Machine learning-enhanced elmore delay modeling under process and environmental variations for robust VLSI timing analysis
This paper introduces a variation-conscious Elmore delay modeling framework with the data. The dataset was generated through SPICE simulations combined with statistically modeled process, voltage, and temperature (PVT) variations. Realistic PVT variations are considered in this process. Resistors (R 1 , R 2 ) and capacitor (C 1 , C 2 ) values were modeled as log-normal with resistance depending on temperature (0.0039/°C) and capacitance depending on voltage (capacitance scaling). A 10-dimensional feature space incorporated in RC cross-products. Raw parameters were trained on 5000 samples to a multivariate linear regression model. Our model achieved R² score of 0.987 and RMSE of 37.6 ps. These results indicate high-fidelity delay prediction across the considered PVT operating conditions. The R 1 *C 1 and R 1 *C 2 terms are used for the analysis of coefficients and sensitivity. The Monte Carlo simulation of 1000 runs gave a mean delay of 496.79 ns with the standard deviation of 54.11 ns and 3σ of ±162.34 ns. The method offers a small, interpretable, and precise delay forecasting framework. This is used in statistical, static timing analysis (SSTA) and PVT circuit optimization.
Authors
- Mandar Jatkar (ORCID: https://orcid.org/0000-0003-2109-9487)
- Nandini Prasad K S
- Vijay Deshpande
Institutions
- Honeywell (India) (IN)
- Dayananda Sagar Academy of Technology and Management (IN)
Publication Details
- Journal
- Next Energy
- Published
- 2026-11-20
- DOI
- https://doi.org/10.1016/j.nxener.2026.101203
- Primary Topic
- Low-power high-performance VLSI design
- Type
- article
- Field-Weighted Citation Impact
- 0.00