Ben Evans

Machine learning for monitoring, understanding and forecasting coupled ice sheet – ocean – climate systems

Research Area

I build understanding of polar environmental systems from a data-driven perspective with a focus on floating ice, whether that be sea ice, icebergs or ice shelves. I am particularly interested in how we can apply machine learning approaches to better understand the influences of floating ice on ocean freshwater distributions. These drive thermohaline circulations and impact primary productivity and biogeochemical cycling, though these impacts are currently poorly understood. Much of my work also serves to support navigational safety and environmentally-aware route planning for scientific campaigns. 

Key challenges in this space that will help to close our knowledge gaps around global climatic impacts, improve numerical simulations and facilitate safe, efficient science and navigation in polar regions include: 
 

• Improved predictability of ice shelf instability and iceberg calving processes, including the effects of hydrofracture and other processes identifiable from satellite observations. 
• Scalable, high-cadence detection systems for icebergs that span sensors and scales. Iceberg tracking systems that will supply full life-cycle insight to dynamics and ultimately support iceberg trajectory, melt and fragmentation forecasting models.
• Improved sea ice predictive capability, including through data-driven models that couple ice, ocean and atmosphere at seasonal and sub-monthly timescales or hybrid or emulator-based approaches to mechanistic sea ice modelling. 

Methodological innovation is needed to address the above challenges. A concrete example might be the surface hydrology of ice shelves or large tabular icebergs and its role in hydrofracture-driven collapse. The triggering processes act at fine scales - metre-scale crevasses, tens-of-metre scale lakes, centimetre-scale firn percolation and drainage events lasting hours or just minutes. The available observations are typically too coarse or too spatiotemporally sparse on their own so we need to exploicomplementary: optical imagerter bodies but fails under cloud and polar night, SAR sees through both and finds buried lakes but usually only has one or two noisy channels, laser altimeters provide sparse-tracks of high-fidelity topography and passive microwave supplies a long term record at very  coarse resolution. This makes the domain well suited to fusing multi-sensor observations and coupling them to emulators. It also allows for the development of advanced downscaling and interpolation models for the crysophere 

Key variables for ice shelves, icebergs and sea ice are only ever partially observed – for example firn air content is essentially unobservable at process scale. This motivates critical probing of whether the machine learning models trained on sparse cryospheric observations learn physically meaningful representations, or simply convenient predictions. a sea-ice-ocean-atmosphere emulator must implicitly represent subsurface ocean heat and ice thickness that satellites resolve poorly or an iceberg dynamics model must represent dynamical fields governing movement and fragmentation. Whether AI models internalise variables and physics that they are never directly shown, and whether they do so faithfully - rather than by exploiting spurious observational correlations - determines how far they can be trusted, particularly when extrapolating to regimes with little training analogue (crucial for cryospheric projections). A growing body of work using tools like probing or representational similarity analysis should allow a student to explore from a more theoretical perspective whether AI models adequately learn the geophysical state they must implicitly reason about. 

My work spans Earth observation, multimodal AI, sparse data assimilation, and machine learning-driven modelling of dynamic interactions at the interface of ice, ocean and atmosphere. While I am trying to develop fully observation-based forecasting systems, I also work closely with numerical modellers, seeking to provide them with better validation data, improved parameterisations and insight to the physics we need to represent. Another key aspect of interest is probing and elucidation of how (whether) machine learning models learn and represent the physics of complex geophysical systems, and what this means for how we should be using such systems. 

Project Interests

I welcome applications looking to address any aspects of the key challenges identified above, as well as thematically or methodologically-aligned ideas that help to improve our understand how ice sheets are coupled to the global climate system via the ocean and shelf seas. I am interested in discussing projects focusing on the Antarctic, Arctic or both.