Surrogate-Assisted Framework for SI-Compliant Interconnect Design Optimization Using the Earth Mover's Distance

This paper introduces a deterministic, interpretable framework for optimizing signal-integrity-compliant PCB interconnects by combining neural surrogate models for waveform prediction, decision trees for physical quality gating, and the Earth Mover's Distance for ranking designs against an ideal reference signal, thereby offering a transparent and efficient alternative to conventional stochastic optimization methods.

The Gist

Imagine you are an architect trying to design the perfect highway for tiny electrical signals (like data) traveling across a computer chip. If the road is too bumpy, too wide, or has too many sharp turns, the signals get distorted, and the computer crashes. This is called a "Signal Integrity" problem.

Traditionally, finding the perfect road design is like trying to find a needle in a haystack by throwing darts in the dark. You guess a design, test it, see if it fails, guess again, and repeat thousands of times. It's slow, expensive, and often feels like a "black box" where you don't know why a design worked or failed.

This paper presents a new, smarter way to design these roads using a "Surrogate-Assisted Framework." Think of it as replacing the dart-throwing with a highly organized, step-by-step detective agency. Here is how it works, broken down into simple steps:

1. The Crystal Ball (Neural Surrogate Models)

First, the team builds a "Crystal Ball" (a type of AI called a neural network). Instead of building a physical road and testing it (which takes a long time), this Crystal Ball looks at the blueprint (the design parameters like wire length and resistor size) and instantly predicts what the road will look like.

• The Analogy: Imagine a master chef who can taste a raw ingredient list and instantly describe exactly how the final dish will taste, without actually cooking it. This saves hours of cooking time.

2. The Strict Gatekeeper (Decision Tree)

Once the Crystal Ball predicts the road shape, the design goes to a "Gatekeeper" (a Decision Tree). This Gatekeeper doesn't guess; it follows a clear, logical checklist of rules (like a traffic cop).

• The Analogy: Think of a bouncer at a club. The bouncer checks your ID against a specific list. If you meet the criteria (no overshoot, fast enough speed), you get in. If not, you are turned away.
• Why it's special: Unlike the old "dart-throwing" methods that just say "this is the best," this Gatekeeper explains why a design is good or bad. It's a "white box" approach, meaning you can see exactly what rules were applied.

3. The "Moving Boxes" Score (Earth Mover's Distance)

Now, imagine we have a list of designs that passed the Gatekeeper. They are all "safe," but which one is the best? To decide, the team uses a metric called the Earth Mover's Distance (EMD).

• The Analogy: Imagine you have a pile of dirt (your actual signal shape) and you want to move it to match a perfect, flat square of dirt (the ideal signal). The EMD calculates the exact amount of "work" (energy) it would take to move the dirt from your pile to the perfect square.
• The Result: The design that requires the least amount of work to look like the perfect square is the winner. It's not just about passing; it's about being the closest to perfection.

4. The Final Result: Speed and Clarity

The paper tested this system on real computer memory designs (DDR3).

• Speed: The new method found the best designs in less than a second. A traditional method (called a Genetic Algorithm) took over a minute to find a similar result.
• Safety: When the team asked the system to look for designs in a "bad" area (where no good roads could exist), the system immediately said, "Nope, nothing works here." The traditional method, however, kept guessing and eventually picked a bad design anyway because it didn't have a Gatekeeper to stop it.

Summary

In short, this paper introduces a system that:

1. Predicts outcomes instantly using AI (no more slow physical testing).
2. Filters designs using clear, logical rules (no more black boxes).
3. Ranks the good designs by how closely they match perfection (using the "moving dirt" math).

The result is a faster, clearer, and more reliable way to design computer circuits, ensuring that the signals travel smoothly without the designer having to guess in the dark. The authors note that while this works well for current technology (DDR3), future versions will need to adapt for newer, faster standards like DDR4 and DDR5.

Technical Summary

Technical Summary: Surrogate-Assisted Framework for SI-Compliant Interconnect Design Optimization Using the Earth Mover's Distance

Problem Statement
Modern Electronic Design Automation (EDA) increasingly relies on Machine Learning (ML) to accelerate Printed Circuit Board (PCB) design, particularly for Signal Integrity (SI) compliance. While conventional surrogate-based optimization methods utilize iterative black-box search procedures (e.g., Genetic Algorithms or Bayesian Optimization) to find optimal parameters, they often lack interpretability and provide limited insight into the underlying physical decisions. Furthermore, these methods typically focus on convergence to a single optimal point without refining the prediction model during the search, and they may struggle with the non-uniqueness issues inherent in inverse modeling. The authors propose a deterministic, interpretable alternative that avoids iterative search and inverse modeling, aiming to support designers in identifying and ranking physically superior SI-compliant solutions within a defined parameter space.

Methodology
The proposed framework follows a sequential, forward-directed logic that integrates three distinct components: neural surrogate models for prediction, decision trees for interpretability, and the Earth Mover's Distance (EMD) for physically grounded evaluation.

1. Data Generation and Feature Engineering:

   • A large-scale dataset was generated using IBIS models and the eCadstar SI simulation tool for DDR3 fly-by networks (single and dual DRAM configurations), totaling over 1.1 million extracted waveforms.
   • Instead of using raw circuit parameters (trace lengths, termination resistances) directly, a physics-inspired feature engineering script transforms these inputs into a larger set of physically meaningful descriptors (e.g., reflection coefficients, harmonic Fourier features, geometric relations). This expands the input dimensionality to $d=62$ for single-module and $d=92$ for dual-module setups to support the learning of complex nonlinear relationships.

2. Neural Surrogate Models (MLP):

   • Waveform Descriptor Prediction: A deep residual Multi-Layer Perceptron (MLP) with Layer Normalization and GELU activation functions predicts waveform descriptors from the engineered input features. This model acts as a deterministic feature generator, replacing computationally expensive field solvers.
   • Wasserstein Distance Prediction: A second, simpler MLP surrogate predicts the Earth Mover's Distance (Wasserstein distance, $W_1$) between the predicted waveform and an ideal rectangular reference signal. This allows for rapid evaluation of signal quality without re-simulation.

3. Interpretable Classification (Decision Tree):

   • A CART-based decision tree acts as a "quality gate." It is trained on the waveform descriptors to classify signals as SI-compliant or non-compliant based on predefined physical criteria (overshoot limits, edge slope, logic levels).
   • To address class imbalance, a cost matrix is employed to weight false negatives (valid designs misclassified as invalid) more heavily. The tree structure provides a transparent, rule-based explanation of which parameter regions yield valid signals.

4. Optimization and Ranking via EMD:

   • Unlike traditional methods that search for a single optimum, this framework first identifies the valid solution space using the decision tree.
   • Within this valid space, the EMD (Wasserstein metric) is used as a similarity metric to rank candidate designs. The EMD quantifies the "work" required to transform a given signal shape into an ideal reference shape.
   • The final output is a ranked list of SI-compliant design parameters, prioritized by their physical proximity to the ideal signal, without requiring inverse modeling or stochastic search.

Key Contributions

• Deterministic and Interpretable Framework: The paper introduces a workflow that avoids iterative black-box optimization. By using a decision tree as a quality gate, the method offers a "white box" approach where the logic for SI compliance is transparent and explainable.
• EMD-Based Differentiation: The integration of the Earth Mover's Distance allows for a physically motivated ranking of valid solutions. This moves beyond merely identifying "safe" regions (as in previous statistical approaches) to differentiating the quality of waveforms based on their morphological similarity to an ideal reference.
• Physics-Inspired Feature Engineering: The transformation of raw design parameters into physically meaningful descriptors significantly enhances the generalization capability of the surrogate models, enabling high-precision predictions of complex waveform characteristics.
• Avoidance of Inverse Modeling: The framework operates strictly in the forward direction (parameters $\to$ waveforms), structurally excluding the non-uniqueness issues and statistical uncertainty models often required by inverse approaches.

Results
The methodology was validated using a dataset of DDR3 fly-by networks:

• Surrogate Accuracy: The residual MLP achieved high precision in predicting waveform descriptors, with median Normalized Root Mean Squared Error (NRMSE) values below 1.0% and Pearson Correlation Coefficients (PCC) of 0.99 across training and test sets.
• Classification Performance: The decision tree demonstrated robust generalization, achieving sensitivities of 98.4–99.3% and specificities of 98.2–99.5% on test data, with Matthews Correlation Coefficients (MCC) exceeding 0.85.
• Optimization Efficiency:
  • In a single-DRAM scenario with 17,384 parameter combinations, the framework identified the top 5 optimal designs in 0.43 seconds, compared to 78.04 seconds for a Genetic Algorithm (GA).
  • Crucially, when presented with an invalid parameter space (where no SI-compliant solutions exist), the framework correctly identified the absence of valid solutions in 0.29 seconds. In contrast, the GA returned a "best" candidate that was physically non-compliant, highlighting the framework's ability to act as a reliable validity gate.
• Validation: Re-simulations of the top-ranked designs confirmed that the selected parameters produced waveforms with adequate rectangular shapes, minimal overshoot, and maximized rise times, consistent with the EMD ranking.

Significance and Claims
The authors claim that this framework provides a computationally efficient and explainable alternative to conventional optimization strategies for PCB development. By combining full design space coverage (via parameter sampling), the transparency of decision trees, and the physical precision of the EMD metric, the method enables designers to:

1. Deterministically identify admissible parameter regions.
2. Transparently prioritize physically superior solutions without relying on stochastic search or inverse modeling.
3. Avoid the "black box" nature of traditional optimizers, thereby offering greater insight into the physical decisions governing signal integrity.

The paper positions this work as a step toward more interpretable AI in engineering, shifting the focus from mere convergence to the physical assessment and differentiation of design solutions. The authors note that while the current implementation focuses on DDR3, the methodology is intended to be extensible to future standards like DDR4 and DDR5, potentially adapting to metrics like eye diagrams and bit-error rates as transient waveform analysis becomes less practical for higher data rates.