Skip to main content

Driving New Developments With Electronic Design Automation Using Machine Learning

Published Date
Author Cadence PCB Solutions

Key Takeaways

  • Algorithmically, machine learning can come in many different flavors.

  • How machine learning can predictively adapt to the core layout activities.

  • The use of machine learning in yielding results in chip design.

Analog signal brain hemisphere input, digital signal AI brain output

Machine learning helps onboard systems operate more like humans using predictive and analytical methodology

Across a wealth of industries, artificial intelligence promises the ability to unlock new paths in product development. Through sophisticated algorithms, machine learning can be utilized to read and predict live datasets. With early results trickling in from real-life applications, there are significant opportunities for early adopters interested in adding machine learning to their research and development operations. For electronic design automation, machine learning is key to revolutionizing traditional workflows.

Styles of Machine Learning

Machine learning is set to advance automation across many industries, and electronic design automation through machine learning is no exception. At the core of machine learning are algorithms intelligently designed to provide system feedback based on datasets. To provide feedback, the algorithm relies on some sort of judgment to help sort or grade the data it is viewing with degrees of confidence. Machine learning is a vast field, and there are many different methods for organizing an algorithm based upon the dataset being analyzed as well as the implementation of the algorithm within the design space itself.

At the highest level of abstraction, the styles of teaching guide machine decision-making. It can be thought of as the motivation for the machine behind its learning process. Less important than how the software incorporates the data is defining what style of learning the machine is engaged in.

Knowledge Acquisition Type

Explanation

Supervised learning

A labeled offline dataset used to train for future datasets; no refinement

Unsupervised learning

An unlabeled offline dataset used to train for future datasets; no refinement

Active learning

Algorithm chooses data from input spaces and refines itself during the search process

Reinforcement learning

Algorithm interacts with data through reinforcement feedback - the goal is to maximize reward

The Optimization of Electronic Design Automation With Machine Learning

Consider the standard design flow of a typical PCB production. After receiving the controlling documents from an engineer, a layout designer is tasked with fulfilling all the requirements of the board in the most optimally designed manner. By following the manufacturer’s recommendations as well as standard rules of thumb, the layout designer ultimately arrives at a proposed solution for the board’s placement, plane design, routing, and more that is an exceedingly unique work in terms of implementation of criteria. Realizing layout and design is very much an art - give two designers the same instructions and they will produce two separate board files that individually meet all of the goals set out by the engineer. This, of course, is where design review allows for feedback that maximizes the operations of the board within the provided parameters.

Differentiating designs will involve optimization of the board’s features in a way that integrates the constraints of the schematic, the fabrication materials, and the manufacturing processes and components available. From placement to routing, machine learning can be utilized in a semi or fully autonomous fashion to guide the layout process:

  • Placement - Initially, machine learning models need to consider the placement of the components. By looking into the geometric relationship between circuits and determining the cost of the design decision, reinforcement systems can use their incentivization scheme to optimize placement. Current placement models are adept at handling the logic of the schematic, but provide suboptimal datapath placement. To remedy this, placement with automatic datapath extraction (PADE) has seen its usage increase. PADE separates the logic of the datapath from the general logic of the circuitry via neural networks and support vector machine algorithms. 

  • Routing - Machine learning requires design rules just as a human operator would. In fact, design flows somewhat backward at this point, as it is routing and its governing rules that define placement more than the other way around. For machine learning solutions, this can take the form of a convoluted neural network for detecting DRC violations all the way to using imaging to analyze pins and predict where routing is likely to become congested in the design. Routing must also account for circuit characteristics such as clocks/timing, power draw, and topology; this is usually best achieved with some common regression-based algorithms such as support vector machines (SVM), random forest (RF), and boosting.

  • Power design - Designers are always looking to reduce power draw. Lightweight power consumption avails boards in the field to higher reliability as well as reduces energy consumption. One solution to machine learning-based power design is to employ a neural network trained to expedite the calculations in the frequency domain. Typically, calculations need to be performed in the frequency domain due to the time domain’s inherent dependency. However, evaluations of signals are best performed in the time domain, like using signal convolution to form an eye diagram. 

Eye diagram on oscilloscope screen

Eye diagrams are useful tools to analyze the power characteristics of a design

Machine Learning Improves Chip Design 

Perhaps unsurprisingly, chip design is also aided by machine learning algorithms. Two production processes, lithography and mask optimization, are areas of research interest due to increases in efficiency:

  • Lithography - Utilizes SVM to detect hotspots in routing and path prediction to further discourage the development of hotspots. As density continues to build in chips, alternative methods such as end-to-end object-detection models may see further development and usage.

  • Mask optimization - Machine learning is focused on increasing the efficiency of a process that matches the finalized chip to the design layout. As is the case with lithography, current methods are running up against the limits of density in semiconductor design. To combat this, a new model known as a generative adversarial network (GAN) produces a lithography image and is read by a discriminator algorithm. To aid the new model, the current mask methodology could be used to pre-train the algorithm.

Image of silicon lithography in semiconductors

Lithography is just one chip design process that stands to be improved by machine learning

With chip designs soon contending against physical limits, machine learning is more valuable than ever throughout the development process. Whether you are autorouting dense designs or utilizing cutting-edge algorithms for chips, PCB design and analysis software offers you unparalleled solutions to electronics design automation in machine learning. See how Allegro PCB Editor offers a powerful toolset to tackle even the most challenging projects.

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.