Collaborative Research: From Adjoints for the Few to Adjoints for the Many: Integrating the Use of Adjoint Methods in Earth System Modeling

伴随模型对于需要状态估计（例如预测或预测）和模型输出相对于模型输入的灵敏度的研究是有用的数学工具。伴随方法允许有效和全面的计算模型的灵敏度非常高维空间的输入。伴随模式可用于有效地量化表面风、浮力通量或海洋混合分布的时空分布变化对南大洋碳吸收等数量的影响。为给定的最先进的地球物理模型开发伴随模型可能与开发地球物理模型本身一样复杂，这是伴随不经常使用的原因之一。 该项目将推进实施伴随模型所需的工具，并将开展三个相互关联的工作包，以克服这些限制：（1）使开源AD工具OpenAD在气候研究和教育中使用的四个开源地球系统建模框架上完全成熟（2）围绕几个科学应用开发配置，其中感兴趣的量的敏感性（成本函数）到输入（控制变量）可以被配置用于实际的运行时，（3）开发一个通用的基于消息传递接口（MPI）的框架，以实现高分辨率前向和低分辨率伴随模型之间的耦合。拟议的地球科学研究将引起计算科学家和气候科学家的兴趣。 这项研究的更广泛影响将使伴随方法更容易在地球科学中使用，并有可能在广泛的地球科学和其他学科中进行数据同化和敏感性分析。

Adjoint models are useful mathematical tools for studies that require state estimates (e.g. forecasts or projections), and sensitivities of model output with respect to model input. Adjoint methods allow efficient and comprehensive computation of model sensitivity to very high dimensional spaces of inputs. An adjoint model can be used to efficiently quantify the effects of changes in the spatio-temporal distribution of surface winds, buoyancy fluxes, or mixing distribution in the ocean, on quantities such as carbon uptake in the Southern Ocean. Developing adjoint models for given state-of-the-art geophysical models can be as complex as developing the geophysical models themselves, which is one reason why adjoints are not frequently used. This project will advance the tools needed to implement adjoint models.The project will pursue three inter-related work packages to overcome these limitations:(1) bringing the open-source AD tool OpenAD to full maturity on four open-source Earth system modeling frameworks used in climate research and education(2) developing configurations around several science applications where the sensitivity of quantities of interest (cost functions) to inputs (control variables) can be configured for realistic runtimes,(3) developing a generic Message Passing Interface (MPI)-based framework to enable the coupling between high-resolution forward and low-resolution adjoint models. The proposed geoscience studies will be of interest to computational and climate scientists. A broader impact of this research will be to enable adjoint methods to be more easily used across the geosciences, with potential to be transformative for data assimilation and sensitivity analysis in a wide range of geoscience and other disciplines.