Water Research Trends
What is heating up and what is cooling down in water science, read from three years of European Geosciences Union abstracts. Then, what it means for you, whichever side of water you work on.
What is rising, what is cooling
Change in each theme's share of water research between 2024 and 2026, in percentage points. Only themes whose trend survives statistical correction are shown. Bar colour marks direction; the tooltip shows how AI-heavy each theme is.
* Rises in the main run but is not stable when the clustering is re-seeded, so treat it as tentative rather than settled.
The rising themes, year by year
Share of water research held by each rising theme, 2024 to 2026. The two deep-learning themes climb steadily every year. Carbon-water coupling and the flood-social theme (dashed) are driven mostly by a single year, so read them with more caution.
* Flood-social theme is not stable across re-seeding (see note above).
The biggest topics, and how much AI has reached each
The largest themes in water research by volume, with the share of each that already uses machine learning. Note groundwater: third-largest topic, yet among the least reached by AI.
Where AI has not reached yet
Every one of the 50 themes plotted by how much of water research it holds against how much of it uses machine learning. The dashed line is the corpus-wide average of 13.2%. Themes to the lower right are the interesting ones: a lot of research, very little AI. Bubble size is the number of abstracts.
The five biggest themes AI has barely touched
Ranked by size weighted against how little machine learning they use.
The semantic map of water research
Each faint dot is one abstract from a 2,800-abstract sample of the 20,416 (plotting all of them would be an unreadable smear), placed so that semantically similar work sits close together — a UMAP projection of the text embeddings. The bubbles are the real thing: all 50 clusters, each sized by how many of the full 20,416 it holds and labelled by its most distinctive words. Blue clusters are rising, amber are cooling, grey are steady. Distance and grouping carry the meaning; the axes have no units. Hover a bubble for detail.
All 50 themes, in full
The charts above show the themes that move most. This is the complete set: every cluster the analysis found, with its size, its AI penetration and its three-year trend. Search it, sort it, and open any theme for its own page.
Showing 50 of 50 themes. Corpus-wide AI penetration averages 13.2%; the bar shows how each theme compares.
large-sample, tws, grace, baseflow, rainfall-runoff, swat
gpp, vpd, nee, deforestation, lai, leaf
pumping, springs, hydrochemical, abstraction, nitrate, mar
ssf, bedload, hillslope, bed, suspended, gully
nexus, governance, wefe, wef, institutional, multi-objective
pluvial, levee, hydrographs, breach, copula, hydrograph
leaf, beech, sap, spruce, oak, transpiration
convection-permitting, gcms, rcms, wrf, cpms, cpm
cyclones, sst, enso, extratropical, rossby, cyclone
speleothem, holocene, miocene, speleothems, cave, carbonate
lst, smap, vod, ssm, microwave, ismn
routing, swat, modflow, nse, sewer, stormwater
spi, spei, deficits, cdhws, drought-to-flood, spei-
insurance, evacuation, multi-hazard, households, agent-based, preparedness
wheat, maize, farmers, rice, rainfed, soybean
compaction, roots, loam, tillage, transpiration, ert
mcss, nowcasting, lightning, pws, imerg, idf
lstm, lstms, differentiable, interpretability, surrogate, ungauged
soc, respiration, biochar, rhizosphere, stocks, fungal
rainfall-induced, slow-moving, lews, deep-seated, insar, geotechnical
subglacial, greenland, meltwater, sheet, supraglacial, proglacial
ism, easm, enso, tibetan, aerosol, jet
pfas, sorption, adsorption, metal, metals, wastewater
lstm, nse, post-processing, xgboost, ann, ungauged
nitrate, phosphorus, doc, wastewater, dom, pharmaceuticals
swot, sar, altimetry, wse, ssc, aperture
heatwave, compounding, multi-hazard, cyclones, dengue, high-impact
nbs, governance, participatory, wefe, nexus, stormwater
mortality, gpp, sar, xgboost, segmentation, peatlands
doc, dom, dic, inorganic, alkalinity, phosphorus
swe, snowfall, avalanche, ros, snow-dominated, sublimation
fault, faults, tectonic, mantle, geothermal, fracture
generative, nowcasting, convolutional, u-net, cnn, adversarial
estuary, estuarine, estuaries, lagoon, slr, adriatic
fluids, calcite, rocks, carbonate, mantle, minerals
peatlands, peatland, peat, rewetting, ghg, bog
methane, ghg, chamber, ozone, respiration, biogenic
tibetan, himalayan, glof, outburst, nepal, himalaya
microwave, earthcare, ghz, gpm, gnss, clouds
antarctic, sheet, greenland, antarctica, smb, subglacial
students, teachers, school, educational, co-creation, game
clouds, aerosol, aerosols, microphysical, inps, droplet
multi-hazard, emergency, warnings, impact-based, language, news
fracture, porous, pore, permeability, solute, fractures
earthquake, resistivity, interferometry, insar, fault, bedload
geothermal, ates, ht-ates, borehole, mines, coal
wildfire, fires, post-fire, burned, fuel, fwi
microplastics, mps, microplastic, plastic, plastics, polymer
auroral, electron, electrons, geomagnetic, energetic, ionospheric
crns, neutron, neutrons, cosmic-ray, cosmic, ray
The vocabulary of water research, grouped by meaning
The 160 most distinctive words across the corpus, sorted into 10 groups by meaning. Words that appear in similar research sit in the same group; a larger chip means the term shows up in more abstracts.
What this means for you
The same data reads differently depending on where you sit. Here is the short version for four kinds of reader.
- Deep-learning streamflow and flood prediction is the strongest riser: its share climbed from 1.7% to 2.8% of water research, and 98.9% of these abstracts use machine learning (LSTM, differentiable, physics-informed).
- Groundwater is one of the least AI-penetrated major topics: just 10.5% of its 3,789 abstracts use ML, against 24.4% for rainfall-runoff. Pairing aquifer physics with learning models is open thesis territory.
- Carbon-water-vegetation coupling rose fast (2.7% to 3.8% of water research) yet only 10.3% uses ML. GPP, VPD and evapotranspiration mark a methods-meets-ecohydrology niche barely touched by AI.
- The classic groundwater cluster (pumping, springs, managed recharge, saltwater) holds 620 abstracts but just 4.7% use ML, and its share is flat. Physics-heavy, AI-light, wide open.
- Drought is the fastest-climbing of the very large topics, up 1.94 points (18.4% to 20.4% of water research), ahead of floods at +1.42. Scarcity is rising up the agenda quickly.
- Extreme and compound hazards jumped +2.4 points (9.2% to 11.6%), the biggest single-topic climb in the whole corpus. Compound heat, drought and flood events are now a core planning concern.
- Groundwater is the largest topic inside Hydrological Sciences: 2,089 abstracts, 27.1% of that division. For India's aquifer stress, the directly relevant science base is deep.
- Water resources management and policy rose +1.7 points to 6.1% of water research, and 19.9% now use ML. Decision-support tools are moving into governance, not just the lab.
- AI methods are among the fastest-rising themes: two of the four top risers are 96 to 99% machine learning. But the biggest topic-level growth (extremes, drought) is largely non-AI, so build on AI where it fits the problem, not everywhere.
- The clearest whitespace is AI for groundwater: 3,789 abstracts make it the third-largest topic, yet only 10.5% use ML, the lowest of the three biggest topics. An uncrowded place to lead.
- Monitoring, data and sensors is rising +1.6 points with 20.2% AI adoption already. Sensor-plus-analytics products ride a growing, partly digitised market rather than a cold start.
- Cooling areas to weigh carefully: crop-water agronomy fell 0.91 points (count down 24%), contaminant work 0.70, geothermal storage 0.51. All three are losing research attention.
- Rainfall-runoff and streamflow prediction has the highest AI adoption of any topic at 24.4% ML. Pairing hydrology with LSTM and CNN skills sits where hiring is heading.
- Remote sensing is both large (18.1% of water research) and the most AI-heavy big topic at 22.4% ML. A strong entry point if you like working with data and code.
- Groundwater work is abundant (3,789 abstracts) but only 10.5% uses AI. Combining hydrogeology with one ML method is rare and valuable, especially in India's groundwater sector.
- Deep-learning precipitation forecasting grew fastest in count of all rising themes, up 90.7% over three years, and 96.1% is ML. An emerging niche skill with little competition.