|
Objective:
Traditional data mining extracts knowledge from static data sets. Simulators offer a richer exploratory environment, with the possibility of generating new data in order to verify patterns and test theories. New tools are needed for deciding which simulations to run and how to transform the output -- possibly terabytes of data -- into knowledge. This research task will develop automated or semi-automated techniques for intelligently sampling a parameter space to best clarify a physical model and its behavior. This application of active, closed-loop machine learning can reduce the hundreds or thousands of computationally intensive numerical simulations required to investigate science problems, helping experimenters maximize knowledge return from their simulation studies. The challenge is to develop a control technique that learns from simulation results -- via an output metric, and with allowance for potentially noisy or chaotic simulations -- to select the next simulation to run in order to maximize the information likely to be gained. Challenges include landscape characterization (determining which conditions lead to a given behavior), model identification, simulator control, and detection and use of trigger events (e.g., to allow backtracking during a simulation or to enable object-centered indexing). The driving application for this research task will be a simulation of asteroid collisions to determine conditions that lead to satellite formation.
Applications:
Numerical and particle simulation studies; earth science (core/mantle, climate, atmospheric, and ocean dynamics); space science (stellar dynamics, solar wind; galaxy and planet formation); engineering design (aerodynamics, structures, propulsion, failure analysis).
NASA Benefit:
Simulators play a fundamental role in investigations by scientists and engineers across NASA, DOE, DOD, FAA, industry, and academia. In many cases, they enable studies that would be infeasible or impossible otherwise. Science examples include studies of the earth's core and mantle dynamics; climate prediction; atmospheric and ocean dynamics; fluid flows in microgravity environments; dynamics of the interiors of stars; interaction of the solar wind with the earth; galaxy and planet system formation; artificial life; and neuronal models. Engineering examples include aerodynamics and flight research; propulsion systems; behavior of flexible structures; and failure analysis. Much of the work in high-performance computing has focused on producing larger, more accurate simulations. The complementary problem of deciding which simulations to run and how to transform the output into knowledge has largely been neglected, despite a large potential payoff. (The simulation process may take months, but the human analysis phase may take years.) This research task will develop machine learning techniques for efficient use by large-scale numerical simulators. Active learning can greatly reduce the number of simulation runs that must be done, by helping to predict which parts of an input parameter space need to be explored. Investigators will focus on particle simulations, a domain of interest to NASA for studying the origins of the planets, the long-term dynamical behavior of comets, Kuiper Belt objects, and other Solar System bodies.
Keywords:
exploratory model analysis, knowledge discovery, numerical simulator control, particle simulation
Images:
PI slides.
|
