CAREER: Accelerating Probabilistic Predictions of Sea-level Rise with Deep Learning

最近在预测地球冰川变化方面出现了一种新的范式，强调对冰的物理学和未来气候方面的未知数进行解释，以便构建一系列可信的未来。这种基于概率的界限对于为海平面上升、生态变化和其他气候反馈做准备至关重要，在这些方面，了解最佳和最坏情况具有实际重要性。这项任务在计算上是困难的：冰川模型是昂贵的，特别是当它可能需要几个月或几年的时间来运行成千上万的模拟，以完全描述可能的结果。然而，了解这种分布对于迎接建设可持续未来的巨大挑战至关重要。该项目将使用深度学习来构建替代模型，替代近似原始模型的冰盖模型组件，但计算成本要低得多。此外，该项目还将举办暑期学校，并制定教育模式，以整合工作的研究和教育部分。特别是，该项目将使用几何和生成式深度学习，结合经典技术在高性能计算中的新应用，构建一个比传统冰盖模型快数百倍的近似（或替代）模型。这样的加速反过来又将使研究人员能够探索前所未有的一系列未来场景，并应对冰建模以了解未来全球影响的挑战。此外，为期九天的暑期学校将汇集来自地球仪各地的教师和学生，学习和分享将深度学习应用于冰川学问题的有效方法。为了帮助高中生学习了解气候变化及其不确定性所需的技能，并使用计算机来应对相关的科学，经济和政策挑战，我们将开发针对高中生探索编程，数据，该项目由地球科学理事会共同资助，以支持人工智能和全球变化。ML在地球科学领域的进步。该奖项反映了NSF的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

A new paradigm has recently emerged in predicting changes to Earth’s glaciers that emphasizes accounting for unknowns with respect to ice physics and future climate in order to construct a range of plausible futures. Such probability-based bounds are essential in preparing for sea level rise, ecological changes, and other climate feedbacks in which understanding both best- and worst-case scenarios is of practical importance. This task is computationally difficult: glacier models are expensive, particularly when it can take months or years to run the many thousands of simulations required to completely characterize possible outcomes. However, understanding this distribution is essential in meeting the grand challenge of building a sustainable future. This project will use deep learning to construct surrogate models, replacements for ice sheet model components that approximate the original model, but which are much less computationally costly. In addition, this project will convene a summer school and develop educational models to integrate the research and educational components of the work. In particular, this project will use geometric and generative deep learning in tandem with novel applications of classic techniques in high-performance computing to build an approximate (or surrogate) model that is hundreds of times faster than traditional ice sheet models. Such a speedup will in turn allow researchers to explore an unprecedented range of future scenarios and meet the challenges of ice modeling for understanding future global impacts. In addition, a nine-day summer school will bring together instructors and students from around the globe to learn and share effective methods for applying deep learning to problems in glaciology. To help high schoolers learn the skills needed to understand climate change and its uncertainties and to use computers to tackle the associated scientific, economic, and policy challenges, we will develop educational modules tailored towards high school students exploring the intersection of programming, data, and global change.This project is co-funded by the Directorate for Geosciences to support AI/ML advancement in the geosciences.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.