on-site: complex room 1006
digital: MS Teams (link shared via e-mail)
Marcus Stiemer: Mathematical Optimization and Machine Learning to Support Electronic Design Automation
Electronic design is a hierarchical process starting from a conceptional system description, which is gradually transformed into a schematic, its physical realization (e.g., as a printed circuit board – PCB, or as an integrated circuit – IC), simulation-based functionality tests, and finally tests in the electrical environment the designed component will be built in. The individual steps in this sequence further subdivide into various operations. The realization of a schematic as PCB within a certain technology framework, e.g., comprises the placement of its parts and a subsequent routing process. Modern electronic components have reached a complexity that makes powerful tools indispensable to support the developer’s work during a design process. Individual steps of the hierarchical process can be modelled and simulated by numerical methods, and various techniques can be employed to search for good parameters on each design step, including mathematical optimization. However, decisions have frequently to be taken on an early design level that will show consequences on a later step in the design hierarchy. An adequate consideration of late consequences is difficult to implement with classical optimization techniques. A prominent example are electromagnetic compatibility issues that remain hidden during a schematic optimization, but will spoil the whole design when they are identified in a test of the component in its electromagnetic environment. Other examples are signal integrity problems appearing in PCB simulations, that have been invisible on the schematic level. Recent research focuses on machine learning (ML) techniques to account for late consequences already on an early design level via the identification of relevant patterns in legacy design data stemming from earlier projects. By learning such patterns, the employed ML methods mimick the experience that a skilled expert brings into an industrial design process. This presentation gives an overview over numerical models, mathematical optimization techniques, and ML approaches that can be used to enhance electronic design processes. Particular focus is laid on the availability of data from earlier designs and how these can be used to improve a current electronic design process via ML. The enhancement of classical design- and identification techniques by hierarchy-overarching, data-based ML-procedures makes it possible to combine classical parameter search techniques from mathematical optimization with information hidden in the design process’ legacy. Certain individual methods will be closer analyzed, and new results will be presented for them. In addition, the concept of a service platform is presented, which administrates data pipelines of legacy electronic design data, algorithm pipelines, further services to efficiently implement machine-learning based EDA, and finally a meta-algorithm to choose the right tools from the service platform for a particular design case. Such a platform has been developed and introduced into industrial electronic design in the project progressivKI ( https://www.edacentrum.de/progressivki/ ), funded by the German Federal Ministry for Economic Affairs and Climate Action. In addition, required computing resources are estimated and adequate ways to make the platform extensively available are discussed.