Research

Uncertainty in water, at three scales

One thread runs through my work: water records are short, noisy, and changing, and the decisions that depend on them are not. I build stochastic and machine-learning models that quantify what we don't know — and act anyway.

Simulate

Stochastic streamflow and rainfall

When a gauge has thirty years of record, planning for a hundred-year event is an act of extrapolation. Stochastic simulation makes that honest: generate thousands of plausible streamflow series that preserve the statistics of the observed record, then plan against the ensemble rather than the single history that happened to occur. My doctoral work introduced a multi-step method coupling PcStream clustering with Markov-chain processes for daily streamflow, validated on river basins and preserving spatial cross-correlation across gauging stations.

Method schematic: clustering to Markov chain to kappa/GEV pulses to an ensemble of streamflow series — Plate J1 PcStream–Markov streamflow simulation (Environmental Research Communications, 2025). PBIAS ±0.41%, d 0.93–1.00 across 1000 ensembles.

Recover

How droughts and reservoirs recover

Water systems fall fast and recover slowly. Across global reservoir records, the return-to-normal phase of a storage anomaly typically runs longer than its development — a persistence diagnostic for supply planning, not a universal law. Under CO₂-removal scenarios, hydroclimatic drought in the Indian monsoon region shows hysteresis: the atmosphere and the land surface recover on different clocks, decoupling by roughly a decade and a half.

Fast decline, slow return — recovery asymmetry schematic — Plate U1 Reservoir recovery asymmetry (under review). Median return ≈ 1.5× decline across 6,631 reservoirs.

CO₂ ramp-up and removal hysteresis loop for drought extent — Plate U2 Drought hysteresis under CO₂ removal (under review). ~68% peak extent, ~17-yr atmosphere–land lag, Hₛ ≈ −0.88.

Drought hysteresis under CO₂ removal: the atmosphere and the land surface recover on different clocks, decoupling by roughly a decade and a half.

Predict

GeoAI for drought propagation

Drought does not stay where it starts. A season-adaptive mixture-of-experts framework learns how drought propagates across Indian catchments, conditioning on season so a single model does not have to average over a monsoon it should be distinguishing — trained across 242 catchments and 6,690 drought events on CAMELS-IND.

Season-conditioned mixture-of-experts schematic for drought propagation — Plate C1 Season-adaptive GeoAI for drought propagation (ISG-ISRS 2025). 242 catchments, 6,690 events, CAMELS-IND.

Active research areas

Stochastic and statistical simulation of rainfall and streamflow; multi-site frameworks preserving spatial cross-correlation.
Drought characterisation, hysteresis, recovery dynamics, and propagation under climate variability and mitigation.
Machine learning and GeoAI for hydrology: mixture-of-experts, season-adaptive and physics-informed models, time-series forecasting.
Climate–hydrology interactions: responses to CO₂ removal, monsoon dynamics, land–atmosphere coupling.
Hydroinformatics: open-source modelling frameworks and production-grade pipelines.

See the publications → · Open-source software →