Addressing deep uncertainty in space system development through model-based adaptive design

Abstract

When developing a space system, many properties of the design space are initially unknown and are discovered during the development process. Therefore, the problem exhibits deep uncertainty. Deep uncertainty refers to the condition where the full range of outcomes of a decision is not knowable. A key strategy to mitigate deep uncertainty is to update decisions when new information is learned. NASA's current uncertainty management processes do not emphasize revisiting decisions and therefore are vulnerable to deep uncertainty. Examples from the development of the James Webb Space Telescope are provided to illustrate these vulnerabilities. In this research, the spacecraft development problem is modeled as a dynamic, chance-constrained, stochastic optimization problem. The Model-based Adaptive Design under Uncertainty (MADU) framework is introduced, in which conflict-directed search is combined with reuse of conflicts to solve the problem efficiently.

The framework is built within a Model-based Systems Engineering (MBSE) paradigm in which a SysML model contains the design and conflicts identified during search. Changes between problems can involve the addition or removal a design variable, expansion or contraction of the domain of a design variable, addition or removal of constraints, or changes to the objective function. These changes are processed to determine their effect on the set of known conflicts. Using Python, an optimization problem is composed from information in the SysML model, including conflicts from past problems, and is solved using IBM ILOG CP Optimizer. The framework is tested on a case study drawn from the thermal design of the REgolith X-ray Imaging Spectrometer (REXIS) instrument and a case study based on the Starshade exoplanet direct imaging mission concept which is sizeable at 35 design variables, 40 constraints, and 10¹⁰ possible solutions.

In these case studies, the MADU framework performs 58% faster on average than an algorithm that doesn't reuse information. Adding a requirement or changing the objective function are particularly efficient types of changes. With this framework, designers can more efficiently explore the design space and perform updates to a design when new information is learned.