Predictive Uncertainty in Simulation
May 4, 2015
- Dr. William Oberkampf
- Consulting Engineer
- Holden Auditorium (Room 112)
- 4:00 p.m.
- Faculty Hosts: Dr. Christopher Roy and Dr. Pradeep Raj
Simulation is becoming the primary tool in predicting the performance, reliability, and safety of engineered systems. To many managers, decision makers and policy makers not trained in modeling and simulation, these simulations can appear most convincing with their captivating video graphics. Terminology such as “virtual prototyping,” “virtual testing,” and “full physics simulation” are extremely appealing when budgets and schedules are highly constrained, or when competitive pressures force project managers to move forward with little or no testing of subsystems or systems. Many contend that higher fidelity physics modeling, combined with faster computers, is the path forward for improved decision making based on simulation. I argue that improved predictive uncertainty is the most constructive path forward. Predictive uncertainty estimation is the emerging field attempting to capture all aspects of uncertainty in the simulation of a system. The tradition in uncertainty estimation is to focus on propagation of input uncertainties through a mathematical model to obtain uncertainty in the system response quantities of interest. In contrast, predictive uncertainty attempts to capture this uncertainty as well as all potential sources of uncertainty. These include numerical solution error, model form uncertainty, and uncertainty in the environments and scenarios to which the system could be exposed, either intentionally or unintentionally. This talk will briefly review the traditional uncertainty quantification approaches that have been developed in fields such as nuclear power reactor safety. An important distinction is made between uncertainties that are random or stochastic (aleatory uncertainties) and those that are due to lack of knowledge (epistemic uncertainties). It is argued that imprecise probability approaches are the most appropriate method to represent the total predictive uncertainty in simulation-based decision-making.
Biography:
Dr. William L. Oberkampf received his PhD in 1970. He has 44 years of experience in research and development in computational mechanics. He was on the faculty at the University of Texas at Austin until 1979. From 1979 until 2007 he worked at Sandia National Laboratories in both staff and management positions. During his career he has been deeply involved in computational and experimental research and development in fluid dynamics and heat transfer. During the last 20 years he has been focused on verification, validation, uncertainty quantification, and risk analyses of high consequence systems. He is a Fellow of the American Institute of Aeronautics and Astronautics. He has over 178 journal articles, book chapters, conference papers, and reports, and has taught 44 short courses in the field of VVUQ. He and Prof. Chris Roy co-authored the book "Verification and Validation in Scientific Computing" published by Cambridge University Press.