Research / Reliable ML / Scientific modelling

From fast predictions to decisions that expose uncertainty.

My primary research asks how an ML surrogate can report more than a point estimate: where error is likely, whether uncertainty is calibrated, and when a prediction should enter a review queue.

RecruiterWhat did the work establish?

EngineerHow does uncertainty change a system decision?

Researcher / studentWhich protocol supports each claim?

01Research focus

A weighted research identity, not a flat keyword list

Reliable ML and uncertainty-aware graph surrogates are the primary body of work. Scientific modelling supports that direction; identifiability is an emerging mathematical inquiry rather than a claimed result.

Primary research

Reliable machine learning

Uncertainty quantification for graph surrogates: ranking likely error, calibrating uncertainty, constructing conformal intervals, and routing uncertain predictions to review.

Uncertainty Quantification
Calibration
Conformal Prediction
Selective Prediction
Graph Neural Networks

Reliable GNN Surrogates for Transport Policy Analysis

Supporting work

Scientific and data-driven modelling

Hydrological uncertainty analysis and bounded forecasting experiments connect model evaluation to physical systems, temporal structure, and explicit validity limits.

Scientific Machine Learning
Data-driven Modelling
Time-series Methods

Uncertainty Quantification in Hydrology Synthetic Streamflow Forecasting Benchmark

Emerging inquiry

Mathematical structure of learned models

Neural-network identifiability is an active mathematical ML direction, presented as an area of inquiry rather than a completed research result.

Neural Network Identifiability

02Primary record

Uncertainty Quantification for ML Models in Transportation Policy Analysis

The submitted master's thesis studies a GNN surrogate for Paris capacity-reduction scenarios, with separate evidence for point accuracy, uncertainty ranking, calibration, conformal coverage, and selective review.

Academic research / Master's thesis submitted

Uncertainty Quantification for Machine Learning Models in Transportation Policy Analysis

Built on the MATSim corpus and PointNetTransfGAT infrastructure of prior work, then evaluated training variants, MC Dropout, deep ensembles, sigma scaling, split and adaptive conformal prediction, selective prediction, CQR, and error-detection diagnostics.

Evaluated scope: 100 scenarios3,163,500 link predictions
Uncertainty ranking: ρ 0.482MC uncertainty vs. absolute error
Selective review: 41.2%Lower MAE at 50% retention

Explore the thesis research Engineering case study Canonical repository

03Supporting work

Scientific uncertainty and time-dependent models

These projects extend the modelling context without being presented as equivalent to the thesis research record.

Group coursework

Uncertainty Quantification in Hydrology

A team seminar connecting HBV rainfall-runoff calibration, local and global sensitivity analysis, and input/output uncertainty propagation.Group coursework; individual ownership of each result is not established

Synthetic demonstration

Synthetic Streamflow Forecasting Benchmark

A deterministic benchmark comparing seasonal-naive, SARIMAX, and gradient-boosted one-step streamflow predictions.Synthetic data cannot establish real-catchment performance

04Learn

Turn the research question into an operational concept

The learning layer explains one decision mechanism at a time, while the thesis record preserves protocols, evidence status, and limitations.

tutorial4 min read

Selective Prediction: When Models Should Abstain

A practical guide to risk-coverage trade-offs: use uncertainty to decide which predictions to accept, review, or defer.

Aug 20, 2026Uncertainty Quantification