Applied AI research on local hardware. Geophysics, materials science, and structural reasoning — from Austin, TX.
ScrollMulti-modal pattern detection on tremor and strainmeter data identifies tectonic precursors weeks before conventional methods.
Seismology → 02Atmospheric reanalysis fields contain detectable structural changes days before hurricane landfall. Pilot results from 8 CONUS events.
Atmospheric Science → 03Seismic tremor, ground deformation, gas emissions, and thermal anomalies — individually noisy, potentially powerful in combination.
Volcanology — In Progress → 04A leaf's branching pattern encodes its mechanical properties. Spectral fingerprints predict traits that otherwise require lab measurement.
Materials & Remote Sensing →Nine events analyzed from 2010–2023 using tremor catalogs (PNSN) and strainmeter data (UNAVCO/EarthScope). Each row shows one event's precursor trajectory metrics.
Slow-Slip Event Trajectories
| Event | Lead Time | Approach Corr | Max Distance | Drift Rate / Day | Signal |
|---|---|---|---|---|---|
| Cascadia 2010 Summer | 17 days | +0.750 | 1.256 | 0.044 | Strong |
| Cascadia 2011 Summer | 17 days | +0.641 | 1.284 | 0.041 | Strong |
| Cascadia 2012 Summer | 13 days | +0.569 | 1.190 | 0.035 | Strong |
| Cascadia 2013 Fall | 17 days | −0.336 | 1.482 | −0.028 | Weak |
| Cascadia 2014 Fall | 16 days | +0.848 | 1.011 | 0.037 | Strong |
| Cascadia 2015–16 Winter | 17 days | +0.838 | 3.016 | 0.161 | Strong |
| Cascadia 2017 Winter | 17 days | +0.637 | 1.191 | 0.044 | Strong |
| Cascadia 2022 Winter | 16 days | +0.572 | 1.623 | 0.054 | Strong |
| Cascadia 2023 Summer | 13 days | −0.611 | 1.839 | −0.030 | Weak |
Approach correlation measures monotonic signal increase toward event onset. Positive = precursor detected. Data: PNSN tremor catalogs, UNAVCO strainmeter network. 14-day sliding windows.
Eight CONUS hurricane landfalls (2000–2024) analyzed against matched controls. ERA5 reanalysis, 500 hPa geopotential height. Each event paired with a same-location, same-calendar-date control year with no storm.
Hurricane vs. Control Signal Separation
| Storm | Year | Cat | Min Pressure | Event Corr | Control Corr | Precursor |
|---|---|---|---|---|---|---|
| Lili | 2002 | 1 | 963 hPa | −0.001 | +0.457 | No |
| Emily | 2005 | 3 | 944 hPa | +0.522 | −0.329 | Yes |
| Ike | 2008 | 2 | 950 hPa | +0.476 | −0.027 | Yes |
| Isaac | 2012 | 1 | 967 hPa | +0.101 | −0.297 | Yes |
| Matthew | 2016 | 1 | 967 hPa | −0.093 | −0.223 | No |
| Dorian | 2019 | 2 | 956 hPa | +0.641 | −0.280 | Yes |
| Ida | 2021 | 4 | 931 hPa | +0.002 | +0.110 | No |
| Helene | 2024 | 4 | 939 hPa | +0.518 | −0.163 | Yes |
Approach correlation: Pearson r between day-offset and distance-from-baseline over 30 pre-landfall days. Positive event correlation + negative control = precursor signal. Data: ERA5 reanalysis (ECMWF), HURDAT2 (NHC). Pilot study — 8 events.
Public data streams we plan to analyze for multi-modal eruption precursor detection. All freely available, spanning decades of observations at well-monitored volcanoes.
Target Data Streams for Volcanic Analysis
| Modality | Source | Instrument / Satellite | Cadence | Coverage |
|---|---|---|---|---|
| Seismic Tremor | USGS / IRIS | Broadband seismometers | Continuous | Kilauea, St. Helens, Rainier, +160 monitored |
| Ground Deformation | ESA Copernicus | Sentinel-1 SAR | 6–12 days | Global, cm-precision InSAR |
| SO₂ Emissions | NASA / ESA | OMI, TROPOMI | Daily | Global column SO₂ |
| Thermal Anomaly | NASA | MODIS, VIIRS | 2–4x daily | Global, sub-km resolution |
| GPS Displacement | EarthScope (UNAVCO) | Continuous GPS | Continuous | Major US volcanoes |
Research direction — analysis not yet complete. Starting with Kilauea (2018 lower East Rift Zone eruption) as the best-monitored first case.
We believe AI should run on hardware you own, cost nothing per query, and get smarter every day. Our research runs on Apple Silicon in Austin. No cloud. No API calls. No per-token fees.
Working on geophysics, materials science, or structural reasoning? We'd like to hear from you.