Teaching water models to adapt
Artificial intelligence is rapidly changing how we understand and manage water systems, but a model only earns the right to inform decisions once you can trust what it says. For Cristian Camilo Gomez Cortes, Machine Learning Engineer at DHI, one of the biggest challenges is keeping models aligned with the systems they represent as conditions change. His work explores a deceptively simple question: how do you keep a model reliable when the world keeps changing?
How would you explain what you do to someone outside your field?
I help computer models learn from real-world data and then try to make those models genuinely useful to the people running water systems.
The analogy I like is a flight simulator. Before a pilot touches a real aircraft, they spend hundreds of hours practicing situations that would be too risky or too rare to learn on the job. But that only works if the simulator behaves like the real aircraft. If it is even slightly wrong, the pilot learns the wrong lessons.
My work sits on both sides of that idea.
The first half is making the simulator trustworthy. A model of a wastewater treatment plant or a water network must be calibrated against real measurements and stay accurate as the real system changes. I develop tools that do this automatically, so engineers spend less time tuning models and more time solving problems.
The second half is putting that model to work. Once you have one you can rely on, you can let an algorithm practice operating the plant inside it, trying strategies across thousands of simulated conditions, including difficult ones that occur only rarely in reality, and gradually learning what works. This is an approach known as reinforcement learning, and it is one of the directions we are actively developing. The two halves reinforce each other: a recommendation is only ever as good as the model behind it.
That matters because urban water models are genuinely hard to operate well. Conditions shift constantly (weather, incoming load, temperature), and control strategies are often built on fixed rules that were sensible when written and never revisited. A system that learns can adapt, and in doing so use less energy, hold environmental performance steady and take some of the guesswork out of running a complex process.
So, one half of my work is about models that describe a system accurately. The other is about helping decide what to do with it.
'AI should work alongside human expertise, not replace it.'
Cristian Camilo Gomez Cortes
Machine Learning Engineer, Denmark
Your career has taken you from Colombia to Denmark. How did that journey unfold and what led you to join DHI?
It started with a question about where I could be most impactful.
I studied civil engineering in Colombia, which is where water management and hydroinformatics sit in the Colombian system. Colombia has abundant water, and yet climate change has made something very clear: the challenge is rarely water availability itself; it is how water is managed and distributed. Some regions face drought while others are flooding, often in the same season. At the same time, relatively few engineers were specialising in water compared to construction or geology. Early on, I realised that if I wanted my work to matter, I should go where the need was growing rather than take the traditional path.
As I went deeper, modelling became the centre of my interest, and it became obvious that the sector needed more advanced, data-driven approaches than we were using. I started training in machine learning through international courses, and that gradually took me into digitalisation and AI.
A PhD was the natural next step. I joined BIOMATH at Ghent University in Belgium, which has a strong reputation in water-sector digitalisation research and matched both my interests and my values. There I worked on digitalisation, model calibration and machine learning for water systems, and became increasingly focused on the gap between advanced research and everyday engineering practice.
DHI kept crossing my path throughout that journey. Its modelling technologies played an important role in my research, and over the years I got to know both the tools and the people behind them. Joining felt like a chance to give something back to a community that had already shaped my own development, and to work on real-world challenges at the scale where research, software development and domain expertise come together.
Is there a challenge you've worked on that demonstrates the value of machine learning in urban water management?
The clearest one: building a reliable model is expensive, and keeping it reliable is even more expensive.
Calibration, tuning a model until it reproduces what we measure, takes weeks of specialist time. Then reality moves: loads shift, sensors degrade, regimes change and the model drifts from the live system. Worse, every recalibration starts from scratch, throwing away what the last one learned. So much of my work has gone into automating that and giving it a memory: each recalibration begins from what earlier ones discovered. In benchmarks, a plant model beat its previous best in half the attempts and, running continuously, caught a seasonal shift and recalibrated unattended.
But the harder problem was not the algorithm. Calibration is not unique to any domain; a treatment plant, a drinking water network and a sewer system ask fundamentally the same question. Yet each has its own tools, data and vocabulary, so each has built its own answer, rebuilding the same machinery around a genuinely different core.
The challenge was not to standardise everyone onto one way of working. A domain expert's knowledge is the valuable part, and a shared tool that flattens it is worse than none. What worked was drawing a line: build the common part once and let each field contribute its specifics as knowledge rather than as another copy of everything. Nothing is hard-wired to one domain; a step uses whatever expertise that field has supplied and stands aside where it has none.
We are still drawing that line, and it moves: each new case shows us something we assumed was domain-specific and is not, or the reverse. But the principle held: share the method, exchange the knowledge, leave room for the detail. That means engineers can build on what already exists instead of starting again.
What challenge in the water sector would you most like to see AI help solve?
Helping operators manage increasingly complex systems under uncertainty.
Utilities are squeezed from several directions at once: climate change, urbanisation, aging infrastructure and stricter environmental requirements. They also have more data than ever. But data volume and decision quality are not the same thing, and turning measurements into actions remains genuinely difficult.
I would like to see AI move us from monitoring systems to actively supporting the decisions made about them, recommending operational actions, adapting controllers as conditions change, helping utilities anticipate problems rather than react to them. The learning-based control I mentioned is one step in that direction, but the wider opportunity is much bigger than any single case.
I want to be precise about what that should look like, though. AI should work alongside human expertise, not replace it. Operators possess knowledge about their systems that has never been written down, and no model has access to it. The most impactful solutions will combine that experience with data-driven intelligence, and that is how we get to genuinely more resilient and sustainable water management.
What do you enjoy doing when you are not working with models and data?
Mostly things that take me away from a screen.
I run and I cycle, and which one it is depends more on the season than on any plan. The point is being outside and moving. Denmark makes that easy; there is always somewhere to go, and the right company makes the weather much less of a problem.
I am also learning Danish. When I arrived in Belgium, it took me some time to find the motivation to start learning Dutch, and I remember exactly what that cost me. So, this time I started early, mostly so I could take part in things sooner. Languages are something I enjoy learning when there are people to speak them with.
And from time to time, I act with an amateur theatre group. It is refreshing to interact in ways you never do in ordinary life. It keeps the mind moving, and it turns out to be a surprisingly creative way of expressing what you feel.
Between the three, that is most of what I need: something that gets me outside, something that keeps me discovering and something that lets me express myself.
