Global earthquake detection and warning using Android phones
As noted by Allen et al. (2025), even though seismologists have a good understanding of where earthquakes will occur around the world, and earthquake engineers understand how to design structures to prevent collapse and avoid loss of life, earthquakes are still a major cause of loss of life and devastating losses to housing and infrastructure in many parts of the world. Access to the knowledge and necessary resources to construct safe buildings are an impediment in many parts of the world, and even with those resources, it would still take many decades to replace vulnerable buildings. Even when earthquake-resilient buildings are in place, there remain hazards from falling objects inside buildings that can cause injuries and fatalities.
To mitigate this risk, earthquake alert systems using standard seismometers have been deployed in locations including the west coast of Canada, China, Israel, Italy, Japan, Mexico, Romania, Taiwan, Turkiye and the west coast of the United States in recent decades. An alert is made possible by the finite (although quite high) speed of seismic waves travelling through the earth. In locations having dense seismic networks, the location and magnitude of an earthquake can be estimated within a few seconds, allowing a warning to more distant locations that strong ground shaking (stronger than MMI Intensity greater than 5, which is potentially damaging) is expected to begin in a few seconds. Such systems have been used for several decades to reduce the speed of “bullet trains” in Japan; to date there have only been two derailments, which both occurred when the earthquake was so close that there was not enough warning time to sufficiently reduce the speed.
There is already a natural warning system in how seismic waves arrive at a site. The first waves to arrive are P waves, which are compressional waves analogous to sound waves. Their amplitudes are much weaker than those of the following S waves, which are shear waves that do most of the damage in earthquakes, and travel at about 60% of the speed of P waves. The “duck, cover and hold on” (DCHO) rule for immediate response to an earthquake is based on this time delay of a few seconds in the case of strong shaking.
Earthquake early warning systems depend on real-time analysis of waves detected at nearby seismic stations. Smart phones contain motion sensors that can be used in place of standard seismic stations, so they can provide alerts in areas without seismic stations. Allen et al. (2025) describe the outcomes of the Android Earthquake Alerts (AEA) system rolled out to Android phones in 98 countries from 2021 to 2024. This detection capability is deployed as part of Google Play Services core system software, so it is on by default for most Android smartphones. As Android phones represent an estimated 70% of all smartphones globally, this system provides an earthquake detection capability wherever there are people, in both wealthy and less-wealthy nations.
In just three years, the AEA system has expanded the existing warning system, based mainly in selected developed countries, to a global scale, and increased access by an order of magnitude to 2.5 billion people. It has issued alerts for 1,279 events over that period, with low errors in magnitude estimation (with a major exception discussed below) and provided warnings consistent with other detection systems but on a much larger scale.
Types of Alerts
Two types of alert messages are delivered by AEA. When the estimated magnitude is ≥4.5, a “TakeAction” alert is delivered to the region expected to experience MMI ≥5, and a “BeAware” alert is delivered to the region expected to experience MMI 3 or 4. For an event with an estimated magnitude of 5.5 and depth of 20 km, the TakeAction alert would extend to 8 km, and the BeAware alert to 197 km. For a M 6.5, the alert regions extend to 78 and 442 km for TakeAction and BeAware, respectively. The same alert distance contours are currently used for all regions. Alerting performance can be improved in the future by accounting for regional seismic attenuation differences.
The TakeAction alerts were designed to prompt users to take protective actions, but by delivering them only when we have predicted strong shaking, we are only able to provide a few seconds’ warning. The BeAware alerts are less intrusive and more informational and often delivered with a warning of many tens of seconds, due to the earthquake being further away. 28% of recipients who received a TakeAction warning reported following the recommended DCHO action, which is higher than in previous surveys. 84% of users responding report that they will trust the AEA system more next time, and only 3% say less. Trust in an alert is one of the key factors promoting compliance with the recommended actions.
Allen et al. (2025, Figure 4) illustrated typical AEA system behavior and warning times with two example earthquakes, one of which they describe as follows:
“The first is the 17 November 2023 earthquake in the Philippines with a magnitude of 6.7 and a hypocenter 40 km offshore at a depth of 52 km (USGS). It took about 12 s for the P wave to reach the closest cluster of phones, and the first alert was generated 18.3 s after OT (the origin time) with an estimated magnitude of 5.5. The magnitude estimate grew over time to a maximum of 6.5, 28.6 s after OT. The warning time for users receiving a BeAware alert ranged from a few seconds for those closest to the epicenter experiencing the strongest shaking, up to ~90 s for those at 400 km experiencing light shaking. Warning times for the strongest shaking (MMI 7 or 8, very strong to severe) ranged up to about 15 s. For the moderate shaking, which also causes damage (MMI 5 or 6, moderate to strong), warning times ranged from seconds up to a minute. Almost 2.5 million phones were alerted for this event, and more than 100,000 received TakeAction alerts. The TakeAction alerts arrived at most phones a few seconds before the S-wave arrival and ~2 to 8 s before peak shaking of MMI 6, 7, and 8.”
Wikipedia (2023a) reported that a maximum intensity of 8 (severe) was felt and at least 11 people were killed while 730 others were injured, including 450 due to panic. No reports have been found as to whether, and if so how, any people reacted to the alert, but it seems likely that alerts could have reduced injuries due to panic.
The other earthquake analysed by Allen et al. (2025) was the M 5.7 earthquake that struck Jajarkot, Karnali Province, Nepal on 3 November 2023. Wikipedia (2023b) reported that it killed 153 people and injured at least 375. The earthquake was widely felt in western Nepal and northern India. No reports have been found as to whether, and if so how, any people reacted to the alert, but it seems likely that alerts could have reduced injuries.
Unfortunately, AEA was in an early stage of development and testing when two major earthquakes struck south-east Turkiye on 6 February 2023: the M 7.8 Pazarcik and M 7.5 Elbistan earthquakes. More than 55,000 people died and more than 100,000 were injured in these earthquakes, and many were asleep in buildings that collapsed around them when the earthquakes occurred. According to Allen et al. (2025):
“AEA was operational in Türkiye and detected both events. For the first event, the initial magnitude estimate was 4.5 at 7.1 s after OT, and the maximum magnitude estimate was 4.9 at 18.7 s. A total of 512,411 BeAware alerts were issued out to a distance of 64 km. In the second event ~9 hours later, the initial magnitude estimate was 6.1, at 24.4 s after OT, which increased to magnitude 6.3 5.2 s later. A BeAware alert was delivered to 3,944,909 phones in this event, with warning times ranging from a few seconds to more than a minute.”
The eventual magnitudes of these two earthquakes were 7.8 and 7.5 respectively, significantly larger than the final eventual AEA estimates of 4.9 and 6.3. Allen et al. (2025) did not mention the more critical “TakeAction” warnings (which are intended to save lives) that were issued at the time, but according to Clayton et al. (2025), only 469 “TakeAction” warnings were sent out for the first earthquake. The less critical “BeAware” warning is designed to inform users of potential lighter shaking and does not override a smartphone on Do Not Disturb.
Allen et al. (2025) describe in detail the many challenges that are presented in early warning of such large earthquakes, and the results of their improved system as follows:
“Using our evaluation system, we ran the new algorithms on the data collected for the Türkiye earthquakes in simulated real time. The algorithm in production today generated an initial magnitude of 4.6, 6.3 s after OT. The magnitude then climbed to a maximum magnitude of 7.4 over a period of 24 s. This generated TakeAction alerts out to 158 km that would have been received by 10 million Android phones, providing up to a 35-s warning, and BeAware alerts to a distance of 604 km that would have been received by 67 million users, providing a 2.5-min warning before the first S wave. For context, the “heavy” MMI 8 shaking in places such as Antakya, Türkiye, did not occur until more than a minute after OT. Work continues to improve the algorithm for large ruptures, when the finite fault effect becomes important.”
It seems likely that many deaths and injuries could have been avoided if the current system had been operating at the time of these earthquakes.
Implications for Earthquake Early Warning Systems
Allen et al. (2025) conclude that AEA demonstrates that globally distributed smartphones can be used to detect earthquakes and issue warnings at scale with an effectiveness comparable to established national systems. Large earthquakes remain the most important and challenging for all earthquake early warning systems, and the global implementation of AEA supports work to improve detection with rapid, large-scale data collection and feedback to algorithms.
Allen et al. (2025) describe the prospects for further applications of AEA. The AEA detection system could also provide additional information and products that could reduce hazards. Small earthquake detections in regions without seismic networks could help characterize faults and associated hazards beneath urban environments. The dense, local, and real-time ground motion observations could provide rapid post-earthquake maps of shaking intensity similar to ShakeMap, which are critical to emergency response. Aggregated observations of shaking over multiple earthquakes could support improved regional hazard models globally. Delivering these products, with the support and trust of users, will help build community awareness of earthquake hazards and strategies to reduce them.
Near Real Time Access to Shakemaps in Australia
Allen et al. (2025) indicate that alerts for shaking of MMI 3 and above were issued in southern Victoria and southwestern Western Australia during the three year period that they analysed. However, no information has been found as to whether, and if so how, Australians responded to these alerts.
Geoscience Australia does not have its own early earthquake warning system (which is instead provided by AEA), but as soon as an earthquake has occurred, Geoscience Australia (2025) generates a ground motion map (“Shakemap”) in near real time (a few minutes). The development of near real time earthquake Shakemaps is described by Allen et al. (2019). As soon as an earthquake is felt, the website earthquakes.ga.gov.au will take you to a map like that in Figure 1, which shows two earthquakes that occurred on September 22 and 29, 2023 with magnitudes of 2.2 and 2.1 respectively. They were too small to have generated Shakemaps, which are requested by clicking on the circled + sign at the bottom of the menu on the right of Figure 1.
These two earthquakes may be aftershocks of the 28 May 2023 Sunbury earthquake, whose magnitude of 4.0 was large enough to generate the Shakemap shown in Figure 2. The maximum MMI intensity was IV, corresponding to a moderate level of shaking and damage that is expected to be very light. This Shakemap immediately indicates that there may be little damage and no need for field reconnaissance.
References
Bradley BA, Cubrinovski M and Wentz F (2022). “Probabilistic seismic hazard analysis of peak ground acceleration for major regional New Zealand locations”. Bulletin of the New Zealand Society for Earthquake Engineering, 55(1), 15–24. https://doi.org/10.5459/bnzsee.55.1.15-24
Gerstenberger, Matt et al. (2022). New Zealand Seismic Hazard Model 2022 Revision: model, hazard and process overview. GNS Science Report 2022/57, September 2022.
GNS (2022). New Zealand National Seismic Hazard Model 2022 Revision. https://nshm.gns.cri.nz/
NZSEE et al. (2022). Earthquake Design for Uncertainty. Advisory jointly prepared by NZSEE, SESOC and NZGS Revision 1, August 2022: Seismic design of building structures.
McVerry, G., M. Gerstenberger, D. Rhoades and M. Stirling (2012). Spectra and PGA’s for the Assessment and Reconstruction of Christchurch. NZSEE Conference, 2012, Paper No. 115.
Milne, Jonathan (2025). Govt set to quash death sentences on hundreds of quake-prone buildings. https://newsroom.co.nz/2025/07/10/govt-set-to-quash-death-sentences-on-hundreds-of-quake-prone-buildings/?utm_source=Newsroom&utm_campaign=99208bee53-Week+In+Review+13.07.2025&utm_medium=email&utm_term=0_71de5c4b35-99208bee53-576269316&mc_cid=99208bee53&mc_eid=a5fefd9213
Risk Frontiers (2013). Are there 15000 to 25000 earthquake prone buildings in New Zealand? Briefing Note 252.
Risk Frontiers (2022). New Zealand National Seismic Hazard Model 2022 Revision and NZSEE Advisory on Buildings. Briefing Note 471.
Risk Frontiers (2012). QuakeNZ: New Zealand earthquake loss model. https://riskfrontiers.com/solutions/model-data-solutions/quakenz/
Stirling M, McVerry G, Gerstenberger M, Litchfield N, Van Dissen R, Berryman K, Barnes P, Wallace L, Villamor P, Langridge R, Lamarche G, Nodder S, Reyners M, Bradley B, Rhoades D, Smith W, Nicol A, Pettinga J, Clark K and Jacobs K (2012). “National seismic hazard model for New Zealand: 2010 update. Bulletin of the Seismological Society of America, 102(4): 1514–1542. https://doi.org/10.1785/0120110170
