A Robust algorithm for the extraction of reduced-order behavioral models from sampled frequency responses is proposed. The system under investigation can be any Linear and Time Invariant structure, although the main emphasis is on devices that are relevant for Signal and Power Integrity and RF design, such as electrical interconnects and integrated passive components. We assume that the device under modeling is parameterized by one or more design variables, which can be related to geometry or materials. Therefore, we seek for multivariate macromodels that reproduce the dynamic behavior over a predefined frequency band, with an explicit embedded dependence of the model equations on these external parameters. Such parameterized macromodels may be used to construct component libraries and prove very useful in fast system-level numerical simulations in time or frequency domain, including optimization, what-if, and sensitivity analysis. The main novel contribution is the formulation of a finite set of convex constraints that are applied during model identification, which provide sufficient conditions for uniform model stability and passivity throughout the parameter space. Such constraints are characterized by an explicit control allowing for a trade-off between model accuracy and runtime, thanks to some special properties of Bernstein polynomials. In summary, we solve the longstanding problem of multivariate stability and passivity enforcement in data-driven model order reduction, which insofar has been tackled only via either overconservative or heuristic and possibly unreliable methods.
29 November 2018 in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
We introduce the concept of Bounded Input Dissipativity (BID) for the characterization from an energy perspective of linearized models of nonlinear circuit blocks. Such linearized models are commonly employed in the design of large systems to approximate circuit blocks that operate in the neighborhood of some well-defined and asymptotically stable bias point and lead to a dramatic reduction in circuit simulation runtime. Even if the original circuit block always behaves as a dissipative system, its linearized model may behave as dissipative or non-dissipative depending on the amplitude of its small-signal port voltages and currents, compared to the corresponding constant bias or supply signal components. This paper presents a theoretical framework for the analysis of such a conditional dissipativity and introduces BID criteria based on the feasibility of Bilinear or Linear Matrix Inequalities. These criteria allow to estimate the maximum small-signal input amplitude that guarantees the dissipativity of the linearized model. Various examples demonstrate the validity of the proposed theory, including a test case showing how spurious and physically inconsistent instabilities may arise if the proposed BID conditions are not verified.
27 July 2020 in IEEE Transactions on Components, Packaging and Manufacturing Technology
Reduced-order models are widely used to reduce the computational cost required by the numerical assessment of electrical performance during the design cycle of electronic circuits and systems. Although standard macromodeling algorithms can be considered to be well consolidated, the generation of macromodels that embed in a closed form some dependence on the design variables still presents considerable margins for improvement. One of these aspects is enforcement of uniform stability throughout the parameter space of interest. This paper proposes a novel parameterized macromodeling strategy, which enforces by construction that all macromodel poles are stable for any combination of possibly several independent design variables. The key enabling factor is adoption of positive definite multivariate basis functions for the representation of model variations induced by the parameters. This representation leads to robust model generation from tabulated frequency responses, at a computational cost that is dramatically reduced with respect to competing approaches. This result arises from a number of algebraic constraints for stability enforcement that depends on the model complexity (number of basis functions) and not on the model behavior as a function of the parameters. As a byproduct, the proposed strategy lends itself to much improved scaling with the dimension of parameter space, allowing to circumvent the curse of dimensionality that may occur when the number of independent parameters grows beyond few units. To this end, we exploit representations based on positive definite radial basis functions. The benefits of the proposed approach are demonstrated through an extensive experimental campaign applied to both passive and active devices and components, comparing the performance of different model parameterizations in terms of accuracy, time requirements and model compactness.
24 February 2025 in Journal of Scientific Computing
In recent years, the Adaptive Antoulas-Anderson () algorithm has been applied extensively for generating Reduced Order Models (ROMs) of large scale Linear Time-Invariant systems starting from measurements of their transfer functions. When used for Model Order Reduction (MOR) of an asymptotically stable system, the ROMs generated applying are not guaranteed to preserve this fundamental property, thus rendering them impractical in many relevant applications. To overcome this issue, we propose a novel algebraic characterization for the stability of ROMs represented under the barycentric structure. We then translate such characterization into a set of convex semidefinite constraints that can be embedded into the optimization routine to explicitly maximize the model accuracy while ensuring its stability. Finally, we generalize the resulting modeling framework to allow for efficient stable MOR of Multi-Input-Multi-Output systems. An extensive set of numerical experiments provides practical evidence for the effectiveness of the proposed approach and compares its performance with those of available state-of-the-art methods.
26 July 2021 in IEEE Transactions on Components, Packaging and Manufacturing Technology
This paper proposes a black-box behavioral modeling framework for analog circuit blocks operating under small-signal conditions around non-stationary operating points. Such variations may be induced either by changes in the loading conditions or by event-driven updates of the operating point for system performance optimization, e.g., to reduce power consumption. An extension of existing data-driven parameterized reduced-order modeling techniques is proposed that considers the time-varying bias components of the port signals as non-stationary parameters. These components are extracted at runtime by a lowpass filter and used to instantaneously update the matrices of the reduced-order state-space model realized as a SPICE netlist. Our main result is a formal proof of quadratic stability of such Linear Parameter Varying (LPV) models, enabled by imposing a specific model structure and representing the transfer function in a basis of positive functions whose elements constitute a partition of unity. The proposed quadratic stability conditions are easily enforced through a finite set of small-size Linear Matrix Inequalities (LMI), used as constraints during model construction. Numerical results on various circuit blocks including voltage regulators confirm that our approach not only ensures the model stability, but also provides speedup in runtime up to 2 orders of magnitude with respect to full transistor-level circuits.