EUVIMEDEUVIMED entdecken
Open Access · CC BY-NC 4.0

Multi-Fidelity Physics-Informed Neural Networks for Fast Electromagnetic Simulation of RF Integrated-Circuit Passives

Marina Petrova, Leo Fernandez

Journal of Computing and Electronic Information Management · 20.07.2026 · 22 (1) · 19-24

DOI: 10.54097/q4brzy12 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of Computing and Electronic Information Management, online erschienen am 20.07.2026. Verfasst von Marina Petrova und 1 weiteren Autor:innen. Die Arbeit nennt 31 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Full-wave electromagnetic (EM) analysis of radio-frequency (RF) integrated-circuit passives is accurate but too slow for the many-query loops of modern design automation, whereas quasi-static approximations are fast but miss frequency-dependent loss and dispersion. We present a composite multi-fidelity physics-informed neural network (MF-PINN) that fuses an abundant low-fidelity (LF) quasi-static model with a small set of high-fidelity (HF) full-model samples to build a fast, accurate frequency-domain S-parameter surrogate over a microstrip design space. The surrogate is regularized by two physical constraints derived from network theory—passivity (energy conservation) and non-negative group delay (causality)—evaluated by automatic and finite-difference differentiation on unlabeled collocation points. Ground-truth data are generated with a validated open-source microstrip model (Hammerstad-Jensen, Kirschning-Jansen, skin-effect loss). On 30 held-out geometries the MF-PINN attains 3.37% relative L2 error (0.11 dB magnitude, 0.55° phase) using only eight HF components, versus 13.17% for a single-fidelity network with the same HF budget. A data-efficiency study shows multi-fidelity training reaches the accuracy of single-fidelity models with 4–8× fewer HF samples, and the physics constraints reduce the mean passivity violation by two-to-three orders of magnitude. Once trained, the surrogate evaluates S-parameters 41× faster than the reference solver. We also report a negative result: aggressive Fourier-feature encoding, useful for oscillatory PDEs, degrades accuracy here because the RF response is smooth in the band, so a compact multilayer perceptron is preferable.

Zitieren

APA

Petrova, M., & Fernandez, L. (2026). Multi-Fidelity Physics-Informed Neural Networks for Fast Electromagnetic Simulation of RF Integrated-Circuit Passives. Journal of Computing and Electronic Information Management, 22(1), 19-24. https://doi.org/10.54097/q4brzy12

Vancouver

Petrova M, Fernandez L. Multi-Fidelity Physics-Informed Neural Networks for Fast Electromagnetic Simulation of RF Integrated-Circuit Passives. Journal of Computing and Electronic Information Management. 2026;22(1):19-24. doi:10.54097/q4brzy12

BibTeX
@article{Petrova2026,
  author = {Petrova, Marina and Fernandez, Leo},
  title = {Multi-Fidelity Physics-Informed Neural Networks for Fast Electromagnetic Simulation of RF Integrated-Circuit Passives},
  journal = {Journal of Computing and Electronic Information Management},
  year = {2026},
  volume = {22},
  number = {1},
  pages = {19--24},
  doi = {10.54097/q4brzy12}
}
Kostenlos starten →