Development of Automation Tools for Automatic Differentiation of Computer Codes for Sensitivity & Uncertainty Analysis

As the DoD increases its reliance on modeling and simulation (M&S) via large and complex computer codes, a major new problem emerges: current M&S codes do not provide an efficient method of estimating simulation sensitivities and uncertainties for given uncertainties in the input parameters themselves. Automatic Differentiation (AD) is an emerging technology that has proven itself in many disciplines as an efficient means of providing sensitivity and uncertainty analysis for very large and complex computer codes. One of the greatest drawbacks of using these AD tools is the difficulty in applying them to complex computer codes. Present AD tools are pre-compilers that only create enhanced code (i.e., containing derivatives) but do not provide access to the new information or compute uncertainties. The derivatives codes produce an enormous volume of new data and the developer faces the daunting task of accessing the I/O variables, their names, physical units, computing uncertainties, etc. OKSI has established a framework to overcome these obstacles thereby making "Automatic" Differentiation truly "automatic." This framework is called Automated Automatic Differentiation (A2D). OKSI will apply the A2D capability to complex codes, such as those comprising the Battlespace Environment Signature Toolkit, to produce user-friendly enhanced codes with full access to all the sensitivity and uncertainty data. The A2D tools will be demonstrated under Phase-I and applied to a suite of simulation codes under Phase-II.