Overview
surmise is a Python library for a modular calibration framework. Bayesian framework allows reducing computer code predictive uncertainty by calibrating parameters directly to the observational data.
The typical Bayesian workflow consists of three main steps:
determination of a prior distribution to express the the background knowledge or expert opinion,
identifying the likelihood function with the observational data,
combining both the prior distribution and the likelihood function using Bayes’ theorem in the form of the posterior distribution.
However, implementing Bayesian calibration for high-dimensional data and complex systems is very challenging. Because the posterior distribution often cannot be obtained analytically, an approximate Bayesian inference should be used. The most popular method for sampling from high-dimensional posterior distributions is Markov chain Monte Carlo (MCMC). However, for even relatively fast computer models implementation of Bayesian inference with MCMC would simply take too long. A very fast approximation to the system code is thus required to use the Bayesian approach. Surrogate models (or emulators) emulate the behavior of the input/output relationship of the computer model and are computationally inexpensive to allow MCMC sampling to be possible.
The purpose of surmise is to provide a modular model calibration software that can input model observations and output draws.
surmise’s work is categorized into three routines:
emulation: Carries out Bayesian emulation of computer model output and generates inputs to
calibrationcalibration: Generates estimates of the calibration parameters based on field observations of the real process and an output from
emulationutilities: Performs different utility tasks such as a sampler (e.g., Metropolis-Hastings) to generate posterior draws of calibration parameter
Examples of how to use emulation, calibration, utilities modules can be found in
the examples/ directory.