Conformal Uncertainty-Aware Physics-Informed Neural Networks for Reliable Electromagnetic Field Prediction in VLSI Interconnects
Academic Journal of Applied Sciences · 21.07.2026 · 2 (2) · 126-132
DOI: 10.54097/56shkz94 ↗Kurz gefasst
Fachartikel in der Zeitschrift Academic Journal of Applied Sciences, online erschienen am 21.07.2026. Verfasst von Sofia Duarte und 1 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY-NC 4.0).
Abstract (Originalsprache)
Electromagnetic field prediction in very-large-scale integration (VLSI) interconnects is increasingly constrained by two competing requirements: field solvers must remain physically faithful near conductor edges and material interfaces, yet they must also provide calibrated reliability indicators for fast design-space exploration. This paper proposes a conformal uncertainty-aware physics-informed neural network (UQ-PINN) for quasi-static electromagnetic field approximation in a two-layer interconnect cross-section. The method combines a random-feature neural representation with boundary, sparse observation, and electroquasistatic residual constraints, and then converts ensemble uncertainty into finite-sample prediction intervals using split conformal calibration. The scope is deliberately reproducible: no measured silicon data are claimed. A deterministic finite-difference solution of div (epsilon grad phi)=0 is used as the reference benchmark, and all reported values are generated by the accompanying code. On the benchmark, the physics-informed ensemble obtains a potential MAE of 0.02057 V and an electric-field magnitude relative L2 error of 17.69%, improving the sparse data-only neural baseline. The conformalized interval achieves 90.60% empirical coverage for a 90% target, whereas the uncalibrated ensemble interval reaches only 66.63%. These results indicate that physics constraints and conformal calibration can jointly support reliable electromagnetic surrogates for early-stage EDA analysis without overstating the substitute for sign-off full-wave verification.
Zitieren
APA
Duarte, S., & Hart, B. (2026). Conformal Uncertainty-Aware Physics-Informed Neural Networks for Reliable Electromagnetic Field Prediction in VLSI Interconnects. Academic Journal of Applied Sciences, 2(2), 126-132. https://doi.org/10.54097/56shkz94
Vancouver
Duarte S, Hart B. Conformal Uncertainty-Aware Physics-Informed Neural Networks for Reliable Electromagnetic Field Prediction in VLSI Interconnects. Academic Journal of Applied Sciences. 2026;2(2):126-132. doi:10.54097/56shkz94
BibTeX
@article{Duarte2026,
author = {Duarte, Sofia and Hart, Benjamin},
title = {Conformal Uncertainty-Aware Physics-Informed Neural Networks for Reliable Electromagnetic Field Prediction in VLSI Interconnects},
journal = {Academic Journal of Applied Sciences},
year = {2026},
volume = {2},
number = {2},
pages = {126--132},
doi = {10.54097/56shkz94}
}
EUVIMED