A friend of ours in Damascus is trying to make his house livable again. The building still stands. For several blocks in every direction, almost nothing does. He has worked at it for months, and the questions he still cannot answer are the ones that matter most. Is the cracked wall he just leaned a shelf against going to hold? Does the main under his street still carry water, or was it cut years ago? Is the wire feeding his block live? He does not really know. He is guessing, carefully, and so is everyone else rebuilding here.
In late 2025 the World Bank priced the rebuilding of Syria at about $216 billion, and said the real figure could reach $345 billion. That is roughly ten times what the whole country produces in a year. Almost every conversation about that number is about where it will come from. To us the harder question is what happens once it arrives. Our friend's problem is the country's problem, repeated millions of times over: you cannot rebuild something well if you cannot first see the state it is in.
Think about a water utility, in Aleppo, in the ruined eastern suburbs of Damascus, or anywhere the front line passed through. It cannot tell you which of its buried pipes still hold pressure, which were cut, which were spliced into with informal connections during years of siege, or, in many areas, where the network even runs anymore. The drawings burned. The control systems were looted. The engineers who kept the rest of it in their heads are now in Gaziantep, in Berlin, or dead. The same blindness covers the buildings (which half-standing wall is still carrying load) and the power grid (which lines are still live). You cannot safely put water, people, or current back into a system whose condition you cannot read. So the real obstacle is not the money. It is that no one can see.
Start with water
Across the Middle East and North Africa, water utilities already lose about 38 percent of the water they produce before it reaches anyone. It leaks out of old pipes or is taken through illegal connections. Engineers call it non-revenue water. Global best practice is around 10 percent, and in parts of Jordan the loss is closer to half. And those are the utilities that still function. Now take away the maps, the meters, the billing records, and the experienced staff, and add ten years of shelling and improvised repairs that no one wrote down. What you are left with is probably the most leak-prone and least observable water network on earth.
The normal answer to this is instrumentation. You cover the network in sensors, build a digital twin of it, and watch it. That is the one thing you cannot do here, and the reason why is the whole technical problem.
Why the tools that already exist do not work here
The leak-detection and structural-monitoring systems sold in Europe and the Gulf are built on four assumptions, and a post-conflict city breaks all four. They assume dense sensor coverage, roughly one device per short length of pipe; here you are lucky to afford one per district. They assume reliable power and a network connection; here both come and go, and a sensor left on a street is stripped for its battery within a week. They assume you already have an accurate map of the network; here the map is the very thing you have lost. And they assume years of recorded, labeled history to learn from; here there is none, and there is no trusted instrument to calibrate the new sensors against.
So this is not a case where the technology exists and only needs to be bought. It is a different problem, and it has not been solved. In plain terms: work out the hidden condition, and often the unknown layout, of a damaged and undocumented network, from very few measurements that are themselves unreliable and unlabeled. Four things make this genuinely hard.
First, it is a badly underdetermined inverse problem. A few pressure readings have to stand in for thousands of unknown pipe conditions, so the math has no single answer unless you constrain it with physics. Hydraulics gives you that constraint. A leak drops pressure in a way that spreads through the network according to known equations, so a pattern of small pressure drops across a few sensors can, in theory, be traced back to the segment causing it. The catch is that this depends on knowing the layout of the network, which you do not.
Second, and because of that, you have to work out the layout and the condition at the same time. Estimating the condition of a network whose map you already have is a standard exercise. Working out the map and the condition together, from the same thin data, is not. That is where the real research difficulty lives, and it is what no existing product tries to do.
Third, there are no labels. You cannot train an ordinary supervised model on leaks that were never recorded. The system has to learn from the gap between what the physics predicts and what the sensors actually report, and then it has to close the loop in the only way the field allows. It points the two or three repair crews you can afford toward the one dig that would tell it the most. Every excavation becomes a piece of training data that did not exist that morning. This is active learning where the budget is measured in hours of digging.
Fourth, the sensors themselves are unreliable, and you have nothing to check them against. In 50-degree heat and constant dust, cheap sensors drift out of calibration within months. With no reference meter to compare to, the calibration has to be done blind, by using the physics and the agreement between sensors to catch when one of them has started to lie.
None of these four is a solved problem. Put together, they are a research program, not a shopping list. And the same four problems (almost no data, unknown layout, no labels, sensors that drift) come back almost unchanged whether you are asking which wall in a damaged building still bears load or which part of a grid is still energized. While the underlying physics engines would require distinct modules for each domain—fluid dynamics for pipes, statics for buildings, Kirchhoff's laws for grids—the data scarcity architecture remains identical. Solve it once, in its hardest form, and you do not have a single product. You have a way to read the physical state of a broken city.
Why it has to be shared across cities
There is one more obstacle, and it points to the solution. A single ruined city does not have enough data to learn these physics-based models well on its own. But corroded pipes, shelled concrete, and overloaded grids do not fail in ways that are special to Aleppo. The same failures repeat in Homs, in Mosul, in Gaza, in Sana'a. The data that is far too thin in any one city is rich when you look across all of them.
So the natural design is a federated one. Each city trains its own model locally and shares only what the model learned, never the raw data. This is not a buzzword added for effect; it is forced by the politics. No government is going to hand over a full map of its water or power network, because that map is a security asset. Federated learning is the only arrangement that lets these cities pool what they learn while keeping what they hold. The shared model gets better with every new deployment, and the next city to come online starts from where the last one stopped.
The engineers who actually know these networks are now spread across a dozen countries. One civil engineer in Homs quietly organized seventy of his displaced colleagues to keep working all through the war. An open, shared system gives people like them something real to contribute to, from wherever they ended up.
How our small team would start
We would not start by building a system for the whole region. We would start by instrumenting one network, in one city, well enough to prove that the loop actually closes.
The first step is recovery: gather whatever partial drawings, old map files, and crew knowledge still exist, and turn them into a rough first guess for the model. The second step is a small amount of cheap, off-the-shelf hardware. Pressure loggers on hydrants and valves, low-cost movement and tilt sensors on damaged buildings, current clamps on the grid feeders. They run on long-range LoRa radio, they are solar-powered with a small battery, and they are plain enough that no one bothers to steal them. Tens of devices, not thousands. The third step is the engine itself: a standard hydraulic solver to provide the physics, a learned model trained on the difference between that physics and the actual readings, and the active-learning loop that turns each crew's next dig into the model's next lesson. The fourth step is to publish the data format, hand the trained model to a second city, and let the network of cities grow from there.
Our team can be small and mostly software. We need someone who understands hydraulics and structures, someone who understands machine learning on graphs and inverse problems, and someone who can keep cheap hardware alive in a hard environment. We also need, without exception, a local partner inside a water directorate or an engineering faculty who can actually get a crew out to the valve.
We should be honest about where we stand. We live in Damascus, in the same city we started this essay with. For a while now we have been designing the system we just described, the soft sensors (algorithms that infer unmeasured states from available physics and sparse data) that estimate what is never measured directly, but so far it lives on paper and in simulation, not in the ground. That is the honest limit of it. The bet, though, is small and clear. The whole thing succeeds or fails on one question: can a model, given almost nothing, tell a crew where to dig and be right often enough to be worth the fuel? Everything after that is scale.
What should exist
The money for Syria, and for the other broken cities of the region, may arrive or it may not. But if it does, it will be poured into systems that no one can currently see, and a lot of it will be wasted. Some of it will do harm, like water pushed back into a main beneath a building that is still occupied. What the region needs and does not have is not more concrete and not more funding. It is a way to know the real state of what is already standing. That is a technical problem. It is hard for reasons we can name. And it is the kind of work that can only be done well by people who know these places from the inside, It would also mean that the next time our friend leans a shelf against a wall, he would not have to guess.