Physical Reliability of Electronics Based on Surrogate Modeling Methods

The thermal simulation of electronics modules have long been used to accelerate the design process by reducing the requirement of building and testing physical prototypes. Thermal simulation is used to determine the best design based on constraints such as device temperature, weight, and cost. A design that minimizes one constraint and satisfies the other constraints would be nominated as a candidate for prototyping. Lower operating temperatures would indicate a better design in terms of functional or lifetime reliability. What is often not considered during thermal simulation is the concept of variability. Variability in operating environment or functional operation, and associated power distribution, is often considered but variability in the physical components effect on reliability is seldom considered. With today’s computational resources, simulation can now be used to drive design decisions beyond predicting zero-hour nominal performance, but also consider reliability due to variability. Simulation based thermal analysis driven systematically, combined with statistics, can be used for reliability assessment. Appropriate use of statistics enables the solution space to be explored efficiently to optimize the design choice, and verify the occurrence of failures are suitably low.This paper introduces the concept of driving design through simulation-based reliability assessment. The study is based on a high power network server ASIC (Application Specific Integrated Circuit) with variances in thermal performance at discrete areas in the TIM1 (Thermal Interface Material) and TIM2 bond lines. Discrete areas of less than nominal performance are used to efficiently model variances in interface flatness and heatsink spring force tolerance. A Design of Experiments on a 3D simulation model is used to develop a surrogate for ASIC maximum operating temperature. The Monte Carlo method with probability distributions is used to exercise the surrogate to study thousands of design variants efficiently to predict the observed failure rate.