Neural network-based real time thermal design expert system [packaging]

Ideally, package thermal performance should be taken into account during the package design process, and each new design would require modeling and simulation to determine package thermal performance. With current methods, this step is computationally intensive, and complete design with several iterations takes considerable time. The design process can be accelerated by more than an order of magnitude if package configuration thermal data is made available in real time. This was accomplished using a neural network-based “thermal calculator”. A thermal calculator was constructed for each package family by running simulations and varying package properties (e.g. die size, pad size, packaging material properties) parametrically throughout the package design domain. The results were used to train a neural net. Subsequent Web interfacing provided real time distributed thermal performance information. This paper discusses the thermal design expert system. TDES provides a higher level of abstraction than thermal calculators, and helps the designer by providing solutions across multiple packages. The tool provides a Web interface in order to determine user requirements. It then uses this information to query the thermal calculators for all packages, and recommends a package to the user based upon certain pre-determined heuristics. All tools have been validated to provide high accuracy, and this process has been shown to substantially improve new package design cycle time. TDES has been written entirely in C++ for easy maintainability. The TDES framework was designed to be extendable to accept other design criteria