30-Second Overview: Following the 1999 Chi-Chi earthquake, Taiwan gradually integrated earthquake observation, real-time communication, and campus disaster prevention. Today’s Earthquake Early Warning (EEW) is not prediction; it utilizes the arriving P-waves after an earthquake has already occurred to issue alerts before the more destructive S-waves arrive. The farther you are from the epicenter, the more seconds you may gain; the closer you are, the less time you may have. The 2024 Hualien earthquake demonstrated that the true difficulty lies not just in speed, but in the ability to correctly estimate intensity when the rupture direction changes.

Wu Yi-min’s Low-Cost Sensors: Starting from a Question

At 1:47 AM on September 21, 1999, the Chi-Chi earthquake pushed Taiwan’s earthquake disaster prevention issues from textbooks to the doorsteps of every family. This magnitude 7.6 earthquake caused over 2,400 deaths and more than 11,000 injuries. Subsequent research and engineering reforms turned the question of "whether we can do one more thing before strong shaking arrives" into a national-level issue. ¹

Figure: 1999 Chi-Chi earthquake intensity map. Public domain; Wikimedia Commons file page. ²

Figure: 1999 Chi-Chi earthquake epicenter location map. Public domain; Wikimedia Commons file page. ³

Professor Wu Yi-min’s research team at the Department of Geological Sciences, National Taiwan University, later developed low-cost P-wave alert devices. Instead of waiting for the entire earthquake to end, the system reads the earliest arriving P-waves to estimate the potential magnitude of subsequent shaking. The key to this idea is not having instruments guess "where an earthquake will happen tomorrow," but transforming an already occurred earthquake into information that is still usable. ⁴

📝 Curator’s Note: The most counter-intuitive aspect of Taiwan’s earthquake early warning is that the less it resembles "prophecy," the closer it gets to technology that can truly save lives.

Starting in 1991, Taiwan gradually established free-field strong-motion stations, accumulating strong-motion data. The Chi-Chi earthquake left behind extensive near-fault records, allowing researchers to place seismic waves, ground response, and building reactions on the same map. A review from National Dong Hwa University points out that these observation data later supported earthquake early warning, seismic-resistant design, and earthquake engineering research. ⁵

P-Alert, which began operation in 2013, formed a high-density network using low-cost MEMS sensors. The Central Research Institute’s Earthquake Science Project page currently lists 786 stations. They do not only issue local alerts but also provide real-time earthquake information and waveform data, allowing researchers to quickly produce intensity maps. ⁶

Here, "low-cost" is not a synonym for cheapness; it is a strategy of density. If every station required expensive instruments, the network would be difficult to lay densely enough. If sensor costs drop, more stations can be placed in schools, research institutions, and different regions, allowing the system to see earlier where the earthquake starts and where it goes.

This density also changes the utility of post-earthquake information. P-Alert does not only answer "where it shook first"; it can piece together real-time signals from various stations into an intensity map, allowing rescue units to see the potentially affected direction first. Research on the 2025 Jiaxi Dapu earthquake shows that P-Alert produced detailed intensity maps 36 seconds after the earthquake. Local sensors within 30 kilometers of the epicenter may also provide notifications earlier than regional alerts. ⁷

Figure: Photo of the Chelungpu-fault crossing a Taiwanese campus runway after the 1999 earthquake. Author Bob Yeats; Source Oregon State University; CC BY-SA 2.0 License; Wikimedia Commons file page. ⁸

Early Warning Is Not Prediction: P-Waves Arrive First, S-Waves Are More Destructive

After an earthquake occurs, P-waves arrive first. They are faster and usually have smaller amplitudes. The S-waves and surface waves that follow often bring stronger shaking. Earthquake Early Warning utilizes this physical gap: stations receive P-waves first, the system quickly calculates location and magnitude, and then notifies potentially affected areas before S-waves arrive. ⁹

The Central Meteorological Bureau (CWB) clearly separates this technology from earthquake prediction. Early warning is rapid notification "after an earthquake has occurred." Prediction is attempting to estimate time, location, and magnitude before an earthquake occurs, which remains a task that cannot be accurately completed in the short term. ⁹

For Taiwan, these few seconds are not abstract scientific units. The CWB’s explanation states that for medium-to-large felt earthquakes within 10 kilometers of the coast of the Taiwan main island, the average time to complete preliminary information and estimated intensity is about 20 to 30 seconds after occurrence. For areas more than 100 kilometers from the epicenter, one may gain more than 10 seconds of early warning time. ⁹

Distance thus determines both the value and the limitations of early warning. In places very close to the epicenter, S-waves may arrive before the system completes its judgment, creating an early warning blind zone. In places further away, although the message arrives a bit later, it may still play a role in train deceleration, factory line stops, gas shut-offs, or personnel taking cover. ⁹

📝 Curator’s Note: The unit of early warning is not just "accuracy," but "what can this ten seconds allow whom to do first."

On-site early warning reads P-waves directly near the user to estimate local subsequent shaking. Regional early warning integrates data from multiple stations near the epicenter, locates and estimates magnitude, and then calculates intensity for other regions. The two can form a hybrid system, prioritizing the result that reaches the release standard faster. ⁵

The CWB’s mobile earthquake early warning service also clearly states that notifications are sent about 10 to 15 seconds after the earthquake occurs, and only uses data from some stations near the epicenter to gain speed; therefore, estimated intensity may differ from actual feeling. ¹⁰

This explains why earthquake alerts cannot rely on a single mobile app. The value of alerts comes from the connection of stations, algorithms, communication, receivers, and response procedures. Without any one segment, the rest may just be a sound on a screen.

The Hualien Earthquake Made "Speed" Encounter Directional Problems

The February 6, 2018, Hualien earthquake became an important case for testing P-Alert. The research team analyzed the performance of the low-cost strong-motion network during this magnitude 6.4 earthquake, demonstrating how high-density stations could quickly produce intensity maps and on-site alert information after the earthquake. ¹¹ Academic reviews point out that during the 2018 Hualien earthquake, after the regional system issued an alert, the P-Alert on-site network still provided about 2 to 8 seconds of early warning in some blind zones. ¹¹

On April 3, 2024, a magnitude 7.4 earthquake occurred offshore of Hualien. Subsequent research indicated that the CWB system estimated a magnitude of 6.8 within 15 seconds. Due to the underestimation, the Taipei metropolitan area did not receive alerts below the original threshold, while the actual intensity felt in Taipei reached 5 (Weak). ¹²

This is not a story of "complete system failure," but rather a more difficult question to answer: when earthquake rupture extends rapidly in a certain direction, if the system only sees partial stations early on, station distribution and data direction may cause magnitude to be overestimated or underestimated. Research found that using only northern stations leads to overestimation, while using only southern stations may lead to underestimation. How stations are distributed directly affects early warning judgment. ¹²

The 2024 Hualien earthquake also reminded people that magnitude does not alone determine disaster damage. Research comparing the 2018 and 2024 Hualien earthquakes pointed out that the 2018 earthquake was shallower, and the Milun fault rupture crossed Hualien City, causing very strong local ground motion. The 2024 earthquake had a larger magnitude but was deeper and lacked similar surface rupture, so the damage patterns of the two earthquakes were different. ¹³

📝 Curator’s Note: A larger number does not necessarily correspond to a larger city wound. What truly determines how people are injured are depth, direction, site conditions, and buildings.

Research after the Hualien earthquake also indicates that real-time early warning and earthquake data sharing networks help facilitate evacuation and response, but they are only part of resilience. Building seismic resistance, school reinforcement, local government decision-making processes, civil organizations, and evacuation experience equally determine whether alerts can turn into action. ¹³

From Stations to Factories: Who Gets the Ten Seconds First?

A High School Freshman Turns Early Warning Blind Zones into an App

When Lin Rui was a high school freshman at Fudan High School in Taoyuan, he began to think, "Is there a way to know faster that an earthquake is coming?" He and Guo Chen-yu did not start from science fair topics or teacher-assigned assignments; instead, they bought materials, assembled accelerometer sensors themselves, and transmitted the measured shaking back to a server. This box, smaller than a power bank, costs no more than NT$800 per unit. Later, the two expanded the stations, forming a cross-school 15-person ExpTech team that created the DPIP Disaster Weather and Earthquake Early Warning App. ^{14 15}

When reported by CNA in 2024, the team had already installed over 150 self-made seismographs in schools and people’s homes. DPIP transmits station signals back to the server, then estimates the arrival time and shaking magnitude of seismic waves based on the user’s location. It does not handle "predicting" earthquakes, but rather turns the time gap before national-level regional alerts complete their calculations into a civilian observation layer that can notify users first. ¹⁶

📝 Curator’s Note: Lin Rui’s story is not "high school students defeating the Meteorological Bureau," but rather breaking down disaster prevention technology into a problem students can start with: build a sensor first, let data come back, and then think about whether it can provide an additional decision entry before the official alert reaches those closest to the epicenter.

DPIP later handled more than just earthquake early warnings. ExpTech’s public repository describes it as a disaster prevention information integration platform, integrating Earthquake Early Warning, real-time intensity, earthquake reports, weather, and various disaster alerts. Its TREM-Net observation network consists of two subsystems: strong-motion observation and micro-seismic observation. This shift is important: an app started in high school ultimately became not just an alert sound on a phone, but placing observation, data services, and user interfaces into the same public disaster prevention problem. ¹⁷

DPIP also makes the question "Can civilian apps replace official alerts?" an incorrect one. Lin Rui and the team still need the CWB’s data, measurement capabilities, and cooperative interfaces. Civilian stations supplement density, on-site signals, and user-end speed. When both know their limitations, cooperation will not package a few seconds into "certain safety." ^{16 14 17}

The application scenarios listed by the CWB are specific: high-speed transportation can decelerate, gas pipelines or life-support systems can be shut off, factory production lines can stop, and computer hard drives can pause reading/writing. These functions do not require people to understand the entire seismology first, but connect alerts to already designed automatic controls. ⁹

Campuses are another application. In the early promotion of Earthquake Early Warning, the system first used national primary and secondary schools as test and promotion targets, because schools have clear response responsibilities and opportunities to practice "what to do when the alert sounds" during normal times. ⁹

A review from National Dong Hwa University mentions that Taiwan has gradually developed on-site, regional, and hybrid early warning, extending messages to application imaginations such as high-speed rail, Taiwan Railways, elevators, gas, electricity, and escape guidance. The question has moved from "whether there is an alert" to "which equipment is connected and which people know the next step." ⁵

More importantly, early warning data can help prioritize when shaking has not completely stopped. Intensity maps do not directly tell rescuers which building will definitely collapse, but they can place station observation, ground conditions, and rupture direction into the same space, marking areas that need verification first. This makes P-Alert’s value not just in the few seconds before the alert, but in the first information map after the alert. ^{6 7}

For the general public, the most important function of alerts is often smaller: stop running, stay away from glass, take cover locally, and move only after the main shaking passes. If all expectations are placed on how early the alert comes, one might ignore that building structures, furniture securing, and normal-time drills are the protection on a longer time scale.

The System Must Clearly State What It Cannot Guarantee Every Time

Earthquake Early Warning has several structural limitations. Areas near the epicenter may fall into the early warning blind zone. If two earthquakes occur in very short succession, the system may have difficulty distinguishing waveforms. Station background noise, communication quality, or human interference may also cause false alarms. The CWB explicitly writes these limitations into technical explanations, rather than packaging the alert as a foolproof guarantee. ⁹

Research on the 2024 Hualien earthquake brought another limitation into the open: the system must make decisions when data is still incomplete, yet must face the uncertainty of large earthquake rupture direction, station spatial distribution, and magnitude estimation. More stable parameters, more comprehensive station configurations, and second-report information are all directions for continued research improvement. ¹²

Therefore, "not receiving an alert" cannot be directly equated with "no earthquake risk," and "receiving an alert" does not equal "knowing where disaster will occur." Alerts are an entry point for risk communication; behind them, building design, community organization, traffic control, medical care, and rescue must still take over.

📝 Curator’s Note: Mature disaster prevention technology does not promise that errors will not occur, but rather marks out where errors might occur first, letting people know where the next layer of protection lies.

What Taiwan Bought Was Not the Future, But Choices

From Wu Yi-min’s team’s low-cost P-wave sensors, to P-Alert’s operation starting in 2013, to the CWB’s regional alerts, what Taiwan has accumulated is not an isolated invention, but an experience of connecting earthquake science to public institutions. ^{4 6 9 7}

This experience also carries Taiwan’s own geographical conditions: the island is not large, earthquakes are frequent, and population and industry are highly concentrated; station density and early warning speed must be required simultaneously. For a society that cannot know in advance when an earthquake will occur, the most pragmatic progress is not finding a magical prophecy, but making every alert closer to reality, making every building more able to withstand shaking, and letting everyone know what those few seconds can do.

When the next alert sounds, the system may only buy a few seconds for certain regions. But if those seconds are already connected to decisions for trains, factories, schools, and families, they are no longer just a countdown. They are Taiwan’s way of choosing to prepare early under the condition that earthquakes cannot be predicted.

References