Kirchoff Law Markov Fields for Analog Circuit Design

Richard M. GoldenNeurIPS 1999

Overview

This paper introduces MRFSPICE, a novel framework that treats analog circuit design as a probabilistic inference problem. By embedding Kirchhoff’s Current Law (KCL) directly into the "energy function" of a Markov Random Field (MRF), the authors created an automated tool capable of optimizing complex, nonlinear circuit parameters while providing a clear "confidence factor" for the resulting design.

The Challenge of Analog Circuit Design

Designing analog circuits is notoriously difficult because they are highly nonlinear and sensitive to small changes. The industry standard, SPICE, is excellent at simulating a known circuit, but it struggles with the inverse problem: finding the right component values among thousands of possibilities to achieve a desired behavior.

Current optimization tools in SPICE often fail when faced with too many free parameters or non-continuous constraints. Furthermore, traditional tools lack a mechanism to communicate how "sure" they are about a solution; they simply output a result. The author, Richard M. Golden, sought a way to incorporate physical laws (like KCL) as soft constraints that guide an automated search toward an optimal, physically consistent design.

The MRFSPICE Methodology

The core innovation of this research is the MRFSPICE algorithm, which bridges Markov Random Fields with traditional circuit analysis through four key steps:

  1. Probabilistic Modeling: The design is treated as a set of discrete random variables (e.g., choosing from 10 specific capacitor values or 20 transistor types). The goal is to maximize a global probability function, $p_G(x)$, which represents the "preference" for a specific design configuration.
  2. KCL as Potential Functions: In an MRF, "potential functions" usually represent local constraints. Here, the author defines these functions based on Kirchhoff’s Current Law. If the sum of currents at a node isn't zero, the "energy" of the system rises, and the probability of that design drops.
  3. The Normalized Current Ratio: To prevent the optimizer from simply choosing components that result in zero current everywhere (a trivial, non-functional way to satisfy KCL), the paper introduces a normalized ratio of average current to current variability. This ensures the circuit is functionally active while remaining physically valid.
  4. Optimization Strategy: The algorithm uses a two-step approach:
  5. Metropolis Algorithm: A stochastic search to sample the design space and find a robust starting point.
  6. Iterated Conditional Modes (ICM): A deterministic "greedy" optimizer that quickly drives the design to a local maximum of the probability function.

Experimental Setup: The JFET Problem

Rather than using a traditional external dataset, the paper evaluates the method using a specific JFET (Junction Field Effect Transistor) design problem.

Circuit diagram for the JFET design problem evaluated using MRFSPICE.

The "ground truth" consists of target characteristic curves (drain-to-source current vs. gate/drain voltage) at both DC (0 Hz) and 1 MHz. The design space included five variables: two diode models (20 choices each), one JFET parameter set (20 choices), and two resistors (15 choices each).

Key Insights

  • Physics as Energy: The most profound idea is treating a fundamental physical law (KCL) not just as a hard requirement, but as a mathematical landscape (a potential function) that an algorithm can "roll" down to find a solution.
  • The Circuit Confidence Factor (CCF): Because the model is probabilistic, it yields a CCF between 0 and 1. A low CCF tells the engineer exactly where the design is "weak," allowing for human-in-the-loop adjustments.
  • Beyond Quadratic Functions: Unlike previous neural network approaches to circuits (like the "Harmony" or "Brain-State-in-a-Box" models) which relied on simple quadratic math, MRFSPICE handles the highly nonlinear equations required for real-world semiconductors.

Key Results

The MRFSPICE algorithm successfully "discovered" component values that matched the desired JFET characteristic curves with high precision.

Comparison of desired and actual currents at DC and 1 MHz.
  • Accuracy: In the test cases, the actual current produced by the generated design was nearly identical to the desired current. For example, at a specific voltage, the target was 1.47 mA and the algorithm achieved 1.50 mA.
  • AC Performance: The model held up at high frequencies (1 MHz), with the observed phase angle (7–9 degrees) closely matching the specified target (~10 degrees).
  • Efficiency: The prototype, written in MATLAB, solved the problem in 2–4 hours on a 200 MHz Pentium processor—a respectable speed for the late 90s given the complexity of the search space.

Limitations and Future Work

While promising, the framework has a few notable limitations:

  • Local Maxima: The fast heuristic used (ICM) is guaranteed to find a local maximum, but not necessarily the global best design. While simulated annealing could resolve this, it would significantly increase computation time.
  • Scalability: The test problem involved only five design variables. The paper does not demonstrate how the MRF approach scales to massive integrated circuits with thousands of components.
  • Discrete Constraints: The model relies on choosing from a finite set of component values. While this mimics how engineers pick parts from a catalog, it may be less flexible for custom silicon design where parameters can be continuous.

Summary updated

BibTeX
@inproceedings{golden1999kirchoff,
  author = {Richard M. Golden},
  title = {{Kirchoff Law Markov Fields for Analog Circuit Design}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {1999},
}
Figure