Infrastructure stress testing with synthetic data: preparing critical assets for extreme weather

Most of the infrastructure we depend on was designed for a climate that does not exist anymore. Bridges, pipelines, power grids and transport corridors were engineered  for the weather of the past, made to withstand the worst conditions recorded in the few decades before they were built. Infrastructure is generally built to the climate norms of the twenty to forty years before it goes up, and not for conditions outside that range. For a long time that was a safe and logical way to design. But it is now slowly becoming a liability, because the range itself is moving.

Two things are happening at the same time, and they reinforce each other. The infrastructure is getting older, and the climate it has to cope with is getting less familiar.

When ageing infrastructure and a changing climate collide

Ageing is the first challenge, and of course it is nothing new. Much of the world’s core infrastructure was laid down decades ago and has been carrying more than it was meant to ever since, with maintenance and replacement often put off for later. A good deal of it would be due for renewal whatever the weather did.

The second shift is that the weather is no longer a stable backdrop. Summers run hotter, rain tends to arrive in heavier bursts, and the swings between dry and wet are wider than they used to be. The European Environment Agency has made the point that the continent’s energy, water and transport systems are closely interconnected, so the way they respond to these conditions is linked rather than isolated. Brazil’s grid shows a similar problem: built around large hydropower across the twentieth century, the network now looks very different, far more reliant on variable renewable generation than the backbone it was designed around. In both cases the underlying situation is the same. Infrastructure built to deal with specific conditions now has to perform under different circumstances.

Where climate stress on infrastructure actually comes from

The effects of this are usually undramatic. Heat softens asphalt and makes steel rails expand and sometimes buckle. Higher temperatures lift electricity demand for cooling just as they make the network carrying that power work harder. Heavy rain brings drainage that was sized for gentler storms to its limits. On their own, most of these are familiar and manageable.

The real problem arises when they combine. The failures that matter most are rarely a single, well-understood stress. They are compound: a dry spell that bakes the ground hard just before a heavy storm, so the rain runs off instead of soaking in; a heatwave that pushes demand to a peak while supply is already stretched; a small fault that trips a protective system and spreads through a connected network before anyone can step in. These combinations are difficult to plan for precisely because they are unusual. There is little history to study, since the particular mix has often not happened before, and a more variable climate keeps producing new mixes faster than experience can build up.

Why monitoring alone cannot predict infrastructure failure

The instinct has been to watch more closely. Sensors on structures, inspections of pipelines, live readings across the grid. Monitoring is genuinely useful and there should be more of it, but it does have a limit. It describes the present and the recent past. It can tell you that a component is warming or a slope is shifting, but it can’t tell how an asset will behave the first time it is confronted by a set of pressures that it hasn’t experienced yet. And by the time a problem is confirmed on the monitoring system, the moment to act has often already passed.

So, in order to address these new issues, we need to design systems that can identify the weak points before they are under pressure. That means knowing in advance which assets are closest to their limits and how a small, local fault might behave and spread once these limits are reached.

Using synthetic data to simulate extreme-weather scenarios 

In order to do this, assets need to put through conditions they haven’t faced yet, so we can understand how they respond. In practice, that is simulation. However, there is an obstacle: data. To simulate an event in a way that means anything, you need data that describes it, and for the events that matter most, that data does not exist, because the event has not happened.

A way to overcome this obstacle is creating the training data synthetically. Artificially created data is generated at high resolution, including the detailed information a model needs to learn from while representing conditions that were never actually recorded. The data is grounded in real physical behaviour rather than simply invented, which is what allows it to stand in for events that have not yet occurred. With this method, there is no longer any need to wait for a rare combination of events to occur before it can be studied, as any combination can be generated, in as many variations as needed. This, then, can be used to test how a landscape, a network or a single structure would respond under the generated circumstances. An extensive drought that precedes a flood, intense heat layered onto peak demand, or a repeated strain on an already tired structure: these could all become scenarios that are examined ahead of time rather than explained afterwards. A model that is shown a wide and deliberately varied range of scenarios learns most from the edge cases, and those edge cases are precisely where infrastructure tends to give way.

From reactive monitoring to predictive planning

The value of this is practical and fairly specific. No operator can reinforce everything, so what matters is knowing where to look first. Simulating how a grid behaves as heat drives demand shows which parts of the network need strengthening. Modelling how a corridor floods under heavier rain shows which crossings are most exposed. Testing how a structure responds to repeated stress turns maintenance from a fixed calendar into a reading of genuine risk. In each case the work moves from reacting to what has already happened towards anticipating what is likely to.

This is the shift critical infrastructure is making, from reactive monitoring to predictive planning. Monitoring keeps you informed about the present. Simulation lets you rehearse the future. Nothing, of course, replaces the engineers, inspectors and sensors already doing the work. But simulation can direct them towards the places that deserve attention, and therefore help save resources where they are most needed.

The conditions our infrastructure faces will keep shifting, and what counts as a normal year will keep changing with them. That is why the historical record, on its own, is no longer enough to plan from. Keeping these assets safe and operational means being able to see, before the fact, how they will hold up in conditions they have not yet met. 

At Another Earth, we generate high-resolution synthetic Earth observation data, designed to train and test the AI models used for environmental, infrastructure and climate work, so that organisations can move from reactive monitoring to predictive planning. It is how we are building the predictive data layer for the physical world.