Combining Data and Simulation to Predict the
Behavior of Complex Systems
Introduction
Predicting the behavior of complex physical systems is a key requirement for many important DOE
applications. However, incomplete physical models and uncertain parameters limit the predictive
capability of even high fidelity simulations. Model and parameter identification are typically
done by finding models and parameter sets that accurately predict the results of controlled
experiments. This often involves a combination of computation and optimization.
|
Senior Investigators
|
Project Goal
The goal of this project is to develop new tools that allow us to use data from
multiple experiments of different kinds in conjunction with simulation.
It has become clear that traditional simulation and sampling methods do not suffice.
The constraints that experiments put on parameters are ill conditioned and complex.
Simulating highly nonlinear experiments is so slow that huge numbers of samples are
not practical.
Measurements often reflect a complex, uncertain relationship between the system state and
measured quantities. For example, laser diagnostic measurements
of a chemical species in a reacting flow are modulated by state-dependent quenching corrections.
Observations from cosmological surveys require modeling of photometric redshifts, corrections for telescope focal plane distortions, and a number of other systematic effects.
In some cases, the combination of imprecisely defined experiments and complex dynamics
means that it is appropriate to consider only
statistical properties of the experiments, not a detailed trajectory.
|
Approach
Our overall approach is based on Monte Carlo sampling within a Bayesian framework.
Monte Carlo (MC) sampling provides a
numerically sound implementation that respects the strong nonlinearity in the target applications.
Our hierarchical approach in this project is based on a combination of implicit sampling (IS),
Markov chain MC (MCMC), and reduced-order models (ROMs), with a focus on application to complex,
high dimensional problems.
In Implicit Sampling, sample generation is guided by the solution of an optimization problem
that is based on the target model; this helps to identify regions of high probability with
respect to the posterior (see Implicit Sampling inset below).
For MCMC, we use techniques based on parallel marginalization to reduce correlation time in the
sampling process, and use an affine-invariant sampling algorithm to improve the
overall efficiency when the parameters span orders of magnitude in scale and variation.
Reduced order models will play an increasingly important role in building a hierarchy of
targets, particularly when the targets span a large range of computational complexities,
and when observables are complex functions of the simulated state variables.
|
Implicit Sampling
|
|
A "Prior" distribution for a parameter is one that is consistent with previous knowledge.
The "Likelihood" represents the parameter distribution that is consistent with the new information.
Sampling will create the "Posterior" distribution that will be consistent with all the data.
In the (desireable) case that the new data strongly informs the parameters, there are several
conditions that make sampling the posterior a difficult task.
|
In (a), the prior and likelihood are peaked and well-separated. A set of samples
based on the prior distribution will rarely result in a good fit to the new data.
Similarly, a uniformly distributed sample set (b) will generate a high
fraction of unlikely samples. With enough samples, the posterior in either case will
eventually be characterized, but with catastrophically poor efficiency.
Implicit Sampling rather proposes a distribution of samples based on a local model of
the posterior (c). Ideally, the local model is centered near the peak value of the likelihood
and is just broad enough to sample the structure of the distribution without wasting
effort sampling regions of low probability. However, this added efficiency is not free.
IS requires that one be able to predict a good generic model for the posterior, and then be able
solve the associated constrained optimization problem to set the parameters of the fit.
We are currently exploring the feasibility of such an approach for realistic applications,
by combining IS with the other approaches discussed above into an iterative adaptive
sampling framework.
|
|
As an example, combustion chemistry models are constructed and optimized with respect to
available experimental data; typically including volumetric reactors (such as calormetric
bombs, flow reactors, shock tubes), and counterflowing flame systems -- all of which are
low-dimensional idealized configurations providing fairly well-characterized data for limited
parameter ranges. However the resulting "validated" combustion models are then
applied to a wide range of systems, from zero-D and steady systems
to fully turbulent combustion environments. Historically, there has been very little feedback from
the more practical combustion systems to the improvements of the underlying models.
Ultimately, we would like to combine all of these technologies into an extensivle
hierarchical adaptive sampling scheme for high-dimensional parameter space, where we
can directly exploit the wide range of computational expense associated with zero-D
systems through turbulent 3D reacting flow systems. The ultimate goal is to create the
methodology to incorporate increasingly complex, realistic and relevant experimentally
observable data into the development and improvement of the fundamental models that
underpin the computational predictions of complex multiphysics systems.
|
Applications
Combustion
Cosmology
Transport in thin-film polymers
Geosciences
|
|