PSimPy: GP emulation-based sensitivity analysis, uncertainty quantification and calibration of landslide simulators

Abstract

Computer simulations are widely used to study real-world systems in many fields of science and engineering, such as earth science, life science, energy engineering, civil engineering, etc. Such simulators may be subject to a variety of uncertainties resulting from many sources, e.g. uncertain input parameters. These uncertainties need to be properly quantified in order to achieve reliable simulation-based prediction and design. However, simulators are often computationally too expensive for uncertainty-related analyses, such as uncertainty quantification, global sensitivity analyses, and parameter calibration. In the recent decades, Gaussian process (GP) emulation has been shown to be effective in overcoming computational bottlenecks.

This work presents the newly developed open-source Python package, PSimPy, for sensitivity analysis, uncertainty quantification, and parameter calibration of simulators using GP emulation. It is built upon recent progress in GP emulation for simulators with massive outputs, GP emulation-enabled global sensitivity analyses, and Bayesian active learning for parameter calibration. The structure of PSimPy is presented herein, and case studies from landslide run-out modelling are performed to demonstrate the feasibility of PSimPy for the above mentioned computationally costly tasks. Due to the data-driven nature of GP emulation, PSimPy is potentially applicable to computationally expensive simulators in many fields.