Volcanic Shockwaves in Home Barometers: Early Warning From 29,334 Weather Station Records

When a volcano erupts explosively, it sends out Lamb waves: acoustic-gravity waves trapped in the lowest layer of the atmosphere that spread horizontally around the globe at the speed of sound. A study built on 29,334 real observation records from consumer weather stations in Japan shows that ordinary balcony barometers can catch those waves. The setup recovers the temperature-dependent wave speed, finds the direction of an unmonitored crater, estimates blast energy and gives between 2.5 and 15 minutes of warning to cities downstream, with no false alarms during violent storms.

Two benchmark eruptions from 2018 anchor the work: the steam-driven (phreatic) blast at Mt. Kusatsu-Shirane and the magmatic explosion at Mt. Shinmoedake. The results are reported as one academic perspective among many on volcanic infrasound and atmospheric wave dynamics, offered to invite more empirical testing and discussion in the geophysics community.

The sparse-sensor problem

Professional volcano monitoring relies on infrasound microphones and seismographs run by government agencies and by nuclear-test-ban monitoring networks. Those stations sit 100 to 500 kilometres apart. That spacing leaves large blind zones, causes severe spatial aliasing and exposes isolated sensors to wind turbulence.

Consumer IoT barometers offer a way round the gap. Thousands of residential Netatmo stations across Japan can be treated together as one continent-sized acoustic antenna. The study's 29,334 observation records suggest that pressure sensors bolted to apartment balconies can warn of an approaching shockwave and its ash before either reaches downstream urban centres.

The research pipeline

The whole project depends on getting data from home sensors into a form where physics can be applied to it. One overview picture captures the idea: an eruption throws out a two-dimensional cylindrical Lamb wave that races along the ground at 1,134 km/h, thousands of residential stations across urban areas intercept it, and cloud analytics compute blast yield in tons of TNT while counting down a warning of 2.5 to 15 minutes.

Continuous ingestion

The foundation is an automated cloud pipeline that runs without supervision. A serverless Google Apps Script trigger fires every 20 minutes and calls the Netatmo Connect API for barometric pressure, outdoor temperature and relative humidity from stations spread across Japanese cities and mountain foothills. Incoming data is indexed, structured and archived into a central Google Drive repository.

Four stages of work

With data accumulating continuously, the project proceeded in four stages that run from raw archive to verified theory.

1. Live ingestion. The script acts as a background engine, polling every 20 minutes to build an unbroken, long-running meteorological archive.
2. Historical event mapping. Past eruptions are laid against that archive by extracting 24-hour windows running from 12 hours before onset to 12 hours after, for both the Kusatsu-Shirane steam blast and the Shinmoedake eruption. Calm and stormy control days provide baselines.
3. Pre- and post-eruption anomaly study. Atmospheric time series before and after each blast are examined for two things: whether faint hydrothermal degassing anomalies appear before rupture, and how the shockwave crosses urban areas afterwards.
4. Theory and validation in a CLI environment. Unified mathematical models, namely Bolton enthalpy, two-dimensional compressible Lamb wave equations, slowness beamforming and cylindrical energy-flux integration, were supplied to Gemini running inside the Antigravity CLI to test whether warnings and crater diagnostics hold up with zero false alarms.

The physics through four everyday pictures

How can an uncalibrated $150 gadget on a balcony register an explosion hundreds of kilometres away? Four familiar analogies explain it.

Lightning and thunder

A flash arrives at once and thunder follows many seconds later, because light moves at 300,000 km/s and sound at about 340 m/s. The warning system exploits the same gap. An eruption's shockwave crosses the ground at roughly 315 m/s, about 18.9 km per minute. Stations 10 to 30 km from the crater feel the initial pressure pulse within 30 to 90 seconds and send their data to the cloud at the speed of light.

An obvious objection: if each station reports in 20-minute batches, how can lead time of several minutes be real? Two principles answer it.

• Travel time over distance. The wave needs real time to cover geography. The delay grows with distance and cannot be avoided: mountain communities 50 km out wait about 2.6 minutes, towns at 100 km about 5.3, Tokyo (180 km from Kusatsu-Shirane) about 9.5, and the Fukuoka transit corridor (280 km from Shinmoedake) close to 15, at 14.8 minutes. Near-crater sensors reach the cloud in under a minute, so alerts arrive well ahead of the wave.
• Staggered reporting. Home stations do not transmit in lockstep, since their 20-minute cycles are scattered through time. In a dense cluster of 100 stations, a fresh packet reaches the cloud every 12 seconds on average, which turns discrete 20-minute updates into an almost continuous stream. Edge-triggered webhooks can add sub-second bursts when a steep pressure gradient appears.

Ripples in a pond

An unconfined explosion spreads energy in three dimensions, and the signal thins quickly, falling with the square of distance. Lamb waves are held against the surface by gravity and layered air density, so they travel as a flat, two-dimensional ring. Energy cannot escape upward, which makes the pulse decay far more slowly, in proportion to distance and not its square. That waveguide effect keeps the signal intact over hundreds of kilometres, which is why an uncalibrated barometer in downtown Tokyo can register a sharp micro-pulse from a crater 180 km away.

Binaural hearing across 150 kilometres

People can shut their eyes and point toward a sound because the brain compares arrival times at the left and right ears down to microseconds. The system does the same over a whole region. By comparing arrival-time differences across 25 to 88 stations spaced along a 150 km baseline, the delay-and-sum beamformer, working in slowness space, lines up the recorded traces for every azimuth around the full circle. The angle at which the waveforms reinforce one another marks the crater, with sub-degree precision and no prior knowledge of its coordinates.

The speed gate and the chorus test

Local wind is the main instrumentation headache. By Bernoulli's principle, gusts across an open balcony cause sharp local pressure drops, and a single station can log hundreds of false spikes on a stormy day. Two physical hurdles remove them.

• The speed gate. Gusts ride along with a weather front at only 10 to 25 m/s, or 36 to 90 km/h. A genuine volcanic Lamb wave moves at the speed of sound, between 300 and 330 m/s, close to 1,134 km/h. Anything slower than sound is rejected immediately.
• The regional chorus test. A gust is a local, uncorrelated event hitting one house at a time. A volcanic shockwave is a coherent wavefront sweeping dozens of homes across a prefecture in order.

Requiring phase coherence across many stations, with semblance of at least 0.70, removed all 378 gust spikes recorded in severe gales. The false positive rate was 0.0%.

A worked example of the lead-time arithmetic

The numbers behind the warning are easy to check. At 315 m/s, a wave covers 18.9 km each minute, so a town 50 km from the crater has about 2.6 minutes between the first detection and the arrival of the wave, and a city at 280 km has nearly 15. What matters for early warning is that detection happens in the near field, 10 to 30 km out, where the pulse registers within 30 to 90 seconds, and that cloud servers receive it in under a minute. Each additional kilometre of distance between the crater and a downstream city therefore adds roughly three more seconds of usable warning. The 2.5 to 15 minute range quoted for the system is simply this geometry applied from the closest to the most distant communities in the study.

The same geometry explains why dense coverage matters more than sensor quality. A single cheap barometer is noisy, yet dozens of them that agree on timing and direction produce a signal that no single instrument could. The consumer devices are uncalibrated, so absolute pressure values are unreliable, but the shape and arrival time of a sharp pulse survive, and arrival times are all that beamforming and speed inversion require.

Guiding AI with physical logic

Generative AI is strong at spotting patterns in high-dimensional data. Applied blindly to noisy weather series, though, an unconstrained model can latch onto spurious correlations and produce artifacts with no grounding. Turning raw model output into verifiable discovery calls for a hybrid: pattern-mining agents coupled directly to first-principles theory and domain logic.

The guiding idea is simple. Feeding a model not just numbers but also the governing physical laws, thermodynamic equations and geometric constraints gives it a structured lens, raising the odds of finding real patterns and filtering out noise. The same philosophy was applied across other disaster-prediction projects. For earthquakes, lithosphere-atmosphere coupling equations and seismic moment tensors were supplied to uncover pre-seismic radon ionization and enthalpy spikes of 5.96 to 15.47 K hours before rupture. For urban torrential rain, 3D Navier-Stokes momentum and helicity equations were used to capture localized convective initiation 45 minutes ahead. For volcanoes, Lamb wave equations, Bolton steam enthalpy and slowness-beamforming constraints separated shockwave patterns and crater bearings from 29,334 noisy records.

Six processing stages

Driven by that human-guided approach, the pipeline carries a derivation from raw readings to physical quantities, as the following stages show. In outline, residential measurements of pressure, temperature and humidity pass through filtering, speed inversion, beamforming and energy integration, ending in a picture of blast size and ash cooling.

1. Pre-eruption enthalpy tracking. Temperature, humidity and pressure become equivalent potential temperature through Bolton's formulation. Subtracting regional trends exposes faint thermal-steam anomalies, 1.5 to 3.2 K, downwind of the volcano hours before rupture.
2. Zero-phase filtering. Pressure series pass through bidirectional Butterworth bandpass filters from 0.0005 to 0.02 Hz, which strip 24-hour tidal cycles while keeping wave arrival times undistorted.
3. Temperature-dependent travel-time inversion. Instead of a constant speed of sound, the model integrates over local temperature fields, accounting for slower sound over freezing mountain plateaus.
4. Beamforming. Each regional cluster of stations has its traces slid in time and summed for every direction on a 360-degree slowness grid. The direction where the sum is strongest and most coherent, measured as semblance, gives the back-azimuth.
5. Cylindrical energy-flux integration. Pressure variance is integrated over the two-dimensional cylindrical surface and scaled by an atmospheric scale height of about 8.4 km, giving total acoustic energy and its TNT equivalent.
6. Ash-shadow cooling. Dense suburban temperature sensors detect solar shielding under drifting ash clouds, seen as local drops of 4.10 to 6.13 degrees Celsius.

What 29,334 records showed

Two contrasting events make a clean test. A freezing winter steam blast and a mild spring magmatic explosion produce different wave speeds and very different energies, so a method that handles both is hard to dismiss as a fluke. The records came from two major eruptions plus two severe-weather control periods, which supplied the stormy baselines against which false alarms were counted.

The validation covered 29,334 real 20-minute records spanning two major eruptions and two severe-weather control events.

• Temperature-dependent wave speed. At Kusatsu-Shirane the shockwave moved at 304.38 ± 4.73 m/s. Sea-level sound speed is near 340 m/s, but sound slows in cold dense air. On that January morning at -5 degrees Celsius across the Gunma highlands, theory predicts 304.8 m/s. The uncalibrated array recovered the thermal slowdown with 99.8% accuracy (R squared of 0.980). At Shinmoedake in 12-degree spring air the speed was 313.27 m/s, within 98.5% of theory.
• Locating an unknown crater from 150 km. Beamforming across Kyushu stations put the Shinmoedake crater at a bearing of 169.49 degrees against a true 167.72, an error of just 1.78 degrees.
• Steam versus magma yields. Radiated acoustic energy was 7.48 x 10^11 joules (178.8 tons of TNT equivalent) for the steam blast and 4.36 x 10^12 joules (1,041.1 tons) for the magmatic one, so magmatic explosions release roughly six times more atmospheric energy than phreatic ones.
• Storm control. A falsification audit over more than 15,000 control records and the 378 gust spikes gave a 0.0% false positive rate.

Three ways to use it

Turning home weather stations into real-time geophysical intelligence supports three kinds of protection, each of which depends on the lead time that distance and network speed create.

Aviation and urban infrastructure

Aircraft and cities face the same threat on different timescales. Pilots need to be steered clear of a rising plume within seconds, while a city needs minutes to prepare its underground and indoor spaces.

Volcanic ash melts inside jet engines above 1,400 degrees Celsius and turns to liquid glass, which can cause total engine failure. A network that estimates yield and crater position within seconds can alert air traffic control and flight management systems to reroute aircraft around rising plumes. Downwind cities get a countdown of 5 to 15 minutes to seal subway ventilation shafts, switch hospital air handling to internal HEPA filtration, pause outdoor construction and begin slowing high-speed rail.

Volcanic tsunamis

Exploding submarine calderas and collapsing island flanks, as at Krakatoa in 1883 and Hunga Tonga in 2022, generate deadly tsunamis. Deep-ocean tsunamis travel near 200 m/s and slow to 30 to 50 m/s in shallow coastal water, while atmospheric Lamb waves move at 315 m/s and outrun them. Coastal barometers that detect the air wave can give 15 to 45 minutes of evacuation lead time before the water arrives, triggering automatic seawall gate closures and rail halts.

Blind crater localization

This case matters most for volcanoes that nobody is watching.

The worst volcanic hazard is a surprise eruption from a dormant mountain or an unmonitored flank fissure with no government instruments nearby. Two methods address it. Before rupture, escaping steam makes small local anomalies of 1.5 to 3.2 K in equivalent potential temperature, and solving the adjoint advection-diffusion equation traces the plume to its source, locating impending degassing days ahead. After an explosion, independent regional sub-arrays in Kanto, Nagoya and Kyushu compute cross-bearing lines with 0.48 degrees of error, which fix the active crater to within 840 metres within minutes.

The AI co-researcher setup

A human principal investigator stays in charge of vision and approval, while a set of specialized agents covers data ingestion, modelling and simulation, code development, falsification and simulated peer review. Each does a narrow job: one audits plans through mathematical dry runs, another evaluates theoretical feasibility, another writes the code, others audit results, guard data integrity and review the theoretical discussion, and the last two check style compliance and typesetting.

Running the continuum-mechanics equations, ingesting the records, performing falsification checks and preparing a publication-ready LaTeX manuscript were all carried out in the Antigravity CLI, powered by Gemini together with an auxiliary research framework called tanaike-lab. The work followed five phases.

1. Strategic prompting by a human. The principal investigator opened with a mathematically constrained research charter and not an open-ended request, with six firm requirements: build a micro-barometric Lamb wave detector from citizen-science IoT series; use 24-hour windows (plus or minus 12 hours) around the two 2018 eruptions, with storm baselines for falsifiability; apply Bolton's equivalent potential temperature to isolate pre-eruptive heat flux; enforce bidirectional zero-phase filtering to keep exact timestamps; invert propagation velocities using spatially varying temperature; and use slowness beamforming to get azimuths and yields from multi-station arrival differences.
2. Plan audit and mathematical dry runs. Specialized subagents checked the plan for mathematical consistency, validated the cylindrical energy-flux model and wrote the temperature path-integral equation to account for mountain terrain.
3. Pipeline development with assertion hooks. A code-development subagent built isolated Python pipelines, each with automated assertions ensuring that Drive ingestion over OAuth2, filtering and beamforming ran on 100% real data with no synthetic mock arrays.
4. Empirical analysis and agent-to-agent discussion. Auditing subagents debated results across the 29,334 records, verified the 304.38 m/s match, assessed crater back-azimuth accuracy and confirmed full rejection of the 378 gust spikes.
5. Peer review and delivery. A simulated five-axis review panel, a style compliance step and a typesetting linter ran pre-submission audits, checked that every in-text citation matched, compiled the LaTeX and produced colour-accessible figures before the human gave final approval.

Unlike a static agent script, the framework captures lessons as it goes. The loop is deliberately automatic: directives from the human and lessons verified during execution are noted on the fly, written at once into skill documentation, and pushed back into the tool without a manual reinstall. Steering directives and verified lessons are recorded during a run, written straight into a SKILL.md file, then exported and reinstalled into the local CLI plugin directory with an immediate Git remote sync, so the toolkit keeps evolving.

Its stated positioning has three parts. It contrasts unsupervised generation, which suffers hallucination and superficial text, with a human-AI synergy model in which the human keeps strategic direction while subagents speed up formulation, code execution, auditing and review. It aims to remove operational friction such as debugging scripts, interpolating uneven time grids, adjusting figures for colour-universal design and fixing LaTeX errors, so researchers can concentrate on high-level reasoning. And it spans three domains: natural sciences such as geophysical modelling, fluid dynamics, quantum chemistry and materials; generative AI engineering including prompt design, RAG retrieval, multi-agent orchestration, token-budget compression and benchmarks; and the Google ecosystem, covering Apps Script libraries and Workspace automation across Drive, Sheets, Docs, Gmail and Forms with quota management.

What it adds up to

Residential weather gadgets, combined with physics-grounded AI, can become a continent-scale early-warning net: Lamb waves and blast energy captured from crowdsourced barometers, with zero false positives against storm baselines. The conclusions are one of many perspectives in volcanic physics, and continued empirical validation and open scientific discussion remain essential. Beyond geophysics, the pattern of a human setting strategy and physical constraints while AI agents handle derivation, coding, auditing and review looks like a scalable template for climate adaptation, multi-agent engineering and automation of Google APIs.