IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS, VOL. CAD-I, 10.
4, OCTOBER 1982 163
Optimized Extraction of MOS Model Parameters DONALD E. WARD, MEMBER, IEEE, AND KYRIAKOS DOGANIS Abstract-A common problem faced by designers in simulating MOS circuits is the specification of model parameters. Typical parameter extraction procedures determine parameters sequentially, ignoring many of the interactions between parameters. The resulting fit of the model to measured data may be less than optimum. Furthermore, the usual extraction procedures are specialized to a particular model; con siderable effort is required to accommodate improvements in the model. This paper describes the application of general-purpose optimization techniques to the problem of extractng MOS transistor parameters. With this method, all parameters are extracted in a single operation, using the Levenberg-Marquardt algorithm to find a least-squares fit of the model to measured device characteristics. The technique is inde pendent of the MOS model used-new models may be incorporated simply by inserting the moel equations into the extraction program. The optimized parameter extraction technique has been implemented and has proven effective. The ease with which the model may be changed makes the program particularly useful for the evaluation of new models. I. INTRODUCTION A COMMON PROBLEM faced by designers in simulating MaS circuits is the specifcation of model parameters. Although most MaS models are based on physical theory, there are always some parameters which do not have physically well-defned values, and others for which the physical values do not give the best ft to actual device characteristics. Thus it is generally necessary to extract model parameters from mea sured transistor data. In a typical extraction procedure, parameters are extracted sequentially, one at a time or in small groups. The value of each parameter is assumed fixed and accurate for use in ex tracting further parameters. A small portion of the model and data from a limited part of a device's operating range are used in extracting each parameter. A linear least-squares fit is often used over each region. As an example, a typical extraction procedure might find values for the threshold- and mobility related parameters in the linear region of operation, and then use these values in fnding the velocity saturation and channel length modulation parameters. Several difficulties may be encountered with this procedure. First, the sequential method of extraction does not account for the interaction of parametcrs, or for the effect of a param eter in regions of operation other than that from which it was obtained. Thus, for example, the effect of channel-length Manuscript received March 8, 1982; revised May 30, 1982. This work was supported by the Army Research Office under Contract DAAG29- 80-K-Q013. D. E. Ward is with the Integrated Circuits Laboratory, Stanford Uni versity, Stanford, CA 94305. K. Doganis is with the Integrated Circuits Laboratory, Xerox Palo Alto Research Center, Palo Alto, CA 94303. modulation on the linear region (in some models) may not be taken into account when extracting mobility, and it is ofen found that the mobility reduction parameters obtained in the linear region are not appropriate in saturation. A second problem is that some parameters which have a large impact on the simulated characteristics may be deeply imbedded in the model equations. The parameter VMAX in the SPICE level 2 model is a good example of this. The result of these two prob lems is that the model does not give the full accuracy of which it is capable. A final difficulty is that parameter extraction routines must be tailored to a specific device model. This can produce signifcant overhead in modifying or replacing models. This paper describes another approach to the parameter ex traction problem -the use of general-purpose optimization techniques to obtain values for a complete set of model param eters in a single operation. This is accomplished by simulta neously adjusting the values of all parameters to produce a nonlinear least-squares fit of the model to a set of measured device characteristics. The result is an optimum set of model parameters. A key feature of this approach is that the model code from the circuit simulator may be used directly in the parameter extraction. This minimizes the effort required to implement a model and insures that the equations used in extraction are identical to those used in circuit simulation. The remainder of this paper describes the use of optimiza tion methods in more detail. Section II outlines some of the functional and numerical details of the algorithm, and briefy describes an implementation of the technique. Section III gives results obtained with the optimization method, and Sec tion IV presents some conclusions derived from this work. II. THE OPTIMIZATION METHOD A simplified flow chart of the optimization method is shown in Fig. 1. Here p is the vector of parameters to be optimized; the vector f(p) consists of the error evaluated at each data point. The quantity to be minimized is the norm of the error vector, given by \\f( p ) W = L f k (P)2 = L ( h(P - Ii )2 (1) k k max ( k' Imin) where h is the calculated drain current and Ii is the measured value at the kth data point. I min is a quantity provided by the user; at currents above I min the relative error is used, while for Ii <Imin the absolute error (scaled by Imin ) is used. Any measure of the error may be used; this definition minimizes the rms error in the fit of the model to the data. Other defini tions might be used to minimize the maximum error, for example. 0278-0070/82/1000-0163$00.75 1982 IEEE