Abstract
Firn densification modelling is key to understanding ice sheet mass balance, ice sheet surface elevation change, and the age difference between ice and the air in enclosed air bubbles. This has resulted in the development of many firn models, all relying to a certain degree on parameter calibration against observed data. We present a novel Bayesian calibration method for these parameters and apply it to three existing firn models. Using an extensive dataset of firn cores from Greenland and Antarctica, we reach optimal parameter estimates applicable to both ice sheets. We then use these to simulate firn density and evaluate against independent observations. Our simulations show a significant decrease (24 % and 56 %) in observation–model discrepancy for two models and a smaller increase (15 %) for the third. As opposed to current methods, the Bayesian framework allows for robust uncertainty analysis related to parameter values. Based on our results, we review some inherent model assumptions and demonstrate how firn model choice and uncertainties in parameter values cause spread in key model outputs.
Introduction: The Firn Modeling Challenge
On ice sheets, snow gradually compacts into solid ice, passing through an intermediate, porous stage called firn. Accurately modeling the rate of firn densification is critical for glaciology. It allows researchers to correctly interpret satellite altimetry data—distinguishing between actual ice mass loss and mere changes in firn volume—and to accurately date trapped air bubbles in ice cores for paleoclimate studies.
Historically, firn densification models have relied on simplified empirical equations with parameters tuned to limited, location-specific datasets (e.g., calibrated only for Greenland or only for Antarctica). Furthermore, these deterministic calibrations lack a rigorous way to quantify parameter uncertainty, leaving researchers in the dark about how much to trust model predictions under varying climatic conditions.
Methods: A Bayesian Approach
The authors tackle model calibration as an inverse problem using a Bayesian framework. Instead of seeking a single "best" value for each model parameter, they treat parameters as random variables and use a Markov chain Monte Carlo (MCMC) algorithm—specifically the Random Walk Metropolis (RWM)—to approximate their posterior probability distributions.
They apply this to three widely used benchmark firn models: Herron and Langway (HL), Arthern et al. (Ar), and Li and Zwally (LZ). This approach yields both the most likely parameter set (Maximum A Posteriori, or MAP) and a 95% credible interval. To prevent overfitting to perfect inputs, the authors injected realistic, normally distributed noise into the climate forcing data and surface snow density during the calibration process.
Datasets and Evaluation
The study utilizes a comprehensive dataset of 91 observed firn depth–density profiles, with 27 from the Greenland Ice Sheet and 64 from the Antarctic Ice Sheet. Of these, 69 cores were used for calibration, while 22 were held out for independent evaluation.
The primary evaluation metric is Depth-Integrated Porosity (DIP), representing the total air content in the firn column. This was measured for the top 15 meters (DIP15) and from 15 meters down to the pore close-off depth (DIPpc). Climate forcing data, including temperature and accumulation rates, were provided by the RACMO2.3p2 regional climate model.
Key Insights and Results
The Bayesian calibration revealed several critical insights:
- Non-linear response to snow load: The original HL and Ar models assumed that firn densification rates scale linearly with the rate of new snow accumulation. The Bayesian posterior distributions strongly rejected this, finding exponents significantly less than 1. This suggests that as firn compacts, the contact area between ice grains increases, making it more resistant to further compaction.
- Accuracy Gains: Using the newly calibrated MAP parameters, the HL model's Root Mean Square Error (RMSE) dropped by 24% for DIP15 and 22% for DIPpc. The Ar model saw even more dramatic improvements: a 45% reduction in DIP15 RMSE and a 61% reduction in DIPpc RMSE.
- Uncertainty Quantification: Parameter uncertainty leads to a ~6–7.5% coefficient of variation in key outputs. However, inter-model uncertainty (structural differences between models) is much larger, jumping to 19.5% for compaction anomalies.
Limitations and Future Directions
While the Bayesian calibration vastly improves parameter estimates, the models' 95% credible intervals still frequently fail to capture exact observed firn densities. This indicates fundamental flaws in the governing physical equations, such as an inability to capture seasonal temperature fluctuations. The authors advocate for developing more parsimonious, physically-based models and emphasize the need for more diverse field data to better constrain the micro-physics of firn compaction.
Summary updated