Passive Simulation Models Gain Accuracy and Scalability

For some time, RF and microwave design engineers have been tapping the benefits of modeling and simulation. Now, however, time-to-market and cost pressures are demanding tighter links between simulation and final design. To shorten design time and lower cost by eliminating iterations, designers are looking for more accurate and reliable software-simulation models. Consequently, third-party software developers, component suppliers, and electronic-design-automation (EDA) vendors are working together to improve the accuracy of passive models while expanding libraries and enhancing simulation tools for both faster turnaround and first-pass success.

To meet the needs of the RF-circuit-design community in particular, software developer Modelithics recently released an enhanced version of its capacitor-inductor-resistor (CLR) library that expands its collection of surface-mount-component models. According to the developer, version 6.0 adds 15 new Modelithics Global Models. Each one represents entire part-value ranges for CLR families from passive-component suppliers like the following: American Technical Ceramics, Murata, Johanson Technology, Coilcraft, KEMET Electronics Corp., Panasonic, Samsung, TDK, and Darfon.

According to Modelithics, there are over 130 surface-mount global models in the library today. Each model features part-value scaling, substrate scaling, representations of higher-order resonant effects, accurate effective series resistance (ESR), pad de-embedding and thorough documentation. Modelithics' President and CEO, Lawrence Dunleavy, points out that alternative RLC models, which include S-parameter-based models, do not represent the design-application environment. They model only to the first self-resonance of a circuit element. Dunleavy adds that RLC models do not facilitate part-value optimization. In addition, they come with little or no documentation.

Each passive model in the Modelithics library is generated from multiple sets of S-parameter measurements that are made with parts mounted on different printed-circuit-board (PCB) substrates (Fig. 1). PCB boards used in the test fixtures include FR4, Rogers, and Taconic soft boards with thicknesses ranging from 4 to 59 mils. Alumina and quartz thin-film boards ranging from 5 to 25 mils also are sometimes used for high-frequency model validations. According to the developer, the loss mechanisms in these RLC models are derived from ESR measurements made with separate equipment, such as resonant lines and impedance analyzers.

Because surface-mount components are considered to be complex strip-like transmission-line components with a PCB backplane as ground, variations in the substrate parameters have a dramatic effect on the component's frequency response. By appropriately scaling affected model parameters, the Modelithics models incorporate data that enables the accurate accounting of substrate-related effects. Most of the models have been validated for a continuous range of substrate thickness. In terms of frequency, most of the new models have been validated to 20 GHz with some to 40 and 50 GHz. Higher-frequency validations are available upon request, according to Modelithics. Pad geometry scalability and selectable horizontal/vertical orientation are other unique features of these models. To further improve high-frequency accuracy, higher-order resonant effects are included in C and L library models.

Models And Accuracy
Recently, engineers from TriQuint Semiconductor used the Modelithics CLR Library within Agilent's Advanced Design System (ADS) to increase efficiency and decrease the costs of the design phase for a miniaturized front-end module. TriQuint's Foundry Engineering Director, Paul Litzenberg, notes, "The CLR Library's advanced and highly scalable models for surface-mount components are adding to designer confidence and helping TriQuint's improved design flow."

To demonstrate the accuracy of its scalable global models, Modelithics has fabricated lumped-element circuits and compared the measured data with simulated results. One such design is a fifth-order, lumped-element low-pass filter using CLR library and Agilent's Genesys software.

According to the application note titled, "Lumped Element Low-pass Filter Design and Optimization Using Agilent Genesys Software," the Chebyshev filter was first implemented in Genesys using ideal inductor and capacitor models. It was then refined through the addition of Modelithics CLR models with the modeling of transmission-line connections. The design concluded that the measured and simulated results were in good agreement, demonstrating a complete simulation-based, first-pass-success design flow.

While working with software developers like Modelithics, passive-component suppliers continue to update libraries with simulation models developed in house using Smith charts, S-parameters, ESR, Q, impedance, and other RF parameters. In doing so, they are working to improve the accuracy and speed of circuit simulation. Despite these efforts, however, some component suppliers observe that errors invariably tend to increase at higher frequencies.

ATC, for instance, has investigated the accuracy of such models for multi-layer capacitors and chip resistive components and observed that they are all generally valid to 6 GHz. Beyond 12 GHz, the errors tend to rise, notes Bob Grossbach, ATC's Vice President of RF Engineering. Below 1 GHz, the supplier also observed that the measurementand hence the simulationerrors for ESRs of capacitors below 2 pF are large. "While a global model is considerably better than an elementary model, it is not a substitute for building and testing RF circuits," emphasizes Grossbach. "It may give you some confidence that the part will work (or not), but it is not the final solution. Parameters of the model do not completely represent variations of the corresponding parameters of the actual production components. Nor do they account for variations in assembly, coupling to other components, or package effects," asserts Grossbach.

Ceramic capacitors cannot be represented truly at high frequencies with simple series-resonant-frequency (SRF) models. After all, the parameters of the model vary with frequency, temperature, and voltage. As a result, KEMET has taken the Spice route to generate its models internally. "Our Spice models offer adaptability to different ambient conditions, frequency, temperature, and voltages," states KEMET's Director of Advanced Applications, John D. Prymak. These models are created based on actual readings at room temperature. Variations in temperature are then applied to these models.

Simulation Tools
To serve myriad RF circuit designers, Modelithics has tailored its CLR global models for a variety of circuit simulators from major EDA vendors. These simulators include Agilent's ADS and Genesys, Cadence Design Systems' Virtuoso Spectre, Applied Wave Research's Microwave Office, Ansoft's Designer, and Sonnet Software's analysis engine.

Meanwhile, KEMET has released its full line of capacitor models to the SIMetrix/SIMPLIS simulation software, which allows signal integrity, power integrity, and board-design engineers to model the decoupling scheme of a PCB using the simulation software. According to the low ESR and ESL capacitor supplier, design engineers who use the SIMetrix/SIMPLIS simulation tool will have the ability to model the time-domain and frequency behavior of the manufacturer's aluminum, film, ceramic, and tantalum SMD capacitors. These models also are available online as Spice models. Prymak summarizes, "KEMET's goal is to get our models to perform with the most accepted, demanding, and accurate simulation software tools. We believe that we accomplish this goal with tools created by SIMetrix/SIMPLIS."

With surface-mount components being squeezed into even tighter spaces, there is more interaction between these miniature components. As a result, multidimensional parameterized analytical models are needed that will take such factors into consideration. To enable vendors to build such advanced models, Agilent has created the electromagnetic (EM)-simulation-based Advanced Model Composer (AMC), which claims to combine the accuracy of EM simulation with the speed of analytical models. AMC provides a method to build multi-dimensional parameterized analytical models for passive planar components, states How-Siang Yap, Product Manager for Agilent's ADS and EM simulation tools. It produces highly accurate analytical models that can be used by all ADS circuit simulators.

This model-generation technique, which is referred to as multi-dimensional adaptive parameter sampling (MAPS), selects a minimum number of EM simulations. It then builds a global analytical fitting model for the scattering parameters of general planar structures as a function of both the geometrical parameters and frequency with predefined accuracy. Data points are selected efficiently and model complexity is automatically adapted. The algorithm consists of an adaptive modeling and sample-selection loop (Fig. 2).

The modeling process starts with an initial set of data points. New data points are selected adaptively in such a way that a predefined accuracy Δ for the models is guaranteed. The process of selecting data points and building models in an adaptive way is called reflective exploration. Yap notes that this approach is useful when the process that provides the data is very costly, which is the case for full-wave EM simulators. Reflective exploration requires reflective functions that are used to select a new data point. The reflective function used in the MAPS algorithm is the difference between two different models. A new data point is selected near the maximum of the reflective function. When the magnitude of the reflective function becomes smaller than Δ over the whole parameter space, no new data point is selected.

If one of the scattering parameters has a local minimum or maximum in the parameter space of interest, Yap points out that it is important to have at least one data point in the close vicinity of this extremum in order to get an accurate approximation. If there is no data point close to a local maximum or minimum, the local extremum is selected as a new data point. For resonant structures, the power loss has local maxima at the resonance frequencies. To get an accurate approximation, a good knowledge of these local maxima is very important, emphasizes Yap.

The scattering parameters of a linear, time-invariant passive circuit satisfy certain physical conditions. If the model fails these physical conditions, Yap emphasizes that it cannot accurately model the scattering parameters. The physical conditions act as additional reflective functions. If they are not satisfied, a new data point is chosen where the criteria are violated the most.

See associated table