Skip to main content

An Introduction to Circuit Design Machine Learning

Key Takeaways

  • How machine learning uses tangible measurements to guide its processes.

  • The incentive system is at the core of many machine learning methodologies.

  • The importance of the design space and how its size and navigability influence machine learning models.

Pad shapes with traces and vias running horizontally

Machine learning offers the ability to trivialize some of the design aspects of circuit CAD

Machine learning is automating many tasks that were long thought to be far too sophisticated for computers alone and circuit design is no exception. Continued advancements in technology, as well as further research and development into data science and supporting fields, have resulted in rapid gains in industries that even a handful of years ago seemed too complex for automation.  

Companies and operators that are able to mix automation into their design workflows are at an express advantage: dense designs can have some or all of their layout reduced. Human oversight still allows for feedback or the user to edit the machine’s performance after running–this is commonly how layout designers will work with autorouter software. Still, circuit design machine learning holds the promise of something much more powerful: a circuit that can be built on a reasonable timescale with minimal oversight, freeing up the designer from the majority of layout work. Designers can therefore more effectively contribute time to the higher levels of design abstraction and leave the rote and tedious work to automation.

How Does Circuit Design Machine Learning Begin?

One of the most difficult aspects of circuit design machine learning is the nature of PCB design. While there are definite pratfalls to avoid, the “best” course of action relies on a variety of inputs, where modifying one parameter inevitably results in the depreciation of another. While there are several training approaches to circuit machine learning, a difficulty compared to more standard data outputs is the length of time it takes to generate data, much less the amount of time to build a large enough dataset for training. 

Further, many training methods are often proprietary and contained to individual businesses; though this allows for multiple methods of advancement, it can sometimes serve to hinder overall progress. To tune the machine learning, it is imperative to determine which parameters form the most basic evaluation of circuit functionality across a wide range of topologies. Fundamental circuit characteristics could include, but are not necessarily limited to:

  • Gain - The ability to amplify an input signal, typically voltage, current, or power. The greater the gain, the greater the boost in signal amplitude; various differential amplifiers are used to increase a signal to meet minimum threshold detection levels.

  • Bandwidth - The distance between the highest and lowest frequency in a range.

  • Power - The rate of energy transfer. Decreasing power draw is the main aim to increase board autonomy in the field where it may have little to no accessibility. 

  • Area - The area devoted to a circuit on a board. Without inducing significant EMI/EMC issues, a smaller area is more conducive to the overall layout, especially in a common dense board design.

Rewards Incentivize the Behavior of Machine Learning

The reward function is scaled to incentivize based on the number of design criteria met each iteration of critique as well as how effectively it improves the parameters used as the basis for evaluation. The last two parameters are of particular importance with denser boards–as areas shrink and densities increase, electrical and mechanical issues may arise from line inductance, substandard thermal dissipation, or other proximal effects. For decades, shrinking dies have been able to support board miniaturization, but with Moore’s Law slowing as chips inch toward the current limits of practical physics, additional resources need to be utilized to maximize design space and efficiency. 

Circuit design machine learning is beginning to bridge the gap, offering a tool to support designers and engineers. At its core, a deep deterministic policy gradient operates on the basis of a two-level system: one sub-routine – the actor –  implements the design or changes to the circuit simulation, while a second sub-routine – the critic – analyzes, grades, and incentivizes the actor’s actions based on some tangible measurement. By gauging the environment of the current state of the circuit, the actor can learn historically to optimize its expected reward. Discussion of the two sub-routines and evaluation/reward model is largely an abstraction–individual circuit design machine learning applications will feature a similar framework, but the exact implementation may differ. 

The Influence of the Design Space on Machine Learning

It is important to recognize the width of the action space involved in performing a task as complex as automated circuit design. The actor contains the encoder-decoder conversation that passes and translates low-dimensional observations into a signal down the line. Due to the sequential action of the data flow, the order in which observations are fed to the decoder will influence the circuit design as well. This method does not require the need for a model, and instead, machine learning can act and be shaped by the most base-level information to provide solutions. Working directly from the data with no “middle-man” helps root out biases that might otherwise lead to inefficiencies in the machine learning circuit design. Every individual parameter of every instance of a component in the topology must be calculated and measured against maximum or minimum constraints for things such as inductance, capacitance, impedance, and more.

For circuit design machine learning, you want a system that is well-integrated with current and evolving machine learning methodology, and Cadence’s PCB design and analysis software is well-positioned to deliver on the promise of a revolutionary technical application to reduce turn time for any design. Included among the Cadence family of products is the OrCAD PCB Designer, a schematic and netlisting program with a simple user interface that belies its power and functionality.

Leading electronics providers rely on Cadence products to optimize power, space, and energy needs for a wide variety of market applications. If you’re looking to learn more about our innovative solutions, talk to our team of experts or subscribe to our YouTube channel.