Calibrated mass loss projections from the Greenland Ice Sheet

Data files are available at: https://arcticdata.io/data/10.18739/A2G737525/ The potential contribution of ice sheets remains the largest source of uncertainty in projecting sea-level due to the limited predictive skill of numerical ice sheet models, yet effective planning for coming sea level rise necessitates that predictions are credible and accompanied by a defensible assessment of uncertainty. Characterization of the likelihood of upper-end contributions are particularly important for developing adaptation strategies. While the use of large ensembles of simulations allows these kinds probabilistic assessments, there is no guarantee that simulations are aligned with observations. Here, we show that calibrating an ensemble of model simulations on observations reduces uncertainties in projecting 21st century mass loss from the Greenland Ice Sheet relative to a plausible a priori distribution of model configurations. We find that jointly conditioning on surface speeds and cumulative mass loss reduces the projected 2100 median contribution and 5--95th percentile by 16-30% and 38-56, respectively, compared to the un-calibrated ensemble, resulting in calibrated sea-level contributions ranging from 4 to 30 centimeters at the year 2100. This data set contains several products: - Surrogate model training data. ~1000 surface speed realizations in netCDF format, prepared with the Parallel Ice Sheet Model (PISM, www.pism.io) - Trained emulators. 50 trained emulators in HDF5 format, prepared with PyTorch (www.pytorch.org) - Posterior parameter distributions. 50 posterior distributions in CSV format. - Time series of projected mass change. Time series of projected mass change from 2008 until 2100 in CSV format, prepared with the Parallel Ice Sheet Model (PISM, www.pism.io), for both the ensemble using the Prior and the Posterior (calibrated) parameter distribution. 500 realizations for each Representative Concentration Pathway (RCP) scenario 2.6, 4.5, and 8.5.

数据集文件可通过以下链接获取：https://arcticdata.io/data/10.18739/A2G737525/。由于数值冰盖模型的预测能力有限，冰盖的潜在贡献仍是海平面上升预测中最大的不确定性来源；然而，针对未来海平面上升的有效规划要求预测结果具备可信度，并附带合理可证的不确定性评估。对高值端贡献的概率特征进行刻画，对于制定适应策略尤为关键。尽管利用大规模模拟集合可以开展这类概率评估，但无法保证模拟结果与观测数据一致。本研究表明，相较于合理的先验模型配置分布，基于观测数据校准的模型模拟集合能够降低格陵兰冰盖21世纪质量损失预测的不确定性。研究发现，相较于未校准的集合，同时结合地表速度与累积质量损失进行条件约束后，预测的2100年海平面贡献中位数以及5%~95%百分位区间分别降低16%~30%与38%~56%，最终校准后的2100年海平面贡献范围为4至30厘米。本数据集包含以下多类产物：
- 代理模型训练数据：采用并行冰盖模型（Parallel Ice Sheet Model, PISM，www.pism.io）生成的约1000组地表速度模拟样本，格式为netCDF。
- 训练好的模拟器：采用PyTorch（www.pytorch.org）构建的50个已训练模拟器，格式为HDF5。
- 后验参数分布：50组后验参数分布，格式为CSV。
- 预测质量变化时间序列：采用并行冰盖模型（Parallel Ice Sheet Model, PISM，www.pism.io）生成的2008年至2100年预测质量变化时间序列，涵盖基于先验参数分布与后验（校准）参数分布的两类集合；针对典型浓度路径（Representative Concentration Pathway, RCP）2.6、4.5及8.5三种情景，每种情景包含500组模拟样本。