Methods of circuit and topological synthesis of analog integrated circuits according to specification using machine learning and differentiable programming methods

The problem of schematic development (netlist generation), which arises in the development of analog integrated circuits, is formulated as an optimization problem for a differentiable smooth function using a combination of differentiable programming and machine learning methods. It is shown that this approach allows one to achieve the specification requirements and propose an optimal combination of circuit templates that make up an analog integrated circuit, without involving combinatorial optimization and reinforcement learning methods. It is shown that this approach provides significant speed advantages compared to traditional methods based on reinforcement learning. The possibility of fully automatic synthesis of an analog IC from specification to topology without expert participation using open-source software is investigated. The advantages and disadvantages of this approach are shown.