ITR: A Prototype to Enable Near Real-time Environmental Characterization on the Grid

0312841马欣塔库马环境特征在实时预测环境事件和制定有效的环境保护、清理、风险评估和减灾战略方面正变得越来越重要。环境特征的一个基本组成部分是从稀疏的测量(例如，来自现场监测或卫星照片的数据)中推断信息的过程。例如，在根据从少量监测井收集的数据描述地下水污染时，理想的结果可能包括地下污染物浓度的三维图和污染物速度场的估计。如果地下特征或污染源的特征描述不佳，分析还可以确定孔隙度场或污染源的数量、位置和强度。根据观测数据解决未知系统特性(例如污染源位置和强度)的问题称为环境逆问题。在这些问题中，污染物的传输和化学由一个偏微分方程组(PDE)组成的数学模型来表示。求解未知问题特征是一个优化问题，需要对偏微分方程组进行成百上千次的求解。基于梯度的搜索技术代表了解决逆问题的最新技术。另一种选择是使用全局优化启发式算法，如遗传算法，以提供更稳健的决策空间搜索。一旦全局搜索识别出解空间的良好区域，就可以使用梯度或非梯度局部搜索技术进行微调。网格计算的出现(例如，NSF的新Tera网格)以及可用计算能力的相关增加产生了新的可能性。由于异质资源的复杂结构，将计算高效地映射到网格资源已被证明对现实应用程序具有极大的挑战性。这项提议的总体目标是研究能够容易地利用网格计算来解决环境表征问题的形式计算方法。为此，我们将开发一个支持网格的软件框架。本文将分别探讨基于MPI和Java的网格版本的两种迭代范例。该框架将应用于地下水和空气污染问题，这两个问题都具有首要的社会重要性。认识到即使在网格环境中，反问题在计算上也是具有挑战性的，因此将探索模型代理(例如，使用人工神经网络的统计近似)。此外，将研究通过人机交互来动态指导搜索过程的建模生成备选方案(MGA)，这是一种识别一组良好但非常不同的解决方案的技术。通过MGA生成的备选方案也将用于解决反问题中常见的非唯一性问题(即，多个特征可以匹配相同的测量数据)。更广泛的影响：将研究的几个核心组成部分整合到现有的高性能计算和工程系统分析的研究生和本科课程中，将促进发现和理解，同时促进教学。

0312841 Mahinthakumar Environmental characterization is becoming increasingly important in real-time forecasting of environmental events and in designing effective strategies for environmental protection, cleanup, risk assessment, and disaster mitigation. An essential component of environmental characterization is the process of deducing information from sparse measurements (e.g., data from in-situ monitors or satellite photos). For example, in characterization of groundwater pollution from data collected from a small number of monitoring wells, desirable outcomes may include a three-dimensional map of underground pollutant concentrations and an estimate of the pollutant velocity field. If the subsurface features or the source of contamination are poorly characterized, the analysis may also identify the porosity field or the number, location, and strength of pollutant sources. Problems in which unknown system characteristics (e.g., pollutant source location and strength) are resolved from observed data are known as environmental inverse problems. In these problems, pollutant transport and chemistry are represented by a mathematical model consisting of a system of partial differential equations (PDEs). Resolving the unknown problem characteristics is an optimization problem that requires the PDEs to be solved hundreds-to-thousands of times. Gradient-based search techniques represent the state-of-the-art for solving inverse problems. An alternative is to use global optimization heuristics, such as genetic algorithms, to provide a more robust search of the decision space. Once global search has identified good regions of the solution space, gradient or non-gradient local search techniques may be used for fine-tuning. The emergence of grid computing (e.g., NSF's new Tera Grid) and the associated increase in available computing power yield new possibilities. Efficient mapping of computations to grid resources has proven extremely challenging for realistic applications due to the complex fabric of heterogeneous resources. The overall goal of this proposal is to investigate formal computational approaches that can readily harness grid computing for the solution of environmental characterization problems. To this end, we will develop a grid-enabled software framework. Two iternative paradigms, based on the grid-enabled version of MPI and Java, espectively, will be explored. The framework will be applied to groundwater and air pollution problems, both of which are of prime societal importance. Acknowledging that, even within a grid environment, inverse problems are computationally challenging, model surrogates (e.g., statistical approximations using artificial neural networks) will be explored. Further, Modeling to Generate Alternatives (MGA), a technique that identifies a set of good yet very different solutions, will be investigated for dynamically steering the search process through human-computer interaction. Alternatives generated via MGA will be used also to address the commonly encountered non-uniqueness issue (i.e., where multiple characterizations can match the same measured data) in inverse problems. Broader Impact: Integration of several core components of research into existing graduate and undergraduate courses in high performance computing and engineering systems analysis will advance discovery and understanding while promoting teaching and learning.