Abstract
Uncertainty quantification (UQ) based on generalized polynomial chaos (gPC) has achieved great success in variation-aware design automation. However, traditional techniques often assume that parameters are mutually independent or Gaussian correlated, which is rarely true in real-world chip manufacturing. This paper presents a new framework to solve UQ problems with many non-Gaussian correlated uncertainties. By proposing smooth basis functions, a tensor-based moment computation approach, and an adaptive sparse solver, the framework achieves significant efficiency gains—up to 3,333x faster than Monte Carlo simulations. Crucially, it accurately predicts output density functions with multiple peaks caused by non-Gaussian correlations, which are typically missed by existing methods.
Introduction: The Challenge of Real-World Variations
In modern semiconductor manufacturing, "process variations"—fluctuations in parameters like transistor gate thickness or laser wavelengths—significantly impact performance. To design reliable chips, engineers use Uncertainty Quantification (UQ) to predict how these variations affect the final product. While standard tools work well for independent or Gaussian variations, real-world manufacturing data is often non-Gaussian and highly correlated because multiple components are affected by the same fabrication steps.
Existing tools for these "messy" cases struggle with the curse of dimensionality—as the number of parameters grows, the computational cost explodes, making them impractical for complex chips with dozens or hundreds of variables.
Methodology
The authors tackle the problem through a three-pronged strategy:
- Smooth Multivariate Basis Functions: They move away from standard polynomials to a new set of basis functions built via a Cholesky factorization of the moment matrix. These functions are smooth, orthonormal, and numerically stable for non-Gaussian inputs.
- Functional Tensor Train (TT) Decomposition: Computing high-order moments in high dimensions is usually an integration nightmare. The authors use a "Tensor Train" approach to break these complex multivariate integrals into a sequence of simple, small-scale matrix-vector products, allowing the method to scale to high dimensions without exponential cost.
- Adaptive Sparse Solver: Instead of running thousands of expensive circuit simulations, they use Compressed Sensing to find the most important coefficients using only a handful of samples. They introduce adaptive sampling (D-optimal, R-optimal, and E-optimal criteria) to intelligently pick the most informative simulation points.
Key Results and Performance
The framework was validated on benchmarks ranging from 19 to 57 non-Gaussian correlated parameters, including photonic filters and CMOS ring oscillators.
- Massive Speedups: For a 57-parameter CMOS Ring Oscillator, the method achieved the same accuracy as Monte Carlo but required 3,333x fewer simulation samples.
- Capturing Multi-modality: Non-Gaussian correlations often create "multi-modal" outputs (distributions with multiple peaks). As shown in Figure 6(d), the proposed framework captures these peaks precisely, matching the "ground truth" Monte Carlo results almost perfectly while using a fraction of the compute time.
- Efficiency: For a Photonic Band-pass Filter, the method reached high precision with just 320 samples, whereas Monte Carlo would have required over a million.
Conclusion
This research provides a scalable and efficient solution for UQ in the presence of complex process variations. By moving beyond Gaussian assumptions and utilizing tensor-based acceleration and sparse recovery, the framework enables high-fidelity performance prediction for modern electronic and photonic ICs. Future work will explore even higher-dimensional systems and the integration of these techniques directly into automated design optimization loops.
Summary updated