Estimating probability distributions from complex models with bifurcations: The case of ocean circulation collapse

Studying the uncertainty in computationally expensive models has required the development of specialized methods, including alternative sampling techniques and response surface approaches. However, existing techniques for response surface development break down when the model being studied exhibits discontinuities or bifurcations. One uncertain variable that exhibits this behavior is the thermohaline circulation (THC) as modeled in three dimensional general circulation models. This is a critical uncertainty for climate change policy studies. We investigate the development of a response surface for studying uncertainty in THC using the Deterministic Equivalent Modeling Method, a stochastic technique using expansions in orthogonal polynomials. We show that this approach is unable to reasonably approximate the model response. We demonstrate an alternative representation that accurately simulates the model's response, using a basis function with properties similar to the model's response over the uncertain parameter space. This indicates useful directions for future methodological improvements.