Analog Integrated Circuits and Signal Processing, 29, 151–158, 2001 C 2001 Kluwer Academic Publishers. Manufactured in The Netherlands. High-Level Fault Modeling in Surface-Micromachined MEMS N. DEB AND R. D. (SHAWN) BLANTON Department of Electrical and Computer Engineering, Carnegie Mellon University, Pittsburgh, PA 15213 E-mail: ndeb, blanton@ece.cmu.edu Abstract. MEMS structures rendered defective by particles are modeled at the schematic-level using existing models of fault-free MEMS primitives within the nodal simulator NODAS. We have compared the results of schematic-level fault simulations with low-level finite element analysis (FEA) and demonstrated the efficacy of such an approach. Analysis shows that NODAS achieves a 60X speedup over FEA with little accuracy loss in modeling defects caused by particles. Key Words: MEMS, NODAS, FEA 1. Introduction MicroElectroMechanical Systems (MEMS) has matured into a multi-disciplinary area with significant applications in industry. One trend in MEMS is towards higher degrees of integration consisting of many mixed-domain devices. As a result, new methods of hierarchical design for microsystems consisting of mixed-domain components are increasingly needed [1]. In addition, robust fault models and test methods are required to ensure high yield and reliability of MEMS products. Hence, analysis and design tools capable of assessing and preventing faulty MEMS behavior are necessary to ensure end-quality, in particular for the the many life-critical applications of MEMS. Surface-micromachined MEMS are devices where the micromechanical structure is fabricated using layers of thin films deposited and selectively etched, a process that is similar to the fabrication of electronic circuits. Most commercial applications use surface micromachining because of its well-developed infrastructure for depositing, patterning and etching thin films for silicon integrated circuit technology. Early applications of this technology include the digital mirror display [2] and the accelerometer [3]. These industrial successes and the existence of developed design expertise, stable fabrication services, and electromechanical modeling tools have made surface-micromachining a natural choice for developing a MEMS testing methodology. MEMS misbehavior can result from a variety of sources including particulate contaminations, stiction, layout curvature, layout etch variations, side-wall angle, and package tilt. In prior work, we have performed three-dimensional finite element analysis (FEA) of particulate contaminations [4]. Other fault modeling and fault classification techniques have also been investigated [5–9]. Although accurate, the amount of time to perform FEA is significant. The aim of this work is to develop particulate fault models using basic MEMS primitives. MEMS misbehavior discovered through low-level process simulation and FEA is represented at the schematic-level using existing twodimensional models of fault-free MEMS primitives. Our approach is similar to that of [6,7] since NODAS uses an analog hardware descriptive language. However, our approach is different from the schematiclevel simulation of faulty comb-drives discussed in [7], where separate one-dimensional models of fault-free and faulty comb-drives are utilized. We use existing “good” models of MEMS primitives to model particulate faults by changing the topology of the nominal design, an approach which is very similar to stuckat-fault modeling used in digital circuit testing [10]. Also, unlike the work in [6], we do not represent non-electrical behavior with equivalent electrical models. Instead, we use nodal models that permit the simultaneous use of electrical and non-electrical models thereby eliminating the need for inter-domain translation. © 2001 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works. 152 Deb and Blanton The rest of this paper is organized as follows. Section 2 gives a brief overview of the schematic-based simulator NODAS. Section 3 presents our case study of a MEMS accelerometer structure. In this section, we describe the different types of defects caused by particulates and our approach to modeling these defects using the “good” models utilized in NODAS. We also compare NODAS simulation results with that of FEA. Finally, Section 4 presents our conclusions and areas of future work. 2. NODAS for MEMS Simulation NOdal design of Actuators and Sensors (NODAS) [11] is a schematic-based simulator and a nodal matrixsolver similar to Spice except that it can simulate nonelectrical devices as well. It uses the concept of a physical system or nature consisting of through and across variables described in Kirchhoffian network theory. For translational mechanics, force is the through variable and displacement is the across variable. For rotational mechanics, moment is the through variable and angular displacement is the across variable. Although these natures are analogous to the electrical current (through variable) and electrical voltage (across variable), NODAS does not integrate mixed-domain behavior into Spice by translating a mechanical nature into an electrical analogue. Instead, NODAS uses modern analog hardware descriptive languages like VHDLAMS and Verilog-A that allow the simultaneous use of electrical and non-electrical natures. Behavioral models of mechanical elements (e.g., beam and plate) and electromechanical elements (e.g., electrostatic gap) have been developed with separate mechanical and electrical terminals. The models are parameterized functions of the geometric and physical properties of the MEMS elements. The investigation of faulty MEMS behavior must ensure that “good” models preserve their accuracy when combined with fault models. In addition, it is desirable that the fault models do not degrade the performance of schematic-level simulation. Our approach so far avoids making new models for representing faulty behavior by using existing beam, electrostatic gap, anchor, and plate models to represent particulate defects discovered through Monte Carlo process simulations [9]. This approach, by definition, avoids the pitfalls described above. 3. Accelerometer Analysis The device used in our investigation is a single-axis accelerometer. The mechanical structures used in this accelerometer are constructed using a MEMS synthesis tool similar to the one described in [12]. It is synthesized for minimum noise and minimum area with the topology shown in Fig. 1(a) using the openly available MUMPs fabrication process [13]. A key parameter of the accelerometer is its resonant frequency f x in the sensitive direction x. Resonant frequency is a function of the physical and topological properties of the accelerometer and is therefore a good indicator for detecting the presence of defects. We use the resonant frequency deviation and the x-direction displacement to gauge the amount of misbehavior caused by particulate defects. However, other operational parameters such as sensitivity, both same-axis and cross-axis, can and should be used as well. The nominal value of the resonant frequency f x for our example design is 5.23 kHz. 3.1. Anchor Defects Particles that create unwanted anchors between suspended parts of the MEMS device and the substrate are called anchor defects. Conductive particles anchored under the beams/shuttle/fingers will short the sense signal to ground, which will obviously cause zero sensitivity. However, insulating particles will be less detrimental (and therefore less predictable) since their electrostatic effect will be confined to adding capacitance to the substrate. Our investigation is therefore focussed on the misbehavior resulting from insulating particles. Since the accelerometer topology possesses fourfold symmetry, a single quadrant of the design suffices for defect analysis. The displacement variation along the sensitive axis x is monitored using the three nodes N0, N1, and N2 shown in Fig. 1. Node N0 is located at the center of the shuttle mass, while nodes N1 and N2 are located at corners of two adjacent U-spring beams. Flexure Beam (FB1) This type of defect is modeled as two beams separated by an anchor as shown in Fig. 2. A particle that is spherical or ellipsoidal in shape has a small aspect ratio, and is therefore sufficiently stiff to be modeled as an anchor. Fig. 3(a) compares the variation of the resonant frequency f x with anchor defects located at various positions along flexure beam FB1 (Fig. 1). The High-Level Fault Modeling anchors fixed fingers 153 movable finger N2 shuttle mass N0 anchor FB2 TB Y FB1 N1 U spring Feedback fingers Sense fingers Feedback fingers X (a) (b) Fig. 1. Topology of a surface-micromachined accelerometer showing: (a) the layout with U-spring beams and electrostatic comb-drives for sensing and feedback, and (b) the corresponding NODAS schematic model. location of the particle is expressed as a percentage of beam length with the terminal anchor serving as the 0% point and the FB1-truss intersection corresponding to the 100% mark. The deviation of f x reported by NODAS matches closely to FEA predictions with the maximum deviation being only 0.5%. Fig. 3(b) illustrates the low-frequency displacement of the three observation nodes N0, N1 and N2 for a constant external acceleration of ax = 1 m/s2 . For all three nodes, the agreement between NODAS and FEA is again very good. It must be noted from Fig. 3 that the displacement of N0 agrees well with the typical behavior of lin- ear second-order systems such as the accelerometer. The relation x/ax = 1/(2π f x )2 holds true for the displacement x of N 0, external acceleration ax , and the resonant frequency f x of the structure. This assumes that for small, realistic displacements the good/faulty structure has negligible non-linearity. Flexure Beam (FB2) Fig. 4(a) compares the variation of f x due to anchor defects located at various positions along flexure beam FB2. The location of the particle is again expressed as 154 Deb and Blanton Fig. 2. NODAS schematic model of accelerometer anchor defects and broken-beam defects: Flexure Beam Anchor—A particle that causes an unwanted anchor under a flexure beam (lower right of the figure) is modeled using two smaller beam elements and an anchor element. Shuttle Anchor—A particle that causes an unwanted anchor under the shuttle mass (middle left of the figure) is modeled by simply adding an extra anchor element. Broken Beam—A beam broken into two unconnected parts (lower left of the figure) is modeled simply by two unconnected beams having total length equal to that of the unbroken beam. Fig. 3. FEA and NODAS comparison of the variation of (a) the principal resonant frequency f x , and (b) the low-frequency displacements of the nodes N0, N1, and N2 for an anchor defect located under flexure beam FB1 for an external acceleration of ax = 1 m/s2 . a percentage of the beam length with the FB2-truss intersection serving as the 0% point and the FB2-shuttle mass intersection corresponding to the 100% mark. The agreement between NODAS and FEA is reasonable and the maximum deviation is 1.7%. The location of the particle has been restricted to less than 30% of the length of FB2 because beyond this point the deviation from the specified behavior of the structure is so great that the fault is guaranteed to be catastrophic. Fig. 4(b) illustrates the low-frequency displacement of the three observation nodes N0, N1 and N2 for a constant external acceleration. For all three nodes, the agreement between NODAS and FEA is good and the maximum deviation is less than 4%. Note that the displacement High-Level Fault Modeling 155 Fig. 4. FEA and NODAS comparison of the variation of (a) the principal resonant frequency f x , and (b) the low-frequency displacements of the nodes N0, N1, and N2 for an anchor defect located under flexure beam FB2 for an external acceleration of ax = 1 m/s2 . Fig. 5. NODAS schematic model of (a) a movable finger affected by an anchor defect, and (b) a pair of adjacent fingers that are welded together due to an inter-finger particle defect. values predicted for node N2 by NODAS and FEA are almost zero for the range depicted. placement to be less than 2 × 10−15 m, which for all practical purposes is zero as well. Shuttle Movable Finger Displacement of node N0 in the x direction for an anchor defect located under the shuttle mass is predicted to be zero by NODAS. FEA results indicate the dis- Here we consider anchor defects that affect movable fingers. An anchor defect located under a movable finger is assumed to have little impact on lateral inter-finger 156 Deb and Blanton Table 1. FEA and NODAS comparison for the accelerometer with a movable finger affected by an anchor defect for an external acceleration of ax = 1 m/s2 . Nodal Displacements fx (kHz) N0 (pm) N1 (pm) N2 (pm) Location of Anchor Defect NODAS FEA NODAS FEA NODAS FEA NODAS FEA Under movable finger tip 9.05 8.89 311 326 160 168 183 180 Table 2. FEA and NODAS comparison for the accelerometer with a broken beam for an external acceleration of ax = 1 m/s2 . Nodal Displacements fx (kHz) N0 (pm) N1 (pm) N2 (pm) Location of Break NODAS FEA NODAS FEA NODAS FEA NODAS FEA FB1 FB2 4.48 4.50 4.50 4.53 1244 1235 1264 1247 701.1 695.9 676.5 667.4 1292 78.7 1343 1.5 and fringe capacitances. Hence, the electrostatic force of attraction is assumed to be unaffected by this defect type. To model this defect, we separate the lumped comb-drive model into two smaller combdrives. Fig. 5(a) shows how a movable finger affected by an anchor defect under its tip is modeled using good element models. The anchored middle beam represents the defective movable finger while the outer beams represent the corresponding fixed fingers. Each gap is modeled by an electrostatic gap element. Simulation results listed in Table 1 show that the accelerometer will suffer a catastrophic frequency deviation due to the defect. Defect locations that extend from the finger tip towards the finger base will cause spring stiffness to increase even more, further aggravating the impact of the fault. Hence, we have not tabulated the resonant frequency and displacement values for these cases. 3.2. Broken Beam Defect The NODAS schematic of Fig. 2 shows our model of an accelerometer with a broken beam. Table 2 lists the comparative simulation results obtained from NODAS and FEA. Both simulations show that resonant frequency is reduced as expected since a broken beam eliminates the stiffness of the U-spring structure. The small difference in f x for breaks in FB1 and FB2 is due to the difference in the effective masses rather than the spring constant change. For the break on FB2, the displacements predicted by NODAS and FEA for N2 do not agree well. Note that both values are quite low indi- cating negligible displacement (<6% of nominal). For all other nodes, NODAS agrees with FEA within 4%. 3.3. Inter-Finger Defect A particle can weld a movable finger to an adjacent fixed finger. If the connection is conductive, the resulting fault will definitely be catastrophic due to signal shorting. Our analysis is for defects made of dielectric material. We decouple the electrostatic and mechanical effects of the defect and model only the mechanical effect. The mechanical behavior of the particle is modeled as a moderately stiff truss beam located between the fixed and movable fingers which are themselves modeled as beams as shown in Fig. 5(b). Fig. 6 gives the variation of f x and the nodal displacements for a particle attached between a pair of movable and fixed fingers. The 0% point is the movable finger tip and the 100% point is the movable finger base. These plots, unlike the previous ones, are nonmonotonic and both display an extremum. This phenomenon is due to the formation of a spring from the welded fingers. As shown in Fig. 5(b), the spring consists of two flexure beams (FF1 and MF1) and one truss beam. This spring model is valid if the connecting particle acts as a sufficiently stiff truss and is much smaller compared to the lengths of FF1 and MF1. The total length of the two flexure beams, L1 + L2, is constant. If the flexure beam lengths are varied, the spring constant reaches its maximum value when the two flexure beam lengths are equal, i.e., L1 = L2. Fig. 6 supports High-Level Fault Modeling 157 Fig. 6. FEA and NODAS comparison of the variation of (a) the principal resonant frequency f x , and (b) the low-frequency displacements of the nodes N0, N1, and N2 for different weld locations for a particle positioned between movable and fixed fingers for an external acceleration of ax = 1 m/s2 . our intuition as the location of the extrema in the plots is approximately the 45% point. The extremum point p, expressed as a percentage of the movable finger length, is given by 1 L non-overlap p= 1− 2 L movable where L movable is the length of movable finger and L non-overlap is the portion of the fixed finger that does not overlap with the movable finger. Using the values of L movable = 302.3 µm and L non-overlap = 23 µm from our accelerometer design, we obtain p = 46.2%, which agrees well with the 45% point indicated in Fig. 6. 4. Conclusions and Future Work Our approach for high-level fault modeling uses NODAS for schematic-level simulation of MEMS affected by particulate contaminations. We have generated schematic-level fault models of particulate defects by modifying the nominal topology with existing models of beam, electrostatic gap, anchor, and plate models. We have demonstrated that NODAS is capable of simulating faulty MEMS with an accuracy comparable to FEA for faulty behaviors resulting from single particulate contaminations, namely, anchor defects, interfinger defects and broken beams. NODAS has a significant speed advantage over FEA. For example, a linear frequency response simulation of an accelerometer using FEA needs 230 s for 82 sample points. Under similar conditions of system load, the same workstation running NODAS can simulate 1800 sample points in 75 s. This is a speedup of more than 60X at the cost of reduced but acceptable accuracy. Our results allow us to claim that schematic-level MEMS fault modeling using NODAS is feasible and even preferable to lowlevel modeling methods like FEA due to the enormous reduction in simulation time achievable with minimal loss of accuracy. Schematic-level modeling of the electrostatic effects of particles located under and between fingers is our next major area of focus. Specifically, particles that fall in the gap without contacting both fingers can have a significant effect on the sensitivity of the device. Vertical stiction, a problem encountered frequently in surface-micromachined MEMS, and other sources of MEMS failure will also be investigated. The analysis of multiple failure sources in the presence of manufacturing variations (such as layout under/over-etch, layout curvature, side-wall angle, etc.) is also required to develop a comprehensive understanding of MEMS misbehavior. References 1. Mukherjee, T., Fedder, G. K. and Blanton, R. D., “Hierarchical design and test of integrated microsystems,” IEEE Design and Test of Computers, pp. 18–27, October–December 1999. 2. Hornbeck, L. J., “Current status of the digital micromirror device (DMD) for projection television application,” in Proc. of International Electron Devices Meeting, pp. 381–384, 1993. 158 Deb and Blanton 3. Payne, R. S., Sherman, S., Lewis, S. and Howe, R. T., “Surface micromachining: From vision to reality to vision,” in Proc. of International Solid State Circuits Conference, pp. 164–165, 1995. 4. Deb, N., Iyer, S. V., Mukherjee, T. and Blanton, R. D., “MEMS resonator synthesis for defect reduction,” Journal of Modeling and Simulation of Microsystems, pp. 11–20, 2001. 5. Kolpekwar, A., Blanton, R. D. and Woodilla, D., “Failure modes for stiction in surface-micromachined MEMS,” in Proc. of International Test Conference, pp. 551–556, October 1998. 6. Charlot, B., Mir, S., Cota, E. F., Lubaszewski, M. and Courtois, B., “Fault simulation of MEMS using HDLs,” in Symposium on Design Test and Microfabrication of MEMS/MOEMS, pp. 70– 77, March 1999. 7. Charlot, B., Moussouris, S., Mir, S. and Courtois, B., “Fault modeling of electrostatic comb-drives for MEMS,” in Proc. of International Test Conference, pp. 398–405, September 1999. 8. Castillejo, A., Veychard, D., Mir, S., Karam, J. M. and Courtois, B., “Failure mechanisms and fault classes for CMOS-compatible microelectromechanical systems,” in Proc. of International Test Conference, pp. 541–550, September 1998. 9. Jiang, T. and Blanton, R. D., “Particulate failures for surfacemicromachined MEMS,” in Proc. of International Test Conference, pp. 329–337, September 1999. 10. Abromovici, M., Breuer, M. A. and Friedman, A. D., Digital Systems Testing and Testable Design. IEEE Press, Piscataway, NJ, 1990. 11. Fedder, G. K. and Jing, Q., “A hierarchical circuit-level design methodology for microelectromechanical systems.” IEEE Transactions on Circuits and Systems II 46(10), pp. 1309–1315, October 1999. 12. Mukherjee, T., Iyer, S. and Fedder, G. K., “Optimization-based synthesis of microresonators.” Sensors and Actuators A70(1-2), pp. 118–127, October 1998. 13. Koester, D. A., Mahadevan, R. and Markus, K. W., MUMPs Introduction and Design Rules. MCNC MEMS Technology Applications Center, Cornwallis, NC, 1994. Nilmoni Deb received his B.Tech (Hons) degree from Indian Institute of Technology, Kharagpur, India, in 1997 and M.S. in Electrical and Computer Engineering from Carnegie Mellon University, Pittsburgh, USA, in 1999. Currently, he is a Ph.D. student in the Department of ECE, Carnegie Mellon University, with broad research interests in CAD for MicroElectroMechanical Systems with emphasis on Test for MicroElectroMechanical Systems. Shawn Blanton is an associate professor in the Department of Electrical and Computer Engineering at Carnegie Mellon University where he is a member of the Center for Silicon System Implementation. He received the Bachelor’s degree in Engineering from Calvin College in 1987, a Master’s degree in Electrical Engineering in 1989 from the University of Arizona, and a Ph.D. degree in Computer Science and Engineering from the University of Michigan, Ann Arbor in 1995. His research interests include the computer-aided design of VLSI circuits and systems; fault-tolerant computing and diagnosis; verification and testing; and computer architecture He has worked on the design and and test of integrated systems with General Motors Research Laboratories, AT&T Bell Laboratories, Intel, and Motorola. Dr. Blanton is the recipient of National Science Foundation Career Award and is a member of IEEE and ACM.