Would training on noise from multiple detectors simultaneously improve transfer, or does each site's geology make that futile?

This is a great question about whether geological site-specific noise can be overcome by training machine learning models on data from multiple detectors simultaneously.

The short answer: it depends on what you're trying to do, and the evidence is mixed.

The case for multi-detector training

Multi-detector approaches *do* work when you're looking for correlated signals across sites [AG-2025.04-450]. By comparing how a signal appears simultaneously at different locations, you can reject noise that's purely local to one site. This is because a true gravitational wave arrives at all detectors nearly in phase (with tiny time delays), whereas site-specific noise doesn't. The Jensen-Shannon divergence method improves sensitivity by 10–20% using this cross-detector consistency [AG-2025.04-450].

The geology problem is real

However, your concern about geology is well-founded. Seismic correlations—the coherent ground motion that couples into detectors—actually *persist across multiple sites* in the 0.01–40 Hz band, even with horizontal separations of several kilometers [AG-2024.02-434]. This means the "interference" you're trying to filter isn't purely local. Newtonian noise from seismic density fluctuations affects all nearby detectors in correlated ways, which could make it harder to disentangle signal from site-specific artifacts when training jointly.

That said, the correlations aren't perfect everywhere—researchers observed significant coherence for "at least 50% of the time" at candidate ET sites [AG-2024.02-434], implying periods when sites do behave independently.

What works better: sensor fusion, not just multi-detector training

Rather than pooling LIGO + Virgo + KAGRA detector data directly, the literature suggests a more nuanced approach: **fusion of *different sensor types* at the same site or local network**. Combining seismometers with distributed acoustic sensing (DAS) strain sensors improves Newtonian noise cancellation because the two methods respond differently to seismic waves [AG-2025.12-345]. Similarly, machine learning on mixed ground and platform sensors *within* a single detector (LIGO) achieves order-of-magnitude improvements in residual platform motion [AG-2025.11-542].

Bottom line

Multi-detector training helps you identify *true* signals (which are global) versus *noise* (which is mostly local). But geology does limit transfer: if you train a model on LIGO Hanford's seismic environment and apply it to LIGO Livingston, you'll lose performance because the soil and topography are different. The provided preprints don't directly address transfer learning across geologically distinct sites, so I can't give you a precise answer on how much performance degrades.