Introduction
Firn densification models (FDMs) are essential tools in glaciology, describing the transformation of snow into glacial ice. This process is critical for interpreting satellite altimetry data; changes in ice sheet surface height must be partitioned between actual mass changes and the compaction of the firn layer. Historically, FDMs have relied on empirical parameters tuned to specific sites, which limits their transferability across diverse climatic conditions. This study addresses these limitations by applying a rigorous Bayesian framework to calibrate three widely used FDMs—Herron-Langway (HL), Arthern (Ar), and Li-Zwally (LZ)—against a global dataset.
Methodology
The researchers employed a formal Bayesian approach to quantify parameter uncertainties and identify complex correlations.
- Bayesian Calibration: The Random Walk Metropolis (RWM) algorithm, a variant of Markov Chain Monte Carlo (MCMC), was used to explore the posterior distribution of model parameters.
- Likelihood Function: Calibration was based on Depth-Integrated Porosity (DIP). To capture different densification stages, the data was split into shallow ($DIP_{15}$, top 15m) and deep ($DIP_{pc}$) components. To account for error propagation with depth, the authors assigned a higher uncertainty (10% standard deviation) to deeper measurements compared to shallower ones (5%).
- Convergence and Validation: Chain stability was monitored using trace plots and the Gelman-Rubin statistic ($R < 1.1$). Residual analysis confirmed that model errors follow a Gaussian distribution, validating the use of a Gaussian likelihood function.
Dataset and Calibration Strategy
The study utilizes a comprehensive dataset of 91 firn core sites. These sites were divided into a calibration set (69 sites) and an evaluation set (22 sites). The selection process was stratified to ensure the evaluation set accurately reflected the broad range of temperatures and accumulation rates found in both the Greenland (GrIS) and Antarctic (AIS) ice sheets.
Key Insights: The Equifinality Problem
A primary finding of the Bayesian analysis is the presence of "equifinality"—where multiple parameter combinations yield similar model outputs. This is driven by strong correlations between physical parameters. For instance, in the HL model, the pre-exponential factors ($k_0^*, k_1^*$) and their respective activation energies ($E_0, E_1$) exhibit correlation coefficients as high as 0.94.
These correlations suggest physical trade-offs within the models. For example, a negative correlation between activation energy and accumulation rate exponents indicates that in warmer, high-accumulation regions, the model must balance accelerated burial with increased thermal densification rates.
Results and Implications
The calibrated models provide a more robust basis for calculating "compaction anomalies"—deviations from steady-state densification. Across 27 Greenland sites, these anomalies were quantified between 0.1m and 1.1m. Such values are vital for correcting satellite measurements of ice sheet height, ensuring that mass balance estimates are not skewed by firn dynamics.
Limitations and Future Directions
While the Bayesian approach significantly improves parameter constraint and error quantification, challenges remain. The high degree of parameter correlation, particularly in the LZ model, complicates the physical interpretation of individual variables. Furthermore, the current study assumes independence between shallow and deep porosity errors. Future work should explore more complex covariance structures and extend this framework to physics-based models that incorporate processes like grain growth and meltwater percolation.
Summary updated