Hybrid Computational Cluster for Ocean Acoustic Inversion

海底属性和海洋声源的声学遥感是地质、岩土工程、海洋哺乳动物和国防(声纳)应用的重要问题。然而，这些应用需要建立和求解非线性逆问题。虽然由于计算复杂而常常被忽视，但定量不确定性估计和模型选择(例如，确定海底沉积物层数或根据数据发出海洋哺乳动物的声音)是有意义的反演分析的关键组成部分。我们正在为海洋声学和地震反演中的模型选择和不确定性估计开发先进的算法。这些算法对后验概率密度(给定观测数据和先验信息的模型参数的概率分布)进行数值采样，并包括对模型选择的概率采样。虽然我们的算法是通用的和强大的，但它们是计算密集型的，因此我们目前在中央处理单元(CPU)集群上使用多个并行采样器的工作是非常有限的。然而，最近图形处理单元(GPU)在计算问题中的应用有可能使非线性(数值)求逆发生革命性变化。特别是，用于计算数据预测的GPU编码的细粒度并行性补充了反演中多个采样器的粗粒度(CPU)并行性；将这些并行化在一个集成的混合计算设施中代表了非常适合我们的非线性推理方法的新型硬件。收购混合GPU/CPU系统将使我们的研究能够充分发挥其潜力，应用于大型和顺序数据集、全波反演问题、在鲜为人知的环境中定位未知数量的源，以及在二维(垂直和横向)变化的海底/地球模型的层析反演中。通过与加拿大地质调查局、加拿大海洋网络、当地海洋技术行业和国际机构的持续合作，为这些应用提供了声学/地震数据和实验。

Acoustic remote sensing of seabed properties and of sources of sound in the ocean are important problems with geological, geotechnical, marine-mammal, and defence (sonar) applications. However, these applications require formulating and solving nonlinear inverse problems. Although often neglected due to computational complexity, quantitative uncertainty estimation and model selection (e.g., determining the number of seabed sediment layers or vocalizing marine mammals consistent with the data) are key components of meaningful inversion analysis. We are developing advanced algorithms for model selection and uncertainty estimation in ocean acoustic and seismic inversion. These algorithms numerically sample the posterior probability density (the probability distribution of the model parameters given the observed data and prior information), and include probabilistic sampling over the choice of model. While our algorithms are general and powerful, they are computationally intensive such that our current work, which employs multiple parallel samplers on a central-processing unit (CPU) cluster, is strongly compute-limited. However, the recent application of graphics processing units (GPUs) in computational problems has the potential to revolutionize nonlinear (numerical) inversion. In particular, the fine-grained parallelism of GPU coding for computational data predictions complements the coarse-grained (CPU) parallelism of multiple samplers in the inversion; combining these in an integrated hybrid computing facility represents novel hardware ideally suited to our nonlinear inference approaches. Acquisition of a hybrid GPU/CPU system would allow our research to reach its full potential, with applications to large and sequential data sets, full-wave inversion problems, localizing an unknown number of sources in a poorly-known environment, and in tomographic inversion for seabed/earth models that vary in two dimensions (vertically and laterally). Acoustic/seismic data and experiments for these applications are facilitated through ongoing collaborations with the Geological Survey of Canada, Ocean Networks Canada, local marine-technology industry, and international institutions.