Designers‘ know-how for mixed analogue/digital systems difficult so far to automate

Crucial in the creation of microelectronic and micro-electro-mechanical systems (MEMS) is the design engineer's knowledge and experience. The development of such systems has long been researched, with improvements and ever more sophisticated automatic design methods being achieved. It has to be said that the engineer's knowledge from experience is not easy to represent visually so that it can be used for automatic design as would be the case if the systems were purely digital. In consequence, systems that are analogue and complex or are mixed analogue/digital are often less than perfect or have incongruencies like poor layout or incorrect test limits which come to light only later, often not until they interact with the components they are serving. More time and money thus has to to be invested at both the design stage and that of validation of the systems after fabrication.

Machine learning at the system design stage for major reductions in cost and risk

Therefore, the IntelligEnt group formed by IMMS and Ilmenau TU have developed application-oriented machine learning concepts in the design of microelectronic devices which can be tacked on to existing methodology and tools. The aim was to make use of the huge potential of machine learning for continued practical and research progress, dramatically reducing both the costs and the risks associated with system design.

This is possible if the incongruencies are found early and obviated to maximum effect. There are many areas in which machine learning has the edge over humans, among them being pattern recognition. If an automated design and characterisation process is integrated into the system design procedure it may, by recognising structures, reduce the total sum of data, sniff out anomalies and greatly improve the existing solution. In the IntelligEnt project, the machine learning algorithms are seen as tools with uses such as regression and classification based on deep learning methodology or for detection of outliers based on self- and semi-supervised learning.

The IntelligEnt work focussed on critical steps in the system design:

• Modelling – creating models of predicted behaviour: Critical for the design quality is the creation of models of the system components and/or the IP. The idea is that integrating features like power consumption and operational areas into system level models will in the end be automated because the system is capable of learning.
• Creating functions when designing analogue/mixed signal circuits: The structure or the topology is what determines how well a mixed analogue/digital circuit performs. To get computers to optimise the structure, a procedure for structure recognition and adaptation has been developed.
• Designing the chip manufacturer‘s construction plan at the analogue/mixed signal circuit layout stage: Layouts that are technically correct may still have incongruencies, among them substrate coupling, field transistors and mismatch. With Ilmenau TU, we have developed an AI-based anomaly detection method to automatically detect non-proven and potentially faulty locations in layouts. Our relatetd plugin for the free design tool KiCad is available on GitHub.
• Checking all steps and functions at pre-fabrication stage by simulation and verification: at each of the above stages, before the chip is fabricated, the system is checked, with the function groups increasing in size at each stage. The simulations carried out for this purpose make use of the models that have been extended by the machine learning methodology.
• Checking manufactured chips with absolute thoroughness by testing and characterisation: to date, optimisation of the test procedures and the choice of critical tests for MEMS and mixed signal systems have had to be done manually, which can mean that redundant tests are carried out. Machine learning algorithms are capable of revealing the factors which are interdependent, so that they can be taken account of. IMMS has developed a machine-learning-based method for analysing measurement data from
  ASICs. It delivers the same results as a manual evaluation, but is 10 – 30 times faster.