27.7172° N · 85.3240° E — Kathmandu, Nepal

Seeking a funded MSc / PhD research assistantship · Spring / Fall 2027

Ayush
Karki

Civil Engineering graduate · Hydrology & Hydraulics · Python & Machine Learning as applied techniques

My major interest is in hydrology and hydraulics — how water moves through catchments, rivers and hydraulic structures. I use Python, machine learning and remote sensing as applied techniques to study these problems, especially where data is scarce and a model needs to stay honest about its own uncertainty. I'm hoping to join a research group as an assistant, contribute to this kind of work, and keep learning from people who do it well.

Download CV GitHub Email me
Portrait of Ayush Karki

Selected work

Under review · Georisk · First author

Landslide Susceptibility Mapping with a Reliability Layer

An ML susceptibility map for the Karnali Highway corridor, paired with an applicability-domain layer that flags where the model should not be trusted.

Read more ↓

Personal project

A Reliability-Aware ML Surrogate for EPA-SWMM

A machine-learning surrogate that approximates SWMM stormwater outputs far faster, with the same reliability layer to flag out-of-distribution inputs.

Read more ↓

Personal project

GR4J Rainfall-Runoff Model with Metaheuristic Calibration

A conceptual daily rainfall-runoff model built from its governing equations, calibrated with differential evolution to KGE ≈ 0.88 and probed for equifinality with an RS-SPSO search, on a monsoon-analogue catchment.

Read more ↓

Final-year thesis

Pre-Feasibility Study of the Manohara Irrigation Project

A full hydrology-to-hydraulic-design workup for an irrigation scheme, from monthly flows and the 100-year flood to intake and canal design.

Read more ↓

Profile

I'm a Civil Engineering graduate from Pulchowk Campus, IOE, Tribhuvan University. My major interest is in hydrology and hydraulics — the movement of water through catchments, rivers and hydraulic structures. To study these problems I apply Python and machine learning, along with remote sensing, as practical tools. I'm drawn to work where physical understanding and data-driven methods meet, and to methods that hold up in the data-sparse conditions common to real-world engineering.

DegreeB.E. Civil Engineering
StandingFirst Division
FocusHydrology & Hydraulics · Applied Python & ML
Based inKathmandu, Nepal

Research & Projects

Work that turns terrain data into decisions.

Under review · Georisk First author

Multi-Scale Pixel and Slope-Unit Landslide Susceptibility Mapping with Applicability-Domain Validation

Georisk (Taylor & Francis) · Karnali Highway corridor (~2,155 km²), western Nepal

Landslides along Himalayan road corridors cause repeated loss of life and long disruptions to access, yet most susceptibility maps never say how far their predictions can be trusted away from the training data. This study maps susceptibility for the Karnali Highway and pairs the map with an explicit reliability layer — a spatial guide to where the model should, and should not, drive engineering decisions.

Read the preprint (EarthArXiv) → doi.org/10.31223/X5C21W View the code on GitHub → github.com/ayushkarkiofficial-lgtm/Landslide_Susceptibility_Mapping

Methods & technical contribution

  • Feature engineering: built nine topography, hydrology, land-cover and road-proximity conditioning factors, with raster workflows for NDVI/NDWI, resampling, masking, raster-stack preparation and prediction-map export (rasterio, GeoPandas, NumPy).
  • Multi-scale design: compared pixel and slope-unit mapping units across moving-window scales of 3×3, 15×15 and 25×25 cells on a 12.5 m DEM.
  • Modelling: trained and tuned Random Forest, RBF-SVM and XGBoost with Bayesian search, Platt-scaled probability calibration, and an equal-weight mean-probability ensemble (scikit-learn).
  • Honest validation: used spatially blocked cross-validation with 1,000 m blocks to strip out spatial-autocorrelation optimism, then reported ROC/AUC, confusion matrices and SHAP feature-importance analysis.
  • The reliability layer: built a Mahalanobis-distance applicability-domain mask that flags where predictions extrapolate beyond the training feature space, published as a trusted / untrusted binary.
  • Deep learning: prototyped CNN workflows with Optuna-based tuning for patch-based hazard modelling (PyTorch).
Complete Personal project

A Reliability-Aware ML Surrogate for EPA-SWMM Stormwater Simulation

Personal research project · EPA-SWMM · Python (pyswmm · swmmio · scikit-learn / XGBoost)

Physics-based stormwater models such as EPA-SWMM are the standard for urban-drainage and overflow planning, yet they are far too slow to run the thousands of times that uncertainty analysis, real-time control or city-scale screening demand. As a personal project, I explored whether a machine-learning surrogate could reproduce SWMM's key outputs much faster, and paired it with the same applicability-domain reliability layer from my landslide work, so the surrogate flags when it is extrapolating rather than returning a silent, out-of-distribution guess.

View the code on GitHub → github.com/ayushkarkiofficial-lgtm/SWMM_surrogate_modeling

Methods & results

  • Training data: Latin-Hypercube sampling across rainfall depth, imperviousness, slope, conduit capacity, roughness and infiltration drove hundreds of headless SWMM runs (pyswmm + swmmio) into a dense parameter-to-response dataset.
  • Surrogate accuracy: Random Forest and XGBoost surrogates reproduce peak discharge at R² ≈ 0.98, classify flood occurrence at ROC-AUC ≈ 0.99, and predict flood volume at R² ≈ 0.92 on cross-validated and parameter-extrapolation hold-outs.
  • Speed-up in action: with the trained surrogate I ran a 50,000-sample Monte-Carlo uncertainty analysis of flood risk — a sweep that is impractical at SWMM's native runtime.
  • The reliability layer: a Mahalanobis-distance applicability-domain gate flags out-of-distribution inputs; surrogate error rises exactly where the gate fires, so the fast model defers when it should not be trusted.
  • Adaptive sampling: AD-guided adaptive sampling targets under-covered regions of the input space, benchmarked against random augmentation.
Complete Personal project

Physically-Based Rainfall-Runoff Modelling with Metaheuristic Calibration

Personal research project · GR4J · CAMELS-AUS · Python (NumPy · pandas)

Conceptual rainfall-runoff models turn daily rainfall and evaporation into streamflow, and they are the workhorses of water-resource planning in data-sparse basins. I built the GR4J model directly from its governing equations and calibrated it against a KGE objective with differential evolution, deliberately choosing a snow-free, monsoon-analogue catchment as a stand-in for Nepali conditions. To probe the equifinality of well-performing parameter sets I then ran a respawning speciation-based PSO (RS-SPSO), and stress-tested transferability with Klemeš split-sample and differential split-sample validation.

View the code on GitHub → github.com/ayushkarkiofficial-lgtm/GR4J-metaheuristic-calibration

Methods & results

  • Model from first principles: implemented the GR4J production and routing stores and the S-curve unit hydrographs directly from the governing equations (Perrin et al., 2003) in NumPy and pandas.
  • Clean data pipeline: turned raw CAMELS-AUS v2 records into a paired daily series of precipitation, Morton PET and streamflow, with no-data handling and a long-term water-balance check.
  • Baseline calibration: fit the four GR4J parameters against a KGE objective by differential evolution (SciPy) with a two-year warm-up, guarding the search against non-physical parameter sets.
  • Honest evaluation: scored with NSE, KGE (with its r, α and β decomposition) and PBIAS, reaching KGE ≈ 0.88 and NSE ≈ 0.76 with PBIAS under 1% on the study catchment (CAMELS-AUS gauge 114001A).
  • Multimodal calibration & equifinality: extended the search with a respawning speciation-based particle swarm (RS-SPSO) that recovers several distinct parameter sets all calibrating to KGE ≈ 0.84–0.87 — a direct look at the equifinality hidden behind a single "best" fit.
  • Transferability testing: applied the Klemeš split-sample and differential split-sample scheme — calibrating on the wettest multi-year block and validating on the driest, and vice-versa, with warm-up handled per split. Calibration KGE ≈ 0.83–0.91 falls to validation KGE ≈ 0.23–0.48 under regime transfer — a candid measure of how far the calibrated model does, and does not, travel between climates.

Final-Year Thesis

Pre-Feasibility Study of the Manohara Irrigation ProjectOpen PDF ↗

A full pre-feasibility workup for an irrigation scheme, completed as our final-year group project: we analysed DHM hydro-meteorological records to derive monthly flows, irrigation water demand and the 100-year flood discharge, then carried the hydrology through to the hydraulic design of the intake, canal, settling basin and distribution structures using HEC-RAS and QGIS.

DomainWater resources & hydraulics
HydrologyMonthly flows · demand · 100-yr flood
ToolsHEC-RAS · QGIS · DHM data
DeliverableIntake, canal, settling basin, distribution design

Capabilities

Every tool here has done real work.

Geospatial & Remote Sensing

QGISPrimary GIS for terrain analysis, map production and the full landslide-mapping workflow.
ArcGISSpatial data preparation, geoprocessing and cartographic output.
rasterio · GeoPandasPython raster/vector pipelines: NDVI/NDWI, resampling, masking, raster-stack assembly, prediction-map export.
DEM / terrain analysisSlope, aspect, topographic wetness index and multi-scale moving-window features on 12.5 m DEMs.
SNAPSatellite imagery processing for remote-sensing inputs.
PostGISSpatial queries and storage of vector datasets in PostgreSQL.

Machine Learning

scikit-learnRandom Forest, SVM and the full training/evaluation stack — pipelines, metrics, cross-validation.
XGBoostGradient-boosted models; strongest performer in the landslide ensemble.
Bayesian tuning · OptunaHyperparameter search for classical and deep-learning models.
Probability calibrationPlatt scaling and calibrated probabilities so model scores mean something.
Spatial cross-validationSpatially blocked CV to remove autocorrelation optimism from reported skill.
SHAPModel interpretation and feature-importance analysis.
Applicability domainMahalanobis-distance masks to flag out-of-distribution predictions.
PyTorchCNN prototypes for patch-based raster hazard modelling.

Hydrology & Water Resources

HEC-RASHydraulic modelling and flood analysis for river and canal design.
Flood-frequency analysisReturn-period discharge estimation, including the 100-year flood.
Hydrologic modellingMonthly-flow and water-availability analysis from DHM records.
EPANETPressurised water-distribution network modelling.
CROPWAT · HydrognomonIrrigation water demand and hydrological time-series processing.

Programming & Scientific Computing

PythonPrimary language — NumPy, pandas, Matplotlib across every modelling and geospatial task.
CFoundational programming and algorithmic work.
LaTeXScientific writing and manuscript preparation.
HTML · CSS · JavaScriptFront-end basics — including this site.

Education

2026

IELTS Academic

Overall Band 8.0 · CEFR Level C1 — English-language proficiency.

Test date: 19 September 2026. View IELTS score report (PDF) ↓
Public copy with personal identifiers redacted. Original report available on request.

2019 — 2024

B.E. in Civil Engineering

Pulchowk Campus, Institute of Engineering, Tribhuvan University, Nepal · First Division.

Hydrology Water Resources Engineering Hydraulics Fluid Mechanics GIS & Remote Sensing Numerical Methods Probability & Statistics Engineering Design Surveying
Show all courses · full curriculum →
2017 — 2019

Higher Secondary (10+2, Science)

Himalayan WhiteHouse International College, Kathmandu, Nepal · CGPA 3.61 / 4.0.

Physics Chemistry Mathematics Computer Science

Get in touch

ayushkarki.official@gmail.com
Kathmandu, Nepal