthArXiv. https://doi.org/10.31223/X5JV68