Invited Paper: End-to-end RFIC Topology Synthesis and Design combining Reinforcement learning and Inverse Design

Distinct from low-frequency analog design, design of radio-frequency, millimeter-wave and terahertz frequency chips follows a complex co-design process between circuits and electromagnetics (EM). These two domains are strongly coupled to each other, and this makes the design process extremely complex. Similar to circuit topologies, EM topologies can be very diverse ranging from lumped elements such as inductors, capacitors to distributed passive structures such as transmission lines, antennas, and all possible combinations, taking one or more elements of the set. The design process from ideation to schematic to layout can be extremely time-consuming relying heavily on expert knowledge, manual interventions, and iterative tuning of predefined circuit and EM topologies. These methods follow a bottom-up approach, starting with fixed geometric templates and using trial-and-error optimization. Other than long design times, this limits innovation and accessible design spaces. Here, we present a series of approaches that introduce a fundamentally different strategy: a universal inverse design method that can generate complex electromagnetic structures and subsequently fabrication-ready RFIC design and layout from specifications. We present a reinforcement learning design framework that canvasses the space for RFIC topology, circuits and circuit parameters and then synthesizes the interface electromagnetic structures through a deep learning enabled inverse design framework. Collaboratively, this creates a specifications-to-GDS design and synthesis flow for RFICs. The designs that emerge through this are often non-traditional and non-intuitive, but can break many of the trade-offs of known topologies. We present fabricated and measured results in silicon ICs demonstrating the feasibility of the presented approaches.