Regression and Bayesian uncertainty analysis are two different
ways to quantify parametric and predictive uncertainties. While the
regression confidence intervals are computationally frugal and easy
to compute, the computationally expensive Bayesian credible intervals
are more accurate for highly nonlinear problems with local
minima. The ability of UCODE_2014 to calculate both kinds of intervals
provides great flexibility in evaluating parametric and predictive
uncertainties.
This work integrates the advanced MCMC algorithm, DREAM,
into UCODE_2014 for Bayesian uncertainty analysis. This enables
UCODE_2014 to efficiently handle high-dimensional and multimodal
problems while taking advantage of UCODE structures to
obtain straightforward and flexible execution. With the template
and instruction files of UCODE, the uncertainty analysis can be
assessed for any process models with ASCII-based inputs and outputs
without modifying source code. With the parallel computing
capability based on UCODE's robust dispatcher-runner protocol, the
Bayesian uncertainty analysis can be evaluated with relatively high
efficiency. In addition, UCODE_2014 provides versatile ways to
initialize the MCMC process and a proper selection can accelerate
the chain convergence and save computational time. It also provides
a variety of parameter prior distributions which makes its
application more flexible. Most importantly, the program can be
used for not only the parametric but also the predictive uncertainty
analysis with little alteration of the input files.
This work uses a groundwater reactive transport model to
illustrate that UCODE_2014 can be used for Bayesian uncertainty
analysis of a variety of environmental problems. When applied to
large-scale inverse problems, e.g., highly heterogeneous problems,
MCMC methods face two major challenges. First, the complex process
models may make MCMC simulation computationally
unaffordable to evaluate posterior probability density for any
parameter sample. In addition, the high-dimensional parameter
spaces make the exploration of the posterior parameter distribution
difficult sometimes even prohibitive. To resolve the first challenge,
currently there are two major strategies: (1) develop advanced
MCMC algorithms to improve computationally efficiency by
reducing the needed number of process model executions, i.e., MTDREAMzs
(Laloy and Vrugt, 2012); and (2) apply model surrogate
methods to improve computationally efficiency by evaluating the
computationally frugal surrogate model instead of the actual process
model, i.e., Zhang et al. (2013) and Laloy et al. (2013). The second
challenge is more difficult because the high dimension usually
means large model nonlinearity and many local minima. For those
problems, currently none of the MCMC methods can guarantee that
all local minima will be found, and surrogate methods are no longer
absolutely computationally competitive because building the surrogate
model can require a large number of process model runs.
Combining the advanced MCMC algorithm with some dimensionality
reduction techniques may be a good solution.