There is growing evidence that machine learning will play an ever more important role in the ability for crucial infrastructure to recover post a major natural catastrophe.

Researchers are increasingly turning to artificial intelligence (AI) to help governments and emergency managers restore critical infrastructure faster and more effectively, and the frequency and severity of natural disasters increases.

A new review has examined how machine learning (ML) methods are being applied to support infrastructure recovery following natural hazards. The study was conducted by researchers from University College London and Tsinghua University and systematically analysed 57 academic studies to understand the current state of the field and its future potential.

They added infrastructure systems, such as power grids, water supply networks, telecommunications, and transportation, form the backbone of modern society. When disasters strike, damage to these systems can disrupt essential services for weeks or months. According to the researchers, faster and more efficient recovery is therefore a key component of infrastructure resilience.

The review found that ML techniques are increasingly being used in three main ways.

First, ML can characterise recovery processes by analysing data sources such as satellite imagery or social media posts to understand how infrastructure systems recover over time after a disaster.

Second, ML models can predict recovery outcomes, such as estimating how long it will take to repair damaged infrastructure or forecasting how traffic flows will change after hurricanes or earthquakes.

Third, and most importantly, ML can optimise recovery strategies. Reinforcement learning algorithms – one of the most widely used approaches identified in the review – can simulate different repair strategies and identify the most effective sequence of actions for restoring infrastructure systems.

The researchers found that reinforcement learning accounted for nearly half of the studies reviewed, reflecting its strong potential for complex decision-making problems. Power systems and transportation networks were the most commonly studied infrastructure sectors.

“Recovery planning often involves multiple interacting systems and stakeholders,” said lead author Zaishang Li. “Machine learning can help decision-makers evaluate many possible strategies and identify solutions that restore services more quickly and efficiently.”

Despite promising progress, the study also highlights several challenges that limit the widespread use of machine learning in disaster recovery.

One major issue is data scarcity. High-quality datasets describing real infrastructure damage and recovery processes are rare, partly because disasters are infrequent and partly because infrastructure data are often sensitive or difficult to collect. As a result, many existing studies rely on simulated data rather than real-world observations.

Another challenge is ensuring that ML models can generalise across different infrastructure systems and disaster scenarios. Infrastructure networks vary widely in design, and recovery decisions also depend on human factors such as resource availability and policy priorities.

To address these challenges, the researchers recommend developing shared open datasets, combining multiple data sources, and applying explainable AI techniques to better understand the factors that drive recovery outcomes.

Looking ahead, the team believes that advanced approaches such as multi-agent reinforcement learning could enable coordinated recovery strategies across interdependent infrastructure systems, for example, restoring electricity, water, and transportation networks simultaneously.

“Ultimately, our goal is to support smarter and more resilient infrastructure systems,” Li said. “By integrating machine learning with disaster recovery planning, cities may be able to restore essential services faster and reduce the societal impacts of future natural hazards.”