Machine learning systems could detect signs of sudden eruptions hours earlier forcing scientists and authorities to rethink the balance between false alarms and public safety
Scientists are exploring whether artificial intelligence and machine learning systems could provide earlier warnings of sudden volcanic eruptions even if doing so means communities may occasionally face more false alarms.
New research suggests that automated forecasting tools capable of analysing subtle changes in volcanic vibrations could identify signs of an impending eruption hours before it occurs. The findings raise an important question for disaster management: how many false alarms should society accept if earlier warnings could save lives?
The issue is particularly urgent for volcanoes capable of erupting with little warning.
In December 2019, Whakaari, also known as White Island, suddenly erupted in New Zealand, killing 22 people and seriously injuring 25 others. The disaster remains the country’s deadliest volcanic event in recent history.
Similar sudden eruptions have occurred elsewhere. In 2014, Japan’s Mount Ontake erupted with little warning, killing 63 people, many of whom were hiking near the summit.
Researchers say conventional warning systems face major challenges when volcanic activity escalates within minutes or hours. Scientists must interpret complex seismic signals and assess the likelihood of an eruption before authorities can issue warnings.
Machine learning systems could help speed up that process.
The researchers re-examined the performance of a machine learning eruption forecasting system across five volcanoes in New Zealand, Japan, Chile and Russia. Using years of seismic data, the system estimated the probability of an eruption occurring within a rolling 48 hour period.
At Whakaari, the system successfully anticipated four out of five eruptions when tested using previously unseen data.
However, the improved warning capability came with a trade off. The system would have generated warnings for approximately 15 days each year when no eruption ultimately occurred.
While that may appear to be a significant number of false alarms, researchers argue that the risks and costs must be compared with the potential benefits of timely evacuation and preventive action.
Using a cost loss model the study found that precautionary measures based on earlier warnings could reduce preventable losses by between 30% and 90%.
The economic consequences of false alarms would vary significantly depending on the location. For example, closing Mount Ruapehu during a busy ski season could result in substantial financial losses and disruption.
But researchers say those costs must be weighed against the possibility of preventing deaths, severe injuries and long-term trauma.
“In some situations, a warning system that ‘cries wolf’ more often may ultimately be safer and economically rational than one that waits for high confidence,” the study suggests.
The researchers point out that other disaster warning systems already operate with a similar level of uncertainty. Tsunami alerts, for example, often result in warnings even when destructive waves do not materialise, yet communities continue to respond because the potential consequences of ignoring a genuine threat can be catastrophic.
Public trust and effective risk communication will therefore be critical if automated volcano warnings are introduced.
Importantly, researchers do not see artificial intelligence as a replacement for volcanologists. Instead, automated systems could act as an additional layer of protection, providing rapid alerts when volcanic activity suddenly changes while experts investigate the situation and determine the appropriate response.
The technology could help protect not only people living near volcanoes but also tourists, hikers, ski resorts, roads, power infrastructure and other critical facilities.
The research also supports the broader goals of the United Nations’ Early Warnings for All initiative, which aims to ensure universal access to multi hazard early warning systems by 2027.
For volcanoes capable of erupting suddenly, researchers say waiting for complete certainty before issuing an alert can itself create serious risks.
The emerging evidence suggests that accepting a higher number of false alarms may sometimes be a reasonable price to pay if it gives communities the most valuable resource during a disaster time.