The first hypothesis was straightforward, at least on paper. Could a machine learning model trained on meteorological data predict a heatwave before it arrived — not in a city with dense sensor coverage, but in terrain that most weather models treat as a gap in their calibration grids? This was the question KG Labs brought to the collaboration with the Vector Institute in late 2024. It seemed like a reasonable place to start.

The framing was institutional from the outset. KG Labs led the research design and analytical framework. The Internet Society Kyrgyz Chapter — which I also cofounded — contributed sensor network data and field datasets from mountain monitoring stations across the country. The Vector Institute provided the modeling architecture and the research team, including engineers Mojtaba Valipour and Yusuf Aydogdu. BE Node — a Turkish civic technology organization, represented by Ceren Zeytonoglu Atici — co-led the project and supported the collaboration from the beginning. Each institution had a distinct role. The overlap was me.

The gap that made the question worth asking

The region in question has a name that is not widely used outside glaciology and hydrology: the Pan-Third Pole. It refers to the interconnected high-mountain system that extends from the Hindu Kush and Karakoram through the Pamirs and into the Tien Shan — covering Kyrgyzstan, Tajikistan, and the southern reaches of Kazakhstan, among other territories. The Third Pole designation comes from the fact that this system holds the largest concentration of glacial ice outside the Arctic and Antarctic. It is the source of river systems that provide water to a significant share of Central Asia’s population. It is also among the least-served regions on earth when it comes to weather forecasting.

The problem is structural, not incidental. Virtually all AI-based and numerical weather prediction models — whether from research institutions or operational forecasting centers — are calibrated for flat terrain. They perform well over plains, coastal regions, and gently rolling landscapes. Once they encounter mountainous terrain, their prediction accuracy drops sharply. This is not a minor limitation. Even the most advanced publicly available AI weather benchmarks, including Google’s WeatherBench 2, document this degradation in mountain zones. The physics of how air, moisture, and temperature interact across rapid elevation changes introduces nonlinearities that standard model architectures struggle to capture at useful resolution.

The second problem is compute. Traditional numerical weather prediction — the kind that powers operational forecasting at institutions like the European Centre for Medium-Range Weather Forecasts — requires high-performance computing infrastructure with advanced GPUs running continuously. This infrastructure exists in Europe and North America. It does not exist in most countries of the Global South, including Kyrgyzstan. The result is a double gap: the models available to the region were not built for its terrain, and the countries most exposed to mountain weather extremes have the least capacity to run the models that might help.

The Vector Institute team developed two architectures to test whether a more efficient AI approach could close part of that gap. GEOFORMER is a lightweight spatiotemporal transformer — designed to learn local patterns in mountain weather data without the computational overhead of large numerical models. The second is a Graph Neural Network encoding spatial relationships at a four-kilometer resolution, letting terrain structure become part of what the model learns rather than something it averages away. Heatwaves have more than doubled in frequency since the 1980s and account for roughly 489,000 deaths per year globally. Even a few additional hours of accurate local warning changes what communities and health systems can do with that information.

The researchers — Yalda Mohsenzadeh, Elham Bagheri, and Joud El-Shawa at Western University and the Vector Institute — were explicit that both architectures remain open questions. The results in the heatwave context are promising. Whether they transfer to the specific conditions of the Tien Shan and Pamirs is the experiment still running.

Heatwaves follow a seasonal logic — they are anticipated, even when they arrive ahead of schedule. Glacial lake outburst floods do not. When one fails, it fails fast.

A shift in the question

Midway through the research, the framing changed. The shift came from the terrain itself, in a sense. Heatwaves are dangerous, but they have structure — a seasonal rhythm, a duration, an arrival that forecasters can anticipate even when the models are imprecise. Glacial lake outburst floods are different. In the Pamirs and Tien Shan, hundreds of glacial lakes sit behind moraine dams and ice walls that thin a little more each season. When one fails, the failure is rapid. A flood may arrive downstream within hours of a trigger event that was invisible to anyone not watching the right sensor at the right moment.

The Ministry of Emergency Situations in Kyrgyzstan knows this pattern from incident records. So do the communities in the Naryn valley and around Issyk-Kul who have maintained their own informal early warning systems — reading water color, listening for sounds from the valley above, watching animal movement — long before any sensor network existed. The question that emerged from the heatwave research was whether a multi-variable AI model could learn to read the same signals, earlier and at greater range.

This is the direction the current team is pursuing. Mojtaba Valipour and Yusuf Aydogdu — engineers at the Vector Institute — are now leading the Third Pole modeling work, with Kyrgyzstan’s mountains as the primary sandbox. The hypotheses involve stacking variables: temperature gradients, precipitation accumulation, soil moisture, snowpack depth, and lake surface reflectance. A GLOF event is not a single indicator. It is an accumulation of conditions converging at a threshold the model needs to learn to recognize before it is reached.

The institutions that make field research possible

No AI model for mountain disaster forecasting runs on satellite data alone, and the expertise required to interpret what the data means is not interchangeable. Each partner in this project contributes a different layer of knowledge.

The Institute of Water Problems under the Academy of Sciences of Kyrgyzstan brings domain expertise in glaciology and hydrology — the scientific grounding to frame the right questions about what the models are trying to predict. The Tien Shan High Mountain Research Center contributes site-specific expertise at elevations and locations where standard monitoring infrastructure does not exist, helping calibrate where sensor placement and model validation need to happen. Neither institution provides datasets directly to the project; their role is expert guidance and domain knowledge that no remote sensing product can replace.

The data itself comes from CAIAG and its partners. The Central Asian Institute of Applied Geosciences, working with GFZ — German Research Centre for Geosciences — operates the SDSS (Sensor Data Storage System), a shared platform providing satellite-derived and in-situ meteorological station data across Central Asia. This is the observational record from which model training is drawn. CAIAG’s 2018 Catalogue of Glaciers of Kyrgyzstan, covering all 22 major river basins using Landsat 8 imagery at 15-metre resolution, establishes the morphometric baseline the model needs to interpret surface reflectance changes as signals rather than noise.

What the Third Pole reveals about the question itself

None of what we are doing is unique to Kyrgyzstan. High-altitude mountain communities dealing with glacial lake risk, limited early warning infrastructure, and AI models calibrated for different geographies — this pattern runs from Tajikistan to Nepal to the Chilean Andes. What Kyrgyzstan’s mountains offer is a specific, documented, and now increasingly instrumented version of a problem that is general.

The experiment is not whether AI can help. It is whether models built here — on this data, in this terrain, with these institutional partners — can later be transferred to similar landscapes without starting from zero. Transfer learning is the technical term for it. The practical question is whether a model trained on the Tien Shan can recognize the same risk signatures in the Pamir Alai, or in a glaciated watershed in Tajikistan, with minimal retraining.

That remains open. The mountain does not give up its patterns quickly. What we know so far is that the resolution matters, the local data matters, and the institutions that hold the long-term observational record matter more than the model architecture. The architecture is the part we can optimize. The rest took decades to build.

---