Consistent uncertainty reduction in modeling nonlinear systems

We present a fairly general time-independent framework for performing systematic and reliable uncertainty analysis of computer-implemented models of complex nonlinear systems. Within this framework, we provide the first formal proof of uncertainty reduction in system parameters and responses, which is achieved by consistently combining model-predicted responses and their associated uncertainties with experimental (e.g., sensor-based) information. MODTRAN - a very large, complex code that models optical radiation transport in the atmosphere - provides an excellent example to illustrate the approach. The sensitivities needed to propagate uncertainties from inputs and parameters to outputs through the complex chain of modules are calculated by automated differentiation.