Everything You Should Know About Machine Learning in Power Electronics
Key Takeaways
-
An explanation of noise and power delivery networks.
-
Design features that can improve the PDN.
-
How algorithms can be found in power design and how they might be of use.
Machine learning helps turn digital signals into predictive analysis
More than any other design criteria, power circuitry is constantly being improved to produce new products with greater functionality. Most electronics users are happy to adopt new revisions that significantly increase battery life and nothing else–efficiency is key. But, there’s more to proper power design than the length of time a battery can go between charges.
Similar to the placement and routing stages of layout design, power plane design is a communicative process that attempts to find the most ideal solution for the design of power planes and other power-related features. Also, like other aspects of design, improvements in one area of the power circuitry come with drawbacks elsewhere.
To best approach this delicate balancing act, machine learning in power electronics has come into greater visibility as a way to predictively analyze the effects of power design choices without the user carrying them to their logical end. As artificial intelligence continues to gain prominence in circuit development, designers will have plenty to gain by improving their grasp of this crucial aspect of board optimization.
Framing the Power Delivery Network and Noise
The power delivery network or PDN is just what it sounds like: the network connecting all of the active components in a design to ground and power planes. The PDN is designed to provide low-noise power and low-noise return throughout the system while accounting for both the intentional addition of impedance as well as the non-idealized inclusion of parasitics inherent to the physical design of components.
Limiting noise inherent to a design helps to solve runtime errors that may detract from the board’s operation. In effect, noise can be thought of as something of a voltage leakage. Using Kirchoff’s Loop Laws, it can be ascertained that the voltage isn’t spontaneous, but rather an effect of the E/M field containing the energy losing some of its containment. This has the effect of drawing down the voltage on the rest of the PDN local to that region, and that can cause significant errors and even operational failure of active components that are not receiving the requisite voltage. Significant noise on power channels can also cause jitter, which can lead to severe timing issues.
Design Practices to Improve PDN Response and Reliability
In practice, it is far too difficult and time-consuming to attempt to measure the individual characteristics of the circuit elements and approximate them with an equivalent circuit. However, a number of workarounds allow engineers to build approximations of varying accuracy. A simple yet powerful analysis is merely calculating the DC resistance of the conductor plane and using Ohm’s Law to calculate the voltage at the point of reference. From here, the PDN voltage can be compared to the ideal - the real value should be somewhere less than a 5% voltage drop to indicate a proper PDN system, though the precise value will vary by system needs.
The three factors providing the greatest weight to the impedance of the PDN at any point are the capacitance between plane layers as well as the parasitic inductance found in capacitor leads and in traces on the board. Some design inclusions that will help reduce the voltage dropoff in the PDN system include the following:
- Optimize placement/plane design to reduce the length of power traces.
- Widen power traces.
- Avoid grouping too many different net vias in a small area of the plane that could choke the plane and expand the return path, creating a larger inductive loop.
- Keep power and ground planes close to the surface for ease of accessibility from the component’s perspective.
- Use via-in-pad on bypass capacitors.
Without a clean signal and reliability, a board is more likely to encounter issues during runtime if components become temporarily starved for power. Unnecessarily convoluted power delivery networks could also result in mechanical failure to components and the board material due to overheating. The efficiency of the PDN will rely on many design aspects, including layout, as well as material values such as substrate Dk/Df.
In practice, all circuit elements contain parasitics, which contribute to inductance and a phase shift of the signal
Applications of Machine Learning in Power Electronics
With a working idea of noise and ways to rectify it during layout design, it is time to consider how machine learning in power electronics can improve the design to reduce noise. New models have begun to predict the voltage drop as opposed to a field solver calculating the voltage drop in real-time. This is a lightweight computational solution (or at least more so than utilizing a field solver) that predicts two forms of voltage drops to arrive at a more predictive value. The passive component of the voltage drop is caused by losses along with the conductive elements of the PDN, while the active component estimates a drop due to switching behaviors and local eddy currents.
Gradient Boosting Algorithms
For this particular application, a system can employ a gradient boosting algorithm to predict the voltage drop due to DC resistance. To start at the beginning, gradient boosting utilizes decision trees and focuses on those it considers weak learners. Despite the name, weak learners are defined as decision trees that show more predictive capability than random chance. In effect, the algorithm places more predictive weight on the weak learner decision trees than on the well-defined trees, as the latter’s predictive ability is fully exhausted as a solved instance of the model.
Continually focusing on the weak learners means the model is further investing its focus into the decision trees that can provide insight the model currently lacks. This weak learner functionality is combined with a loss model and an additive model that introduces a weak learner one at a time to minimize the loss model.
With sophisticated algorithms, machine learning allows the system to isolate and spool mixed data to determine the next course of action
While automating PDN is only one aspect of machine learning in power design, and perhaps a less-heralded one at that, Cadence is committed to improving the machine learning in power electronics through a host of powerful PCB design and analysis software.
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.