End-to-End Neural Modeling of EM Response and Design Performance for Free-Form RFIC Passives

Yuhao Mao; Chenhao Chu; Martin Vechev; Hua Wang

eess.SPOct 25, 2026

End-to-End Neural Modeling of EM Response and Design Performance for Free-Form RFIC Passives

Authors: Yuhao Mao, Chenhao Chu, Martin Vechev, Hua Wang

Organizations: Department of Computer Science, ETH Zurich, Switzerland · Department of Information Technology and Electrical Engineering, ETH Zurich, Switzerland

Abstract

Passive matching networks are critical components in radio-frequency integrated circuits (RFICs), but their design is bottlenecked by expensive finite-difference-based electromagnetic (EM) simulation. We propose an end-to-end neural framework for free-form RFIC passives that predicts both EM response and downstream design performance under varying frequency and circuit contexts, combining a frequency-conditioned neural simulator with a circuit-conditioned neural solver. To improve design performance prediction, the simulator outputs a neural EM response consisting of both the explicit EM response and a complementary response, together with a dependence regularizer that encourages the two responses to capture orthogonal information. To evaluate the proposed framework, we construct a large-scale dataset of free-form passive layouts, synthesized from both expert-designed contours and random paths. Our 21M-parameter model achieves high accuracy across a broad frequency range, circuit contexts and passive matching settings, with $R^2>0.98$ for almost all targets, while providing orders-of-magnitude faster evaluation than traditional EM simulation. Code available at https://github.com/ETH-IDEAS/RFIC-NeuralSim.

Figures & tables

Figure 1: Solver stability under imperfect EM response. The ground-truth (GT) solver achieves

R^{2}=1

with noiseless input, but its design-performance

R^{2}

becomes negative even at

R^{2}_{\mathrm{EM}}\approx 0.9984

. In contrast, the learned neural solver remains stable given imperfect EM responses. Appendix E details the evaluation protocol and illustrates the magnitude of the perturbations.

Figure 2: Typical RFIC PA with three passive matching networks: input matching, interstage matching, and output matching.

Figure 3: Example layouts from the two passive layout distributions.

Figure 4: Overview of the proposed model. The frequency-conditioned neural simulator maps the layout

{\bm{M}}

and frequency to a neural EM response, consisting of an explicit EM response and a complementary response. The circuit-conditioned neural solver then combines this neural EM response with the circuit context to predict design performance. Frequency information is injected through modulation across four residual blocks.

Figure 5: Predictions of the 21M input matching model on a frequency grid with 0.001 GHz spacing. Wider curve means the prediction changes less smoothly.

Figure 6: Per-target

R^{2}

over frequency for the 21M and 5.4M input matching models.

Figure 7: Per-target

R^{2}

over frequency for the 5.4M input matching model across training epochs.

Figure 8: In-distribution data scaling for input matching. Both models are trained and tested on RandPath-seeded layouts with a fixed training budget of 52,450 optimizer steps. Each panel reports test

R^{2}

for one target. Panel-specific

y

limits differ.

Figure 9: Effect of the complementary response for the 21M input matching model.

Figure 10: Generalization to unseen mask distributions between the Contour-seeded and RandPath-seeded families. Each cell reports $R^{2}_{\mathrm{design}}$ . Models are trained for 100 epochs for this analysis.

Figure 11: Effect of dependence regularization.

\mathrm{Corr}(\hat{S},Z_{\mathrm{comp}})

is the average correlation.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 12: Per-target

R^{2}

over frequency for the 21M and 5.4M interstage matching models on the full distribution.

Figure 13: Per-target

R^{2}

over frequency for the 21M and 5.4M output matching models on the full distribution.

Figure 14: Per-target

R^{2}

over frequency for the 5.4M interstage matching model across training epochs on the full distribution.

Figure 15: Per-target

R^{2}

over frequency for the 5.4M output matching model across training epochs on the full distribution.

Figure 16: Predictions of the 21M interstage matching model on a frequency grid with 0.001 GHz spacing, which is

100\times

denser than the training grid.

Figure 17: Predictions of the 21M output matching model on a frequency grid with 0.001 GHz spacing, which is

100\times

denser than the training grid.

Figure 18: Per-target

R^{2}

over frequency for the 21M input matching model under different dependence-regularization strengths on the full distribution.

Backbone	Parameters	$R^{2}_{\mathrm{EM}}$	$R^{2}_{\mathrm{design}}$
Plain CNN	5,347,884	0.98839	0.94128
Residual	5,383,468	0.98792	0.94386

Appendix

Table 3: Simulator-backbone comparison for input matching after 100 training epochs. Both models retain the conditional simulator–solver framework.

Matching	5.4M	21M
Input	$7.118\pm 0.036$	$3.925\pm 0.027$
Interstage	$8.170\pm 0.039$	$4.134\pm 0.028$
Output	$13.874\pm 0.049$	$6.890\pm 0.036$

Appendix

Table 4: Passivity-violation rates in percent. Error bars are one IID Bernoulli standard error over layout–frequency pairs, computed as

100\sqrt{\hat{p}(1-\hat{p})/N}

for the observed violation fraction

\hat{p}

. This calculation treats pairs as independent and does not account for dependence among frequencies of the same layout.

Example	Before	After
1	$-0.4457+0.2717j$	$-0.4396+0.2774j$
2	$-0.1701+0.3195j$	$-0.1654+0.3267j$
3	$-0.3637+0.3805j$	$-0.3585+0.3726j$

Appendix

Table 5: Example S-parameter values before and after adding 4% noise. Here

j

denotes the imaginary unit.

Neural Surrogate Modeling

Neural surrogate modeling uses learned models to approximate costly simulations, enabling faster prediction, design, control, and inference.

See all 412 papers

Alpha-RF: Automated RF-Filter-Circuit Design with Neural Simulator and Reinforcement Learning
Feb 18, 2026Nhat Tran, Chenjie Hao, Alexander Stameroff +2Electronic Design Automation Neural Surrogate Modeling
University of California, Davis · Keysight Technologies, Inc.
K-TRAIL: Simulator-Guided Generative Design of EM/RF Circuits
Sep 19, 2026Piyush Saha, Evan Newell, Hanna O'Leary +2Inverse Design Electronic Design Automation
School of Electrical Engineering and Computer Science, Oregon State University, Corvallis, OR 97331 USA · Department of Electrical and Computer Engineering, Yale University, New Haven, CT 06511 USA
Inverse Design of Multi-Layer Sub-Pixel-Resolution RF Passives Through Grayscale Diffusion with Flexible S-Parameter Conditioning
May 6, 2026Tommaso Dreossi, Christopher M. Bryant, Hao Liu +4Conditional Diffusion Models Inverse Design
Arena Physica New York, NY, USA

Feb 18, 2026J/K move · Enter open · S save

Alpha-RF: Automated RF-Filter-Circuit Design with Neural Simulator and Reinforcement Learning

Nhat Tran, Chenjie Hao, Alexander Stameroff, Anh-Vu Pham+1

University of California, Davis · Keysight Technologies, Inc.

Electronic Design Automation Neural Surrogate Modeling

Open paper AskPDF arXiv

Accurate, high-performance radio-frequency (RF) filter circuits are ubiquitous in radio-frequency communication and sensing systems for accepting and rejecting signals at desired frequencies. Conventional RF filter design process involves manual calculations of design parameters, followed by intuition-guided iterations to achieve the desired response for a set of filter specifications. This process is time-consuming due to time- and resource-intensive electromagnetic simulations using full-wave numerical PDE solvers. This process is also highly sensitive to domain expertise and requires many years of professional training. To address these bottlenecks, we propose an automatic RF filter circuit design tool using neural simulator and reinforcement learning. First, we train a neural simulator to replace the PDE electromagnetic simulator. The neural-network-based simulator reduces each of the simulation time from 4 minutes on average to less than 100 millisecond while maintaining a high precision. Such dramatic acceleration enable us to leverage deep reinforcement learning algorithm and train an amortized inference policy to perform automatic design in the imagined space from the neural simulator. The resulted automatic circuit-design agent achieves super-human design results. The automatic circuit-design agent also reduces the on-average design cycle from days to under a few seconds. Even more surprisingly, we demonstrate that the neural simulator can generalize to design spaces far from the training dataset and in a sense it has learned the underlying physics--Maxwell equations. We also demonstrate that the reinforcement learning has discovered many expert-like design intuitions. This work marks a step in using neural simulators and reinforcement learning in RF circuit design and the proposed method is generally applicable to many other design problems and domains in close affinity

Abstract

Neural Surrogate Modeling

Explore similar work

Alpha-RF: Automated RF-Filter-Circuit Design with Neural Simulator and Reinforcement Learning

K-TRAIL: Simulator-Guided Generative Design of EM/RF Circuits

Inverse Design of Multi-Layer Sub-Pixel-Resolution RF Passives Through Grayscale Diffusion with Flexible S-Parameter Conditioning