TECHNIQUES FOR NONLINEAR CIRCUIT SIMULATION Large signal periodic time-domain circui simulation K.R. Whight Indexing terms: Time-domains, Signal effects, Simulation Abstract: A novel time-domain algorithm has been applied in the field of circuit analysis to model large signal effects under periodic excita- tion. Conventionally, the need to integrate out initial transients over many cycles had effectively ‘made time-domain simulation impractical. The new algorithm, however, neatly avoids this problem, yielding the periodic steady state directly by simulating over just one cycle. A prototype simulator has been written that implements the algorithm within a circuit context, and it has been ‘used in a preliminary study of large signal harmo- nic effects in both driven and autonomous circuits. In particular, the following circuits are analysed: a single-stage transistor amplifier, a varicap tuner, a full wave rectifier and a Colpitts oscillator. The nonlinear effect of jump resonance is observed in the varicap tuner circuit, An extension of the basic algorithm beyond harmonic analysis into the realms of mixers and other possibly noncom- mensurate frequency circuits is proposed. Pos- sibilities regarding mixed-mode circuit/device simulation are also discussed. 1 Introduction Large signal periodic effects are an important class of problem that in the past have not been amenable to direct time-domain simulation. Transient time-domain simulation, although easy to perform for a circuit with any number of strongly nonlinear elements, suffers from the fact that the system being modelled may have time constants that are large compared to the fundamental excitation frequency. Thus many cycles may need to be integrated to overcome the effect of initial transients and to arrive at the final periodic steady state. To avoid the problem of initial transients the author proposed a novel approach, within the field of semicon- ductor device modelling, which was implemented in the TESSA simulator [1]. This approach requires the system ‘equations to be recast in a spacetime form, as it considers the time axis effectively to be additional dimension to which periodic boundary conditions are applied, icc. periodicity is forced by the simulation mesh. In the Circuit context this requires, in its simplest form, that the circuit nodes be automatically repeated on timeplanes. © IEE, 1994 Paper 1244G (E10) frst recived Ist October 1993 and in revised form 29th March 194 ‘The author is with the Philips Research Laboratories, Cross Oak Lane, Redhill, Surrey RHI SHA, United Kingdom 1b Proc-Circuits Devices Syst, Vol. 141, No.4, August 1994 Each node on each timeplane is distinct and has a unique voltage variable associated with it. Each timeplane has aan associated fundamental phase, but there are no nodal, phase variables. Taking a spacetime viewpoint, the simu- lation mesh has a cylindrical topology with time running around the circumference of the cylinder, the distance around the cylinder being equal to the fundamental period. A more general description of the underlying ‘method as applied to semiconductor device modelling is ‘given in the TESSA paper [1]. The problem now becomes that of relating the current flowing between spatial node neighbours, i. on the same timeplane and connected by a circuit clement to the volt- ‘ages on relevant spacetime nodes. Once this has been achieved a matrix equation can be constructed, by apply- ing Kirchoft's current conservation law at each spacetime node, in which the unknowns are all the spacetime nodel, voltages. This process results in m equations in the 7 unknowns, which, on solution, yields directly the periodic steady state of the system. ‘The above method has been implemented in a proto- type large signal circuit simulator (LARGESS) and the program has been used in this preliminary study of large signal effects in both driven and autonomous circuits. ‘The circuits studied include a simple single-stage ampli- fier, a varicap tuner circuit, a full wave rectifier and Colpitts oscillator. The results of these simulations are reported following a more detailed description of how circuit elements are treated in the method and an illustra- tion of the matrix assembly process, using as an example aa simple linear LCR series resonant circuit 2 Spacetime formulation of circuit equations To implement the proposed algorithm the current flowing through any circuit element at a given instant needs to be related to relevant spacetime nodal poten- tials. There are three distinct types of time dependence within current/voltage relationships that need to be con- sidered, namely, those that have no time dependence (archetype: a resistor), those that have a time dependence that is a first-order differential of voltage (capacitor), and those that depend on an integral of voltage over time (inductor), 2.1 Resistor From the definition of resistance we have V=IR a from which the current flowing from node i to node j on timeplane 0 connected by a resistor of value R is ve—ve, I= 2 =e Q ass22 Capacitor ‘The definition of capacitative current is av cE 8) Taking a first-order approximation of the time derivative we may write the current flowing from node i to node j on timeplane 0 through a capacitance of value C as: ree (Win M=W 400) o a were the ~ superscript refers to the tlevant eguvaent node on the previous timeplane. 23 Inductor The dfn ofindutv erent is v te fast 6 where the integral represents the change in current flowing through the inductor, of inductance L, occurring during the time step dr and i is the constant of integra- tion. The constant of integration may be obtained by associating a small ‘sense resistor’ in series with the inductor. Again to first-order only, we may write, the current flowing from node i to node j on timeplane 0 as Wp vivae ve L + R where k and Lare the nodes to whieh the sense resistor, of value R, is connected. The direction of current through the sense resistor needs to correspond to the direction through the inductor, and all the current flowing through the inductor must also flow through the resistor. It is not absolutely necessary to employ a sense resistor but it makes the matrix assembly process easier and, a all real inductances have series resistance, itis not considered to bean unreasonable requirement. o 2.4 Example circuit equations for an LCR series resonant circuit To illustrate the matrix assembly process a linear series resonant circuit will be considered (Fig. 1). In this Figure ut Wy —— ~~~ unre u A ~—— wie R voreerf —nlavey — Fig. ‘Spacetime view of series resonant LCR circuit the circuit is seen to be repeated on timeplanes; however, the cylinder (of period 1) has been cut at ¢=0 and its, surface laid flat. There are a total of n free nodes at which the potentials (V; for i= 1, 2, ..., n) satisfying the circuit equations and boundary conditions are to be calculated simultaneously. There are also n Dirichlet boundary con- dition nodes, of which n/2 are earthed, the remaining ‘ones being at given fixed potentials Uy, Uz, .... Uys Note that the periodic temporal boundary’ condition, implies that nodes on timeplane ¢ = x ~ Ar are to be con- 286 sidered as in the immediate past of nodes on timeplane £0. Also note that the resistor already present in the Circuit will be used as the inductor’s sense resistor, as all the current that passes through the inductor at any instant also passes through this resistor. ‘The matrix assembly process will be illustrated by writing down the equations necessary to satisfy Kirchoff’s current conservation law at nodes 1 and n/2+ 1, the equations for the remaining nodes being similar to one or other of these. Node 1: From eqns. 2 and 6 we may write (Waa —Yoad At, Yaa Ve) , ear Vs (Baste Haat) Hasp—t 0 where the first term represents the inductive current out of the node and the second term the resistive current, o Node n/2 + 1: From eqns. 2 and 4 we may write Vix Vere , CW, Vara) Roe F 0 ®) where the first term represents the resistive current flowing from node n/2+1 and the second term is the capacitative current. Note that in this second term only two variables appear, as the other relevant nodes, required by eqn. 4, are in this case earthed Dirichlet nodes, ie. at zero potential ‘The above process, when continued for the remaining free nodes, results in a system of n equations in the n free nodal potentials which, being linear, is directly soluble for the periodic steady state. For nonlinear elements equations may be written down as above. However, the necessary Newton linearisation process would generate additional terms in the matrix when differentiating the nonlinear functions L, C or R with respect to the free nodal potentials. 3 LARGESS simulations: driven circuits ‘The circuit simulator LARGESS is written in C and uses data structures similar to those reported for TESSA [1]; the linear algebra routines are also similar, ie, LU decomposition without pivoting. The direct solver option is used exclusively in this work and a Newton linearisa- tion method is used to solve the nonlinear circuit equa- tions. The program is currently restricted in the type of circuit element, e.g. only an Ebers-Moll transistor model is available, but it is capable of general topologies. The spatial circuit layout is described using a simple input language, A rudimentary postprocessing routine has been written that allows the output of either raw data, as a function of phase, or numerical Fourier analysis data, both in a form suitable for plotting. All simulations were performed on an Apollo DN10000 machine working at approximately 2.5 MFLOPS, ie. unvectorised code, 3.1. Single-stage transistor amplifier AS a first example of the use of LARGESS the simple single-stage transistor amplifier circuit shown in Fig. 2 was simulated with an input sinusoid of $V at 10 kHz and a 10 V DC supply. An Ebers-Moll transistor model ‘was used with diode characteristics given by = 1-1) for -1<0<05V 0) where I, = 1.0 x 10° amp. Outside the above voltage range the characteristic was linearly extrapolated with IRE Proc-Cireuits Devices Sys, Vol. 141, No.4, August 1994continuous first-order derivatives. The gains of the ‘model's current-controlled current generators were a. 0.99 and a, = 0.5. In Fig. 3 the output voltage at node 6 1 ov Vout lima Fig. 2 Simulated amplifier circuit 0 02 os 06 OB io fundamental periods Fig. 3 Amplifier ouput waveform O25 imeplane, 3 Btimeplnes is shown as a function of time using 25 and 49 equi- spaced timeplanes spanning the 10 kHz period, the simu- lation CPU time was 5's (25 timeplanes). It can be seen that the output is similar on both meshes, and severely distorted. The first (negative voltage) half period has a “bite out of it caused by the collector-base junction becoming forward biased. The latter half period exhibits clipping caused by the emitter-base junction becoming, reverse biased. Conventional transient simulation tends towards the waveform shown in Fig. 3, but takes in excess of 15 min CPU time, ic. although the general waveform amplitude converged in a few seconds, the ‘bite’ was still not fully stable after 15 min CPU time. Further, when the conven- tional transient was driven by a sine rather than a cosine input the convergence of the ‘bite’ occurred from opposite directions, ic. one would initially overestimate the ‘bite’ and the other would underestimate it. The simu- lator used in these conventional transient simulations was an in-house simulator called PSTAR. The long PSTAR runtime is not due to an inefficient solution process. PSTAR itself has been compared to the com- ‘mercially available circuit simulator SPECTRE [2] and is judged to be equivalent in terms of runtimes. 3.2 Varicap tuner circuit ‘The next circuit analysed was that of the varicap tuner shown in Fig. 4. The varicap diode was modelled by a parallel combination of a nonlinear capacitor and a non- linear resistor (diode form, eqn. 9). The variable capac- itance was fitted by a 12th-order polynomial in voltage IEE Proc-Circuits Devices Syst, Vo. 141, No.4, August 1994 and was valid in the range 0-23 V. The coefficients were Co = 2.276121 x 10°C, = —9.940123 x 10-1? Cy = 4.600225 x 10°"? Cy = ~1.602318 «10-1? C4 = 3839086 x 10° Cy = ~6:311486 x 10° Coq = 1187672 x 107! Cy = 5.700524 x 10" Cy = 3132829 x 10°17 Cy = ~1.168078 x 107" Cio = 2.816101 x 10-29» C,, = ~3.956508 x 10-7? Cia = 2.458037 x 10-4 ent o4TpF asepr Ke Fig. 4 Simulated caricap tuner circuit LARGESS was used to simulate the output voltage as a function of frequency and input voltage amplitude at 10V DC bias; the results are plotted in Fig. 5. Bias 06; Q S00 6l0 20 —<6s0—~=~C~SKSSC*«C frequency, Mhz Fig. 5 Vu, at 10V DC, as a fetion of frequency and AC input ‘amplitude ON VAC Ha0V AC = TOV AC, USVAC 2o0V AC points for this figure were computed at 0.5 MHz steps, i. 100 per curve. The initial guess for each frequency sweep was the DC solution; subsequent points used the solu- tion at the previous frequency as the starting guess. The shape of the resonance curves depicted in Fig. 5 is seen ‘experimentally and shows the classic form of a jump resonance [3, 4] which develops as the input amplitude increases. Indeed, for the 2 V input curve convergence problems were experienced in the region of 629 MHz, ‘and the plotted curve could only be obtained from two runs, ie. one starting at 600 MHz and increasing fre- quency and one starting at 650 MHz and decreasing fre- quency. Runtimes averaged approximately 3 per bi point, using 25 timeplanes. Increasing the number of timeplanes to 49 results in approximately a 1 MHz reduction in resonance frequency. 23733 Full wave rectifier circuit ‘The next example circuit is a full wave rectifier, as shown in Fig, 6. This circuit was chosen for simulation as it was mF 0.633mH Your 20. some Fig. 6 Simulated fll wae rectifier circuit believed to be a severe test of the accuracy and con- vergence of the spacetime approach. The diodes in this circuit were modelled purely as nonlinear resistors. Small signal analysis would predict that the only ripple fre- quencies existing in the output, ie. the voltage difference between nodes 2 and 4, are at 2n times the input fre- quency, where n is an integer. In this example the input signal frequency was 50 Hz and two simulation mesh results will be discussed, a 25 Hz mesh spanned by 100 timeplanes and a 50 Hz ‘mesh spanned by 20 timeplanes. The amplitude of any subharmonics could therefore be observed on the 25 Hz ‘mesh, as well as the effect of mesh density on output amplitudes. Fig, 7 shows how the DC and 100 Hz signal levels are predicted to grow in the output, with increasing input ° © 20 30 7 input voltage, ¥ Fig. 7 Harmonic amplitudes as a function of input voltage at 80 He nt 10 ines 5 He meh 10 cele 50 He mesh 20 meplanes| amplitude for both meshes. Subharmonics and odd har- ‘monies of the 50 Hz input were all absent, ie. amplitudes in the numeric noise (1,0 x 10°! V), and it can be seen that the 20-timeplane mesh yields a solution close to the 25 Hz mesh (50 timeplanes per 50 Hz. cycle). Runtimes were approximately 10 s for the 50 Hz 20-timeplane mesh and 500 s for the 25 Hz 100-timeplane mesh. ‘Additional investigations with mesh densities of 50 and 100 timeplanes per 50 Hz cycle yielded better than 288 10% agreement between the two meshes up to the 8th harmonic (400 Hz), and within 20% to the 12th harmonic at signal power levels approximately —80 dB down on an input of 40 V (50 Hz), ie. harmonic signal levels of order 10 x 10°3 V. ‘Note that an even number of time planes have been used in these simulations. This was deliberate, as it has been observed that, for odd numbers of timeplanes, sub- harmonics of the $0 Hz input signal were generated on the 25 Hz mesh that only slowly decayed to zero ampli- tude with increasing mesh density, Even numbers of time- planes yielded zero amplitude subharmonics irrespective of the mesh density. This effect is not fully understood and is the subject of further study; it may be related to the chosen first-order discretisation. 4 LARGESS simulations: autonomous circuits Autonomous circuits present two mathematical dif culties to the large signal algorithm, First, and most seri ‘ously, as an oscillator runs from a DC supply and is not driven, the phase of any oscillation on the simulation ‘mesh is not uniquely defined. We therefore have an infin- ity of possible solutions and the resulting system of equa- tions cannot be solved. Secondly, oscillators have unstable DC steady states; in reality’ noise ensures that these states are not achieved, but in the noiseless world of the circuit model these states need to be actively avoided, Both these difficulties are simultaneously overcome in the following manner. When attempting to impose a particular phase on an oscillator solution, the first thing to note is that, as we have n equations’ (current @servation at.n spacetime Circuit nodes) in n +1 unknowns (the n spacetime nodal voltages and the frequency of oscillation) any method that avoids scanning the frequency must bring in addi- tional equations and possibly unknowns. Therefore, a resistive switched DC voltage source is introduced into the system and arbitrarily connected to one of the circuit nodes. The switch is closed for one timeplane only (at 0) and is open at all other times. This artifice essen- tially turns the system into a driven system, which can be solved. To determine the free oscillation of the original oscillator, the potential of the switched voltage source and its Switching frequency are adjusted for a zero ‘exchange of energy acrosss the switched link at the only instant in the spacetime-domain when itis closed. In itself the zero energy transfer condition is not suffi- cient to uniquely determine the free oscillation, and in addition a requirement of zero differential’ of the exchange energy with respect to frequency is necessary. ‘The addition of this last condition ensures that if the forcing of a real system is disconnected, the system cannot change frequency, as any such change requires external energy to be supplied. In the above, note that we have introduced two requirements (equations) ie. vy At a (9) and ae F ay where E is the exchange energy, ¥, is the forcing voltage of the additional switched DC supply, Vis the voltage at IEE Proc-Cireuits Devies Sys, Vol. HI, No.4, August 1994the spacetime node to which the switch is connected, R, is the resistance of the link, Ar is the time period for which the switch is closed, and, f is the frequency. ov imi 2558pH 3 10 1 2 3 frequency, Mie Exchange eneray as a fction of frequency and forcing voltage SOY 1 180 200 xo 780 30 transistor gain Fig. 10 Fundamental amplitude asa function of transistor gain We also have an extra unknown in our system — the forcing voltage — and therefore appear to have n +2 equations in n+2 unknowns, ic. a soluble system. However, when attempting to solve eqn. 11, estimates of the differential of the nodal voltage V, with frequency are required, Such estimates can be obtained by stepping the frequency toward the required solution, but a direct method can be obtained by solving in parallel two ‘copies’ of the oscillator at slightly different frequencies f and f+ Af, where f is a variable, ie. two spacetime IEE Proc-Circuits Deices Syst, Vol. 111, No.4, August 1994 cylinders are used, yielding an overall system of 2n +2 equations in 2n + 2 unknowns, which is directly soluble. ‘The above forcing process, which can be carried out experimentally, not only imposes the required phase con- dition on the solution but also avoids the unstable DC steady state by a suitable choice of initial value for the forcing voltage. The direct method, ie. solving the 2n +2 equations, can in practice only be applied from an initial condition close to the answer. This close initial condition is obtained by frequency-stepping the problem consisting of nn current conservation equations in the n unknown nodal potentials while adjusting the forcing voltage after each ‘sweep to obtain a minimum in the exchange energy when plotted against frequency. Once this initial condition is known the procedure for generating the parallel problem is performed by LARGESS automatically, ie. the user is not aware and does not have to guide the process, apart from an initial specification of a value to be used for Afi This algorithm for autonomous circuits will be illus- trated by analysing the behaviour of a Colpitts oscillator, 4.1 Colpitts oscillator circuit Fig. 8 shows the circuit to be analysed, in which the tran- sistor was again represented by an Ebers-Moll model as in the amplifier circuit of Fig. 2. Included in Fig. 8 are the switch elements that are used to obtain the solution, ‘These do not form an integral part of the Colpitts oscil- lator itself, A small signal analysis of the above circuit yields an oscillation frequency [5] of 2) (13) ‘A small signal analysis cannot yield any information on the amplitude of the oscillation, as this is determined by the nonlinearities in the system. In a real circuit the frequency of oscillation would. also depend on the amplitude of oscillation via the transistor junction capac- itances. However, the simple Ebers-Moll model used here has no capacitative elements. ‘The first step in the solution procedure is to scan fre- quency f while adjusting the value of the forcing voltage ¥,. Fig. 9 plots the resulting exchange energy for the case where = 200, ie. energy supplied to the oscillator via the link, as a function of f and V,. As ¥, is progressively decreased from 2.7'V to 25 (the corresponding DC value of ¥, is 2.82966 V) we sce the development of a ‘volcano’-type energy curve in which a shallow minimum forms. This minimum progressively deepens as the outer ‘volcano walls’ grow, until the minimum value becomes negative (for V, = 2.55 V), We are searching for the values of ¥, and f for which the minimum value is exactly zero. The method proceeds by loading in a solution from within the minima of one frequency scan that exhibits a negative exchange energy, then creating and solving a parallel problem at a slightly different frequency but the same V, (Afif= 0.001 in this case). These two problems are then combined and the variables V, and f added to the system, so that we have 2n current conservation ‘equations + eqns. 10 and 11, in 2n +2 unknowns, which is then solved, yielding the amplitude and frequency of free oscillation. For this case, where f = 200, we obtainan amplitude of 1.31 V at an oscillation frequency of 10.15 MHz. Fig, 10 plots the amplitude of oscillation as a function of transistor gain. The simulations were run using 30 ‘imeplanes and each point on a frequency scan (0.1 MHz steps) took on average approximately 2's of CPU time. Coupled solutions would take of the order of 10's. 5 Discussion ‘This study represents a first look at the proposed time- domain circuit analysis method, and as such the imple- mentation was kept simple, ie. uniform fixed timeplane separation, simple active element models and the use of a direct solver. More sophisticated active element models, although affecting the detail of the solutions obtained here, should not significantly influence the convergence properties of the algorithm, provided they do not intro- duce additional strong nonlinearities. ‘As required for nonlinear systems, the method makes no assumptions on superposition. Further, the accuracy of the solution can, in principle, be arbitrarily increased by increasing the number of nodes in the time dimension. I has already been demonstrated that relative distortion figures can be accurately determined to low levels on remarkably few timeplanes in the particular case of diode distortion [6]. Analytic distortion estimates for a diode, valid for small signal levels only, compared extremely well with TESSA [1] predictions. Transistor distortion, at the —80dB power level, has also been accurately reproduced on 10 timeplanes using TESSA [1}. It should be stated that the spurious points (Fig. 4 in Reference 1) have since been traced to an implementation error in a mobility model, and the modified code now yields a smooth third harmonic curve at high collector currents. Future developments could lie in the direction of simulation meshes that are adapted during the solution and in which nodes do not have to lie on timeplanes. A general mesh in which circuit nodes do not necessarily all hhave the same number of temporal incarnations spanning the fundamental period would allow optimisation of the ‘mesh in which high- and low-frequency portions of a circuit could be simulated together using a minimum of spacetime nodes. However, both new data structures need (o be designed and the elemental circuit equation formulations would need to be generalised, so that current could be apportioned appropriately between neighbour nodes not occurring at the same time instant, With regard to the lincar algebra, given the closed- Joop nature of the time axis a waveform relaxation iter- ative solution method would seem a good approach for further investigations. Iterative schemes based on conju- gate gradient methods, used successfully in semicon- ductor device simulation, were not found to be very successful All the problems tackled in this paper have been har- monic. However, it may be possible to efficiently tackle both noncommensurate frequencies and mixer-type prob lems which require high-order harmonies to be resolved, and would therefore, using the current formulation, need, prohibitively many timeplanes. This could be achieved by further generalising the technique used to solve the problem of autonomous circuits. In particular, the solu- tion space can be extended to a spacctime-frequency domain by including additional spacetime cylinders rep- resenting additional harmonic series, and noting that the voltage at a node at any instant of time is a sum over the spacetime cylinders used to simulate the circuit. Linear 20 interpolation of the voltage between temporal nearest neighbours on each cylinder would allow a soluble system of equations to be built up, as before, for the com- plete system. In effect, the problem matrix itself is inte- ‘grated by allowing time to increase simultaneously on all, cylinders during the matrix assembly process, and each time a node is encountered on a cylinder appropriate entries are added to the matrix. This process stops, yield ing » equations in the n unknowns, for mixer problems when the fundamental time period has passed, the nodes on higher-frequency cylinders having been visited many times. For noncommensurate frequencies a potentially soluble system of equations is obtained when the period of the lowest frequency present has passed. However, if, time is increased further, additional unique terms can be added to the matrix and therefore different solutions obtained. It would be interesting to study the variations in the solution with the matrix integration time. The con- vergence properties of these matrix equation systems are, of course, not yet known; also, it should be noted that the matrices are potentially dense in both frequency and time. The algorithm can be developed in the direction of mixed-mode circuit/device simulation as the approach is, by design, compatible with that taken within the semi- conductor device simulator TESSA [1]. This would enable the impact on circuit performance of changing device structure to be directly calculated without need of fan intermediary circuit model of the active device. Full advantage of mesh adaptation possibilities could also be taken to minimise runtimes. Alternatively, a circuit model could be optimised to a new device design within the same program. Periodic time-domain approaches to nonlinear circuit analysis have been proposed before; however, they have generally either been extrapolation approaches [7] or similar [8, 9]. All involve the repeated solution of an initial value, ie. a transient problem followed by mathe- matical manipulation to speed convergence to the required periodic steady state. The method described here is clearly distinct, as at no time is an initial value problem solved. The periodic boundary condition enforced by the cylindrical spacetime topology effectively turns the governing equations into elliptic, ic. closed boundary, form as there is no boundary in time with such a topology. Every timeplane is equivalent, none acts as an origin in time and all free spacetime nodal poten- tials are computed as one system of simultaneous equa- tions. Further, given the defining equations of spacetime ‘current (e.8. eqns. 2, 4 and 6 for resistors, capacitors and. inductors, respectively), the solution of a spacetime periodic steady-state problem is exactly equivalent to solving for the DC steady state in space. Indeed, the DC steady state is a special case of the periodic steady state, whereby the spacetime domain, of arbitrary period, is spanned by a single timeplane. Each node is therefore its ‘own future and past incarnation. Frequency domain techniques have also been pro- posed, the most notable of these being the harmonic balance technique [10]. This requires decomposition of a Circuit into linear and nonlinear subcircuits, each having the same number of interconnection ports. The division process is normally a compromise between the need to keep the number of ports to a minimum and the opposing need of a large number of ports to ease the description of the nonlinear network [11]. No such decomposition is required in the method described here: IEE Proc-Cireuits Devices Syst, Vol. MI, No.4, August 1983the circuit is analysed directly, as described in the circuit layout. 6 Conclusions The application of a simple and elegant algorithm for time-domain harmonic analysis of nonlinear circuits has been proposed and outlined in this paper. The algorithm ‘overcomes the classic objection to time-domain analysis in that initial transients are completely avoided and the Periodic steady state is computed directly. It also com- plies fully with the lack of a superposition principle for nonlinear systems. The only assumptions made are those implicit in the definitions of circuit elements. ‘A prototype circuit simulator, LARGESS, has been written based on the algorithm. LARGESS has been suc- cessfully applied in the simulation of both driven and autonomous circuits. The method is easy to use in that, siven a circuit layout, inputs are applied and an output is obtained, which is then analysed. No manipulation of the circuit or expert knowledge is required to perform a rea~ listic simulation, and there are no restrictions on the number of nonlinear elements, their types oF the strength of their nonlinearities, In particular, the known and experimentally observed nonlinear effect of jump resonance has been reproduced in the simulation of a varicap tuner circuit. Also, a minimum internal energy approach has been developed to solve autonomous circuit problems using the algo- rithm, In general, the algorithm converges well to a solu- tion, except when jump resonance is occurring IEE Proc-Cireits Devices Syst, Vol. 41, No.4, August 1994 The method has possible future extensions into the realms of noncommensurate frequencies and mixed-mode circuit/device simulation. 7 References 1 WHIGHT, KR. GOUGH, PA, and WALKER, P. ‘Large signal periodic time-domain simulation’ Int. J. Num. Modeling” Electron. ‘etworks, Devices and Fels, 1992, 8, pp. L]-21 2 SPECTRE: Cadence Design Systems, Inc. Corporate Headquarter, '555 River Oaks Parkway, San Jose, California 98134, USA 3 LEIGH, .R-: Essentials of nonlinear contol theory’, IEE topics in ‘control series 2 (Peter Peregrinus Ld, London, 1983) 4 JORDAN, DW, and SMITH, P- "Nonlinear ordinary difeental ‘equations, 2nd eda" (Clarendon Press, Oxford, 1987) 5 GHAUSI MS. ‘Electronic citcuits, models, functions, a design’ (Van Nostrand Reinhold, 1971) 6 WHIGHT, KR, GOUGH, P.A, WALKER, P. and BATTERSBY,, S.J: ‘Spacckime diseretisition ofthe semiconductor equation and its mn to distortion analysis. Proc. NASECODE VII, Copper 3. Colorado, USA, Sth- 12th April 191, pp. 247-249 7 SKELBOE, S.: "Computation ofthe periodic sieady state response of nonlinear networks by extrapolation methods, IEEE Trans, C= ‘ute Syst, 1980, CAS-27 pp 161-175 8 APRILLE, TJ, and TRICK, TIN. ‘Steady-state analysis of nonlia- ee ctcuits with periodi inputs, Proc. IEEE, 1972, 60, pp. 108-14 9 COLON, FR, and TRICK, TN.: ‘Fast periodic steady-state analysis for large-signal elecionic circuits, TEE J. Sol-State Cir- ents, 1973, SC, pp. 260-260 10 KUNDERT, KS, and SANGIOVANNI-VINCENTELLI, A. ‘Simulation of nonlinear circuits in the frequency domain’, 1EBE Trans. Computer-Aided Design, 1986, CAD-S, pp. 521-835 M1 RIZZOLL. V, and NERI. A.-'Stateof the art and present trends in nonlinear microwave CAD techniques’, IEEE Trans. Microwave ‘Theory Tech, 1988, %6, pp. 43-365, ysis and