Chilly models

Flying over Greenland and looking down on land previously covered by the island’s massive ice sheet, Stephen Price got a bird’s-eye view of part of his shared challenge: a rippled topography of knolls, troughs and valleys.

“This is exactly what an ice sheet looks like underneath the ice, and this has a huge impact on the ice dynamics,” says Price, a former hands-on glaciologist turned computational scientist with the Fluid Dynamics Group at Los Alamos National Laboratory (LANL). Price imagined imposing the grid of a computational model, with each section or cell several square kilometers across, on the terrain.“You realize that we’re really simplifying this complexity in our current land-ice models.”

Creating a more detailed model that accurately captures complicated land-ice-ocean interactions is Price’s job as co-principal investigator (co-PI) of PISCEES – Predicting Ice Sheet and Climate Evolution at Extreme Scales.

The collaborative effort includes researchers at four Department of Energy (DOE) labs, four universities and the National Center for Atmospheric Research (NCAR). PISCEES is at the global forefront of modeling to predict the loss of land-ice mass and the resulting rise in sea levels, with a focus on Antarctica.

“What happens with the Antarctic ice sheet over the next century is the biggest unknown in terms of what’s going to happen with future sea-level rise, and we don’t yet have models that can make accurate predictions about that,” Price noted in a presentation this past summer at the annual meeting of principal investigators in DOE’s SciDAC-3 (the third phase of the Scientific Discovery through Advanced Computing program).

PISCEES will form the Earth-based ice component of DOE’s new Accelerated Climate Model for Energy (ACME) project, meeting a key goal of accurately coupling ice-sheet and ocean models.

“We’re applying the DOE’s most sophisticated mathematical tools to solving a problem that’s relevant to everyone,” says Esmond Ng, PISCEES co-PI and head of the Applied Mathematics and Scientific Computing Department at Lawrence Berkeley National Laboratory (LBNL).

With access to DOE’s top computers at Oak Ridge National Laboratory’s Leadership Computing Facility and LBNL’s National Energy Research Scientific Computing Center, PISCEES might soon help make the first reliable estimates of future sea-level rise due to ice-sheet mass loss.

This would be a major international accomplishment. The 2007 Intergovernmental Panel on Climate Change (IPCC) Fourth Assessment Report said that land ice models weren’t good enough to predict ice sheets’ contribution to sea-level rise.

“For our community, this was a big glaring gap,” says Price, part of LANL’s Climate, Ocean and Sea Ice Modeling (COSIM) group. The group has an unmatched record in contributing both the ocean and sea-ice models presently used in the Community Earth System Model, an IPCC-class climate code DOE co-developed with NCAR, and will lead similar activities for ACME.

A fundamentally different way of modeling

PISCEES incorporates two distinct land-ice models, or dycores (dynamical cores), that aim to improve model accuracy using widely spaced meshes of computational data points: BISICLES, which uses a regular, structured mesh; and FELIX, which uses an unstructured mesh.

Unlike earlier simplified models, both codes include a more complete description of the relative stresses involved and focus closely spaced meshes and computational resources on the fastest-moving, most dynamic parts of the ice sheet.

BISICLES, a joint effort with University of Bristol researchers in Great Britain, was developed in response to a 2009 call for land-ice model initiatives from DOE’s Advanced Scientific Computing Research program – issued in response to the land-ice modeling hole the IPCC report identified. The model draws on LBNL’s expertise in block-structured adaptive mesh refinement (AMR) – a technique to zoom in and focus finely-spaced meshes where needed to follow changing conditions within a larger-scaled grid.

“BISICLES was designed from day one to attack the resolution issue,” Ng says. This is critical in dynamic features that can affect the ice flow deep into the interior of the continent, like ice streams, outlet glaciers, floating ice shelves and grounding lines, the point at which seaward-flowing ice begins to float due to buoyancy.

Unlike BISICLES, FELIX uses an unstructured mesh. It “is a fundamentally different way of modeling,” Price says. “With BISICLES and FELIX, we’re covering both different meshing approaches and different approximations to the governing stress balance. No one knows at the moment what the better approach is going to be and it is likely that the answer to that question will vary depending on specific circumstances.”

Both the FELIX and BISICLES dycores rely heavily on tools from FASTMath, a SciDAC institute on applied mathematics. FELIX is built on Trilinos, a framework supported by FASTMath members at Sandia National Laboratories. And BISICLES is built on two FASTMath-supported tools: the Chombo AMR framework and PETSc. “Without the FASTMath components we’d have a hard time writing these codes,” says Ng, who also is LBNL’s FASTMath Institutional PI.

BISICLES already is helping significantly improve coupled ice sheet and climate model simulations by, for example, coupling with a modified version of the Parallel Ocean Program model developed at LANL and the Naval Postgraduate School.

“Our effort is the only one that has a large-scale, high-resolution, eddying ocean circulation model coupled to a high-resolution, large-scale ice-sheet model,” says Price, describing a collaboration with former LANL post-doctoral researcher Xylar Asay-Davis, now at the Potsdam Institute for Climate Impact Research. “Both recent observations and modeling argue that the interaction with the surrounding ocean is probably the really important process in driving ice mass loss.”

A massive pocket of relatively deep, salty and warm water forms a belt outside the continental shelf around Antarctica. Periodically this warmer water rises onto the continental shelf and flows underneath the floating ice shelves, accelerating sub-shelf melting. That affects the movement of ice farther upstream, including grounding-line retreat and accelerated ice flow and melting.

“We’re at the stage where we can run an ocean model that circulates under these ice shelves and calculate melt rates that affect the ice-sheet model, which changes the ice-shelf geometry and feeds back into the ocean model in a realistic way,” Price says.

Similarly, while BISICLES runs on a rectangular structured grid (resembling graph paper), the PISCEES team is working to further fine-tune the code to better capture ice dynamics at the grounding line.

“One of the challenges we’re tackling now is to develop the AMR technique to more accurately represent the natural shape of the grounding line, rather than having to fit to the edges of our grid boxes like a stair-step pattern,” Ng says.

Given the rapid rate of development, the PISCEES team focuses on verifying and validating models – ensuring they run as expected and provide the correct answers.

“The code is changing every day, so we test it every night on at least one big DOE machine” at LANL, LBNL or Oak Ridge, Price says, and that access “is a huge factor in enabling us to develop project momentum.”

Along with support for verification and validation, PISCEES collaborators from the Institute for Sustained Performance, Energy and Resilience (SUPER, a SciDAC institute) regularly measure model performance to identify and correct any changes that retard the code. And in the past year, PISCEES collaborations with another SciDAC institute, QUEST (Quantification of Uncertainty in Extreme Scale Computations) have helped create a roadmap for developing the algorithms and software needed to quantify uncertainties in model outputs, like ice sheet mass loss and sea-level rise.

To a casual observer, PISCEES’ latest BISICLE visualization of ice velocity across the Antarctic continent is a remarkable doppelgänger of ice movement rates measured from remote sensing. “It’s a coarse validation,” Price says. Always with an eye to land ice’s complexity, he notes that the model’s still not ready for prime time as part of a fully functioning Earth system model.

“We haven’t quite gotten there yet,” he says, weighing his words. “But we’re close.”