08:30
SPUX - a Scalable Package for Bayesian Uncertainty Quantification
Jonas Sukys | Eawag: Swiss Federal Institute of Aquatic Science and Technology | Switzerland
Show details
Author:
Jonas Sukys | Eawag: Swiss Federal Institute of Aquatic Science and Technology | Switzerland
SPUX (Scalable Package for Uncertainty Quantification in "X") is a modular framework for Bayesian inference and uncertainty quantification.
The SPUX framework aims at harnessing high performance scientific computing to tackle complex aquatic dynamical systems rich in intrinsic epistemic uncertainties, such as ecological ecosystems, hydrological catchments, lake dynamics, subsurface flows, urban floods, etc. The challenging task of quantifying input, output and/or parameter uncertainties in such stochastic models is tackled using Bayesian inference techniques, where numerical sampling and filtering algorithms assimilate prior expert knowledge and available experimental data.
The SPUX framework greatly simplifies uncertainty quantification for realistic computationally costly models and provides an accessible, modular, portable, scalable, interpretable and reproducible scientific workflow. To achieve this, SPUX can be coupled to any serial or parallel model written in any programming language (e.g. Python, R, C/C++, Fortran, Java), can be installed either on a laptop or on a parallel cluster, and has built-in algorithmic and performance metrics and automatic reporting.
I will briefly present the key SPUX concepts to showcase a simple random walk example, and discuss recent results for a realistic hydrological model with uncertain parameters, rainfall input, and data observation processes.
08:50
Korali: a high-performance framework for Bayesian uncertainty quantification and optimization
Daniel Waelchli | ETH Zurich | Switzerland
Show details
Authors:
Daniel Waelchli | ETH Zurich | Switzerland
Sergio Martin | ETH Zurich | Switzerland
Georgios Arampatzis | ETH Zurich | Switzerland
Petros Koumoutsakos | ETH Zurich | Switzerland
We present Korali, a high-performance framework for Bayesian uncertainty quantification and optimization of engineering models. Korali implements a number of state-of-the-art algorithms in these fields, with particular emphasis on their optimization for massively parallel computing architectures.
In addition to performance at scale, Korali distinguishes from existing UQ software in the following: (i) it provides multi-language support for computational models programmed in C++, Python, or Fortran as well as stand-alone and legacy applications. (ii) it provides a language-independent interface for the description of statistical problems that allows users to solve a problem with different algorithms effortlessly, (iii) it is a modular, open-source software that is readily extensible. Scientists can build and test novel statistical methods, while benefiting from Korali's parallel engine.
We discuss key aspects of Korali's design and show examples in solving Bayesian Inference problems for parameter identification in molecular dynamics and micro-fluidics models. We also show how Korali integrates existing HPC software such as LAMMPS (CPU-Based) and Mirheo (GPU-Based) and scales on up to 2048 nodes (24'576 cores) of the CSCS Piz Daint supercomputer.
09:10
Randomized Low-rank Ensemble Kalman Filter for Nonlinear Networks
Yue Qiu | ShanghaiTech University | China
Show details
Authors:
Yue Qiu | ShanghaiTech University | China
Sara Grundel | Max Planck Institute for Dyanamics of Complex Technical Systems | Germany
In this talk, we consider the state estimation problem for nonlinear networks with random inputs using the ensemble Kalman filter (EnKF). Such a network is modeled by a nonlinear differential algebraic equation (DAE). We propose a randomized low-rank approach to approximate the state ensembles at each time step. The advantage of this randomized low-rank method are twofold. First, the number of forward model simulations at each time step is reduced from Ne to rk, where Ne is the size of ensembles for the standard EnKF, rk is the reduced rank, and Ne >> rk. Second, our randomized low-rank approach further reduces the computational cost of the EnKF update step. Numerical experiments show that the performance of our randomized low-rank EnKF is comparable with EnKF of an ensemble size Ne while the computational cost is reduced dramatically.
09:30
Robust Bayesian Inference under Limited Information and its Application to Atomic Spectra for Atmospheric Entry Systems
Anabel del Val | von Karman Institute for Fluid Dynamics | Belgium
Show details
Authors:
Tathagata Basu | Durham university | United Kingdom
Anabel del Val | von Karman Institute for Fluid Dynamics | Belgium
Jochen Einbeck | Durham university | United Kingdom
Matthias Troffaes | Durham university | United Kingdom
Thierry Magin | von Karman Institute for Fluid Dynamics | Belgium
Complex physical processes often need efficient uncertainty quantification techniques to extract knowledge from experimental data that can appropriately inform the proposed models. A physical model that accounts for uncertain measurements leads to an improvement in the design of potential engineering systems in terms of reliability. Bayesian inference is one of such uncertainty quantification techniques. However, Bayesian inference is based on expert elicitation and often provides different results based on the choice of priors. The different choices of priors can also produce degenerate results, especially when the available information is scarce and poses severe uncertainty. In this work, we use techniques derived from imprecise probability theory to derive a robust Bayesian inference approach. We use a set of priors for the prior specification, giving us a set of posteriors and a region of estimates, i.e. a lower and upper bound for our posterior expectation. We apply this framework to atomic spectra measurements where we infer temperatures and concentrations of a reactive mixture of gases in the general case of thermal non-equilibrium. This measurement is widely used for the validation of ablation models which play an important role in the design of spacecraft atmospheric entry systems.
09:50
- CANCELED - Hybrid Projection Methods for Large-scale Inverse Problems with Mixed Gaussian Priors
Jiahua Jiang | Virginia Tech | United States
Show details
Authors:
Jiahua Jiang | Virginia Tech | United States
Julianne Chung | Virginia Tech | United States
Taewon Cho | Virginia Tech | United States
Recent developments on generalized hybrid iterative projection methods have enabled the efficient computation of solution estimates for Bayesian linear inverse problems where it is impossible or undesirable to compute and store the prior covariance matrix. In this paper, we develop hybrid projection methods for inverse problems with mixed Gaussian priors where the prior covariance matrix is a convex combination of matrices and the mixing parameter is not known in advance. Such scenarios may arise when data is used to generate a sample prior covariance matrix (e.g., in data assimilation) or when the solution contains very different qualities and a linear combination of matrices is used for the prior covariance matrix. The proposed hybrid methods are based on the generalized Golub-Kahan bidiagonalization and can be used to efficiently compute regularized solutions. A distinctive feature of the proposed approach is that both the regularization parameter and the weighting parameter for the covariance matrix can be estimated during the iterative process. Furthermore, various data-driven covariance matrices (including learned covariance kernels) can be easily incorporated. Numerical examples from tomography demonstrate the potential for these methods.
10:10
- CANCELED - Bayesian Inference and Markov Chain Monte Carlo Sampling for Lagrangian Particle Tracking in the Ocean
Samah El Mohtar | King Abdullah University of Science and Technology | Saudi Arabia
Show details
Authors:
Samah El Mohtar | King Abdullah University of Science and Technology | Saudi Arabia
Omar Knio | King Abdullah University of Science and Technology | Saudi Arabia
Ibrahim Hoteit | King Abdullah University of Science and Technology | Saudi Arabia
Estimation of Lagrangian particle tracking (LPT) model parameters is challenging due to the irreversible nature of processes affecting the particles. In the ocean, several methods have been proposed to tackle this problem, but each has its limitations. In this work, we present a generic method that employs Bayesian inference and Markov chain Monte Carlo (MCMC) sampling to construct the probability distributions of parameters of interest. The forward model used within the MCMC machinery enables the use of existing application-specific modules built on top of LPT models. Furthermore, various types of observations can be considered depending on available data. This framework allows the method to serve different applications of LPT in the ocean. The method has been tested using a simple LPT advection-diffusion model in a double-gyre synthetic flow field. Different data types were tested, ranging from the simple location of particles, to measurements of concentration and contours of spills. Inference of the location of the source and time of release is presented as a probability distribution that quantifies uncertainties and describes correlation between parameters.