Surrogate Models to Accelerate High-Fidelity Physics Based Simulation

Simulations are designed to reproduce the behavior of end-to-end systems or their critical components. When designed under a stochastic framework, they enable us to identify ensembles of possible system evolutions and their associated uncertainties and sensitivities of outputs relative to the inputs. When the intent is to reproduce sensor observations they need to account for the sensor and data processing limitations as well. It may not be feasible to emulate an end-to-end system using existing physics-based computational models when these processes are too complex or strongly nonlinear or if the physics-based models impose too great a computational burden. Further, it may be challenging to obtain enough data to train machine learning (ML) surrogate models to sufficient accuracy. To address these limitations, we propose to develop hybrid (deterministic and stochastic) models in which we reduce complexity and, at the same time, manage the error/uncertainty in our modeling. using error subspaces defined by deviations of the evolving dynamical system from modes of our stochastic surrogate models. Tracking this error subspace enables us to constrain it where the uncertainty may be growing. In addition, when emulating a system evolution for which data has been collected, such data can be assimilated efficiently into the dynamically orthogonal subspace to improve the skill of the underlying models. Such information sources are efficiently incorporated using principled Bayesian data assimilation in the subspace. Existing commercial EO-IR-Radar scene generation systems use high-level tools such as UNREAL or UNITY, developed to produce 4D gaming environments, to define interactive time-evolving scenarios in terms of frame-by-frame wire mesh descriptions, in which all surfaces are represented by facets. Facets are assigned materials with reflectivity properties. Objects are assigned temperatures so that IR signatures can be formed for them. Sensor observations are synthesized by ray tracing all light, IR, and Radar sources from source to sensor aperture. For scenarios as complicated as missile defense, where multiple missiles are incoming, with some being exploded or incinerated, and the debris and gaseous plumes serve to obstruct the observations of additional still-potent missiles, we anticipate that surrogate models will be needed to synthesize realistic physically accurate sensor observations in a real-time hardware-in-the-loop configuration. We anticipate that our proposed surrogate models will be integrated into and replace selected components of the rendering stage provided by such commercial scene generation systems. Approved for Public Release | 22-MDA-11339 (13 Dec 22)