# How Accurate Is Switching Converter Modeling vs. Lab Measurements?

### Abstract

Power supply design models and simulations are usually confined to typical values. A more complete picture of the end design process requires consideration of potential silicon variations that a model doesn’t account for. In this design solution, we’ll examine nonidealities like capacitor or inductor variations that contribute up to ±10mV (0.3% of nominal 3.3V) deviation to the droop voltage, illustrating the need to understand and account for physical component variations during simulation.

### Introduction

In a previous article, we looked at hand calculations for buck converter error budgeting and saw how they stacked up against simulations from EE-Sim® tool, Maxim's online SIMPLIS power supply design and simulation tool. This naturally leads to a deeper question: How closely do the models match bench measurements? Because at the end of the day, it's the physical silicon that matters, and all the calculations and simulations in the world won't help if they don't adequately model your real power supply.

So, how do the models stack up to bench measurements? For the quick answer to that, check out Figure 1. It shows the MAX17242EVKIT output responding to a load step, and the MAX17242 SIMPLIS model responding to the same load step in Maxim’s new OASIS (Offline Analog Simulator Including SIMPLIS) design tool. The correlation, especially with regard to peak surge/sag voltage and recovery time, is remarkable.

Granted, this didn't come from the typical model, and for good reason: each piece of silicon will respond a little differently, while each instance of the model is bound to respond identically. Model parameters are correlated to typical silicon parameters, so comparing them to a random IC is bound to introduce discrepancies. In order to establish the accuracy of the models, the model we used to generate the above step response has been tweaked slightly to match the values of the silicon on the bench, a process which we'll explain in detail below.

Determining the necessary adjustments requires a quick trip to the frequency domain for Bode plots, but once we've established that we can trust the model’s results, we can use it to show the range of step responses predicted by process and component variation.

### Correlation Process

When it comes to diagnosing issues and finding discrepancies in a power supply, there are few tools in a power designer's toolbox more powerful than the Bode plot. The presence of ringing on a transient waveform hints that a problem exists, but it does little to identify the source of the issue. Moving to the frequency domain provides more information about where the issue is coming from and how to fix it. When it came to correlating the model parameters to the hardware, we worked exclusively in the frequency domain. Once the Bode plots lined up, the time domain correlation displayed in Figure 1 happened by itself.

Maxim uses SIMPLIS, in part because you can't take a quick trip to the frequency domain with SPICE. You either have to stop and take the additional time to create (and correlate) a linearized average model to run in a SPICE AC simulation or work tediously in the time domain by settling and running the time domain model over and over again with different test signal frequencies in long time domain SPICE simulations. You would then perform FFT post-processing for each output and finally, manually stitch together the FFT results into Bode plots.

On the other hand, the SIMPLIS simulator included with Maxim’s new EE-Sim OASIS design tool provides a periodic operating point (POP) analysis that quickly settles the time domain model. It includes a proprietary linearizing/averaging algorithm that seamlessly generates bode plots directly from extracted frequency domain data. AC and transient data are effectively derived from the same model in a single simulation, speeding up the correlation process by orders of magnitude. Speaking of speed, in addition to the bode plot advantages, the switching converter time domain simulation in SIMPLIS is simply faster than in SPICE. Importing the model into Maxim's EE-Sim OASIS design tool provides the ideal environment to rapidly tweak and test the model in both time and frequency domains.

The first thing we did with the MAX17242 evaluation board was to run AC analyses on the control loop, which we compared to the simulations from a typical model placed in an identical circuit (Figure 2). The first of the three bode probes in the schematic allows us to capture the full-loop transfer function, while the other two allow us to split it up into the individual compensator (OUT to COMP) and power stage (COMP to OUT) transfer functions. This setup was mimicked on the bench measurements, and the model and bench results are overlaid in Figure 3. The plots are certainly similar, but there are clear discrepancies with certain poles and gain values.

By splitting the control loop, the source of the discrepancies between the model and the hardware becomes obvious. The power stages of the model and the board are almost identical, but in the compensation stages the DC gain of the model is too high, and the first pole happens too early. By pushing the transconductance gain and output resistance of the error amplifier away from their typical values (but still within spec), we can push the model's behavior to mimic what we observe on the bench. The output resistance and transconductance gain were the only parameters we touched, and the result is shown in Figure 4. The measurement and the model neatly overlap one another, with gain never deviating by more than 2 dB and phase response typically within 5 degrees or closer through the crossover frequency.

Once the measured and modeled Bode plots lined up, we moved on to the transient analysis. We began the testing on the hardware, taking careful measurements of both the load current and the output response. To keep the stimulus identical and eliminate as much discrepancy as possible between the model and the bench, we used the measured load current to create a custom current source in the simulation environment. This allowed us to test the model using the same load step that the hardware experienced, rather than an idealized approximation.

At this point, having correlated the Bode plots and supplied the same stimulus for the physical and modeled circuits, the transient responses lined up without any extra work required. Figure 1 is evidence that, providing for variation within silicon, the SIMPLIS models used within EE-Sim provide exceptionally accurate responses and is a reliable tool for the design and simulation of real-life power supplies.

### Extrapolating Results

Establishing a high level of trust in the model allows us to not only design and simulate typical power supply designs, but also sweep specific parameters to understand the range of potential transient responses across component and silicon variation. Figure 5 illustrates this, showing the expected load step responses and bode plots for ±10% tolerance in the load capacitance, ±20% tolerance in the inductor, and ±3dB variation in error amplifier transconductance.

These results match up with intuition: increasing the output capacitor or decreasing the inductor value will decrease the magnitude of the voltage droop after the load step, but does little to impact the overall settling time. Increasing the error amplifier transconductance, on the other hand, has little impact on the voltage droop, but gives a quicker recovery to the nominal output at the expense of more ringing during the recovery.

### Conclusion

In order to effectively use our design tools, we need to trust them, which requires us to establish when they are accurate and when they are not. When it comes to the SIMPLIS models used in EE-Sim design tools, comparisons to bench data show that we have a high degree of confidence that our simulation results reflect the expected behavior of the part. The nature of simulation, however, means our models and simulations are confined to typical values, and a more complete picture of our end design requires us to consider potential silicon variations that a model won’t account for. In our situation, the transient droop voltage predicted by the compensated model matched the measurement to within tenths of a millivolt. However, nonidealities like capacitor or inductor variations contributed ±10mV (0.3% of nominal 3.3V) deviation to the droop voltage, illustrating the need to understand and account for expected physical component variations during simulation.