Simulation in electronic design: Workshop for Model management & Machine Learning

Résumé : This thesis was written in the context of simulations carried out by Intel to estimate the power consumption of its future products. The company, which often reuses similar electrical circuits in its products, does not take sufficient advantage of the knowledge acquired in the many simulations carried out previously. Simulation models are poorly catalogued, rarely reused by the engineers in charge, and it's often necessary to start from scratch with a job only partially done in the past.

Our aim is to offer the company methods for better management of its power consumption simulation models, to facilitate reuse of the knowledge capitalized on in previous simulations. This involves, for example, detecting models simulating standard circuits (memory, cores, etc...), which are very often created, in order to offer engineers ready-to-use libraries for these models. We have chosen to focus on the extraction and exploitation of data from models, to enable the company to implement PLM later.

To achieve these goals, we employ machine learning methods to exploit the metadata attached to the models, and the data contained in the models. We first propose an algorithm that takes advantage of three metadata attached to models to evaluate the distance between each pair of simulation models. We then use these distances, which can be weighted, to propose groups of similar models to simulation engineers, using hierarchical clustering. For the data contained in the models, we propose to use a mathematical language processing algorithm. In particular, we exploit the equation describing the power consumption of the modeled circuit, to quantify the distance between two simulation models. Again, we use this distance to group similar models according to this criterion, using the OPTICS clustering algorithm.