High-performance Geoacoustic inversion cluster

高性能地声反演簇：沉积物层状结构、声速、密度、衰减和散射强度等海底地声参数对海洋声场的传播有很强的影响。因此，需要了解这些参数来理解和模拟浅水环境中的声音传播，这是许多声纳应用（以及岩土工程和地质目的）所需要的。从海洋声学测量中估计海底参数提供了一种方便的替代直接采样（例如取心）的原位方法，具有与声纳相关的灵敏度。然而，这需要解决一个本质上非唯一的强非线性逆问题（即，一系列解可以接受地拟合数据）。定量不确定性估计和模型选择（例如，确定适当的海底参数化）是有意义的反演分析的关键组成部分，但由于非线性问题需要大量的计算量而经常被忽视。我们正在开发一个贝叶斯推理公式和一个数值（统计）抽样方法，为未知参数建立概率分布，并根据数据和物理约束（先验信息）评估模型选择的证据。这种方法是严格的和完全非线性的，但计算密集，因此计算能力目前是我们研究的限制因素。我们的反演算法已经或正在开发中，以利用多个计算节点上的并行处理，具有低内存和节点间通信要求。基于多核处理器的新一代相对便宜、大规模并行的机器非常适合此类应用。获得一台大型并行计算机将使我们的研究充分发挥其潜力，应用于全场声学测量的地球声学反演、声学混响、海底反射率、环境噪声和船舶噪声，以及将反演方法转移到地球物理反演问题，如大地电磁、表面波色散和接收函数分析。

HIGH-PERFORMANCE GEOACOUSTIC INVERSION CLUSTER: Seabed geoacoustic parameters, such as sediment layering structure, sound speed, density, attenuation and scattering strength, strongly affect the propagation of ocean acoustic fields. Hence, knowledge of these parameters is needed to understand and model sound propagation in shallow-water environments, as required for many sonar applications (as well as geotechnical and geological purposes). Estimating seabed parameters from ocean acoustic measurements provides a convenient in situ alternative to direct sampling (e.g., coring), with a sensitivity relevant to sonar. However, this requires solving a strongly nonlinear inverse problem that is inherently nonunique (i.e., a range of solutions can acceptably fit the data). Quantitative uncertainty estimation and model selection (e.g., determining appropriate seabed parameterizations) are key components of meaningful inversion analysis, but are often neglected due to the high computational effort required for nonlinear problems. We are developing a Bayesian inference formulation and a numerical (statistical) sampling approach that builds up probability distributions for the unknown parameters and evaluates evidence for model selection based on the data and physical constraints (prior information). This approach is rigorous and fully nonlinear, but computationally intensive such that computing capacity currently represents the limiting factor in our research. Our inversion algorithms have been/are being developed to exploit parallel processing on multiple compute nodes, with low memory and inter-node communication requirements. The new generation of relatively inexpensive, massively-parallel machines, based on multi-core processors, are ideally suited to such applications. Acquisition of a massively parallel computer would allow our research to reach its full potential, with applications such as geaocoustic inversion of full-field acoustic measurements, acoustic reverberation, seabed reflectivity, ambient noise, and ship noise, as well as transfering the inversion methodologies to geophysical inverse problems such as magnetotellurics, surface-wave dispersion, and receiver function analysis.