Statistical learning for data that are incomplete, heterogeneous and dynamic

Research pillar 01

Learning under complex data conditions

How does informative missingness change what can be learned from labelled and unlabelled data?
When can partially observed data improve inference rather than weaken it?
How can robust procedures remain reliable under outliers, heavy tails and distributional heterogeneity?

Research pillar 02

How can difficult high-dimensional and non-convex objectives be solved more efficiently?
How should hyperparameters be selected without excessive tuning or unstable validation?
How can iterative and stochastic algorithms balance speed, accuracy and robustness?

Research pillar 03

How can deterministic and probabilistic forecasts remain calibrated under changing regimes?
How can physical knowledge be combined with statistical and machine-learning models?
How can interventions and causal mechanisms be studied in dynamic systems?

Current directions

Current work extends these themes into trustworthy AI, physics-informed learning and uncertainty-aware prediction.

Understanding when the mechanism that determines what is observed carries information that should be modelled directly.

Combining governing structure with noisy, incomplete or large-scale observations in environmental and engineering systems.

Developing models that quantify uncertainty across energy, groundwater, climate and environmental applications.

Examining robustness, transparency and responsible deployment of AI in geoscience and other scientific domains.

Research collaboration

I am particularly interested in collaborations where statistical structure, computational difficulty and domain knowledge must be addressed together.