TL;DR
This paper introduces a novel approach to analyze how manufacturing imperfections (process variations) in MEMS devices can lead to unpredictable performance. They combine probabilistic (Monte Carlo) and deterministic (robust optimization) methods, integrated into the SUGAR MEMS simulator, to provide designers with a comprehensive understanding of potential performance deviations. This is crucial for ensuring reliable MEMS device functionality in real-world applications.
What Problem Does This Paper Address?
MEMS (Micro-Electro-Mechanical Systems) are increasingly complex, with millions of devices fabricated using fine-line lithography. However, the fabrication process is not perfect, leading to variations in the physical dimensions of these tiny structures. These geometric variations, even if small, can significantly alter the device's performance from its intended "nominal" design. Existing MEMS design tools often struggle to account for these uncertainties, leaving designers without a clear picture of the potential performance spread. This paper aims to bridge this gap by providing methods to estimate and bound these performance variations, allowing for more robust and reliable MEMS designs.
Methods
The authors propose a two-pronged approach, integrated into the SUGAR MEMS simulation tool:
- Probabilistic Approach (Monte Carlo): This method treats geometric variations as random variables with defined probability distributions (uniform, Gaussian, or uniform corner). By sampling these distributions many times, the simulator estimates the range and likelihood of different performance outcomes. This is computationally intensive but provides a statistical understanding of performance spread.
- Deterministic Approach (Robust Optimization/Ellipsoidal Calculus): This technique focuses on finding the "worst-case" performance by considering bounds on geometric variations. It uses semidefinite programming (SDP) approximations to compute an "ellipsoid of confidence" that encloses all possible equilibrium points of the system. This approach is particularly useful for safety-critical applications where guaranteed bounds are essential.
The core of the modeling involves representing MEMS structures as interconnected atomic elements (like beams) described by nonlinear second-order dynamical systems. For DC equilibrium analysis, the mass and damping terms are neglected, simplifying the system to a stiffness matrix dependent on design variables and displacements. The paper details how to derive this stiffness matrix and represent the uncertainty in an affine form suitable for SDP.
Datasets
Not discussed in the paper. The authors use simulated MEMS structures (dog-bone suspension, folded spring) as test cases.
Key Insights and Ideas
- Complementary Nature of Approaches: The Monte Carlo and robust optimization methods are not mutually exclusive but rather complementary. One provides statistical insights, while the other offers guaranteed worst-case bounds, catering to different design needs.
- The "Worst Case" is Key: For safety-critical applications, understanding the absolute worst possible performance deviation is paramount. The robust optimization approach directly addresses this by finding bounds on performance parameters.
- Affine Representation for SDP: A crucial step for applying SDP is representing the uncertainty in the system matrix (stiffness matrix) in an affine form. The paper shows how to achieve this for beam elements by considering variations in area and moment of inertia.
- Ellipsoidal Calculus for Bounds: The concept of an "ellipsoid of confidence" is a powerful visualization and analytical tool for understanding the set of all possible equilibrium points due to variations.
- Trade-offs in Robustness: The ellipsoidal calculus can sometimes be "overconservative," meaning the calculated bounds might be wider than necessary, leading to designs that are more robust than strictly required.
Key Results
The paper demonstrates the application of both Monte Carlo and ellipsoidal calculus methods on simulated MEMS structures (dog-bone suspension and folded spring).
- Monte Carlo: The simulations show that even for symmetric structures, process variations can lead to non-zero displacements in unintended directions (e.g., x-axis displacement in a y-axis loaded dog-bone suspension). Different distributions (uniform, Gaussian, uniform corner) yield different patterns of these variations.
- Ellipsoidal Calculus: The method successfully computes confidence ellipsoids that enclose the sampled performance outcomes. The paper notes that the ellipsoid can be tight (closely fitting the data) or overpessimistic, depending on the structure and the distribution used. It also highlights that the ellipsoid approach captures correlations between different displacement variables.
Limitations and Future Work
- Computational Cost: The Monte Carlo method can become computationally intractable for a large number of variation variables, as the number of scenarios scales exponentially. The SDP solver for the ellipsoidal calculus, while polynomial, has significant overhead.
- Over-conservatism: The ellipsoidal calculus might provide overly conservative bounds, potentially leading to designs that are more robust than necessary.
- Independent Variables Assumption: The paper acknowledges that assuming variations in area and moment of inertia are independent is an approximation that can lead to infeasible solutions and thus over-pessimistic ellipsoids.
- Future Work:
- Introducing different geometric bounds for individual variables.
- Analyzing variations in material properties (Young's modulus, Poisson ratio).
- Extending the analysis to 3D structures.
- Developing more efficient SDP solvers or alternative methods for the ellipsoidal calculus.
- Extending the ellipsoidal calculus to design synthesis, i.e., tuning nominal dimensions to improve performance despite variations.
Summary updated