KG LABS

Kyrgyzstan · glacier projections 2020–2100

Kyrgyzstan's glaciers, year by year to 2100

Every glacier is drawn with its mapped outline as a sheet of ice on the relief. As the years run, the ice thins to the share of area that global glacier models keep, and turns to a dark scar in the year the models say the glacier is gone. Choose a warming level and press play.

Global model projectionVan Tricht et al. 2025Randolph Glacier Inventory 7.0
2050
+2.7 °C · published years
Building the terrain…

Warming levels are the rise in global mean temperature over 2071–2100 above pre-industrial, as simulated by each CMIP6 climate model run. They come from climate models, not from satellite or ERA5 observations. Each level pools the runs that fall in a band: +1.5 °C covers 1.1–1.6 °C, +2.0 °C covers 1.5–2.45 °C, +2.7 °C covers 2.2–3.4 °C and +4.0 °C covers 3.7–4.4 °C. Observations enter elsewhere: the glacier models were calibrated with satellite measurements of glacier mass change for 2000–2019 and driven by ERA5-based reanalysis climate for the past. Ice thickness is exaggerated.

Global projections, Kyrgyz evidence

These are global projections. We are fitting them to Kyrgyzstan and the Pan-Third Pole.

The map shows global glacier models run for every glacier on Earth with global climate and glacier data. KG Labs is fine-tuning their calibration for Kyrgyz mountains — and for the other high-mountain countries of the Pan-Third Pole, which share the same ranges and the same gaps in measurement — with four kinds of evidence.

Real-time data links

LoRaWAN, Wi-Fi HaLow and satellite IoT carry readings out of remote valleys with no mobile coverage. At Adygene, a station already sends environmental readings every 10 minutes, all year, over LoRaWAN and cellular networks.

Ground sensors

Low-power stations for precipitation, air and ground temperature, water level, soil moisture and snow, placed where a measurement changes a decision and owned by someone who maintains it.

Geospatial data

Sentinel-1 and -2, Landsat, elevation models, glacier and lake inventories and recovered Soviet-era catalogues, processed the same way for every glacier, lake and basin in the country.

AI models and frameworks

Earth-observation foundation models, AI weather models and time-series models, each tested against Kyrgyz observations and a simple baseline before it is used.

The aim is glacier, water and hazard projections that hold up against measurements in Kyrgyzstan and across the Pan-Third Pole, not only against global averages.

What you are looking at

Four layers on one relief model

The glaciers

Each glacier keeps its real shape from the glacier inventory: long valley tongues, branching tributaries and small cirque glaciers. Its ice thickness shows the share of its 2000 area still left, and its colour runs from deep blue for low glaciers to pale ice for the highest. When the projected extinction year passes, it turns to a dark, flat scar.

The glacier inventory

The pale fill is every outline in the Randolph Glacier Inventory 7.0 whose centre lies in Kyrgyzstan: 11,181 glaciers, about 5,800 km², including some 4,000 small ones the models did not run on. The inventory was released in 2023 but describes glaciers around 2000, the common starting point for glacier models. Its Kyrgyz outlines were traced from satellite images taken in 1994–2008, three quarters of the ice area from 2002.

The permafrost band

Above about 3,200 m the ground in Kyrgyzstan's ranges can stay frozen all year. Ochre marks 3,200–3,500 m, where field studies in the Ala-Archa valley found patchy permafrost; rust marks continuous permafrost above 3,500 m. It is an elevation band from those studies, not a map of where permafrost has been found.

Terrain and lakes

Valleys run from green through tan and brown to grey rock and snow, shaded by a low north-west sun. Issyk-Kul, Son-Kul, Chatyr-Kul, Toktogul and the other lakes and reservoirs larger than 1 km² keep their natural colour from cloud-free Sentinel-2 satellite imagery.

By warming level

How many of 7,336 glaciers are gone

Published years come from the authors' released code. Corrected years are our recomputation with a fix to that code, which moves the median year about three years later.

Warming by 2071–2100Glaciers gone byStill standing in 2100Ice area left in 2100
2030204020502100

Each cell shows published years, then corrected years in grey. Ice area left is the models' median area as a share of the 2000 area of the 7,336 glaciers.

Method

How this was made

The projections are from Van Tricht and colleagues (2025, Nature Climate Change), who ran three global glacier models (GloGEM, OGGM and PyGEM) with many CMIP6 climate model runs and reported, for every glacier, the median year in which it is effectively gone. We took the glaciers whose centre lies in Kyrgyzstan and reproduced the authors' published years exactly.

The models ran on the older RGI 6.0 inventory, so each modelled glacier is linked to the RGI 7.0 outline it overlaps most and drawn with that outline. Where one RGI 7.0 glacier was split into several pieces in RGI 6.0, each piece is drawn with its own RGI 6.0 outline.

We found a defect in the released code: a rule meant to act at 1% of the initial ice volume acts at about 3.25%, which makes published years early. The corrected years are our recomputation; the switch on the map shows both. Glaciers whose median year is 2101 count as gone by 2100.

These are global model projections. KG Labs is fine-tuning their calibration for Kyrgyzstan and for the other high-mountain countries of the Pan-Third Pole, against measurements taken in those mountains.

KG Labs research

AI and climate research for Kyrgyzstan

KG Labs Public Foundation builds the research and computing layer for forecasting climate dynamics and natural hazards in Kyrgyzstan: national datasets, reproducible pipelines, tests of AI and weather models, and briefs for the institutions that act on them. Official warnings remain with the responsible state agencies.

46research experiments run
11,481glaciers with a national data file
2,592glacial lakes tracked
929,924lake-weeks of satellite measurements and reconstructed lake states
1,769recorded floods, mudflows, outbursts and landslides, 1846–2026
152,143river discharge observations from 116 stations
22,542national glacier image tiles analysed by AI, with no failed tiles
≈250,000chip-epochs of model training on H100 GPUs
9 × 6hazard types × forecast horizons (24–168 hours) in a daily research issue

Figures from KG Labs records on 7 October 2026; image-tile and GPU-training figures are from the September 2026 national run.

Collaborate

Work with us on Kyrgyzstan's glaciers and hazards

We are looking for partners in field monitoring, sensor networks, data sharing and research: government agencies, universities, telecom operators, donors and international programmes.

Email [email protected]
WhatsApp +996 990 777 700