WorksComputer Science
Route-Aware EM Surrogate for Automated LC Matching Network Optimization
Abstract
This paper proposes an open-source automated LC matching network optimization framework that combines schematic-level design optimization with electromagnetic (EM)-aware post-layout refinement. An ML-based surrogate predicts the frequency-dependent complex admittance of individual routes, enabling reuse across matching network topologies and physical layouts for the same technology node without retraining. These route models guide post-layout passive retuning to recover matching performance while preserving the layout footprint. Evaluated in a commercial 65nm CMOS technology, the surrogate achieves 1.94-4.91% mean absolute percentage error with respect to a full-wave electromagnetic solver, with no detected passivity violations. Across nine LC matching networks operating at 5-95GHz, our framework recovers S_11 from -6.37dB, for schematic-level optimization after post-layout degradation to -22.27dB after EM-aware post-layout refinement. All designs meet matching targets under full-layout EM verification in optimization runtimes below 70s.
Introduction
Modern wireless communication systems increasingly operate at radio-frequency (RF) and millimeter-wave (mm-wave) frequencies to support high data rates and wide bandwidths. At these frequencies, impedance matching is essential to reduce reflections and maximize power transfer between circuit blocks with different source () and load () impedances [1, 2]. Inductance-capacitance (LC) matching networks are widely used to realize this at a target operating frequency. Their design commonly begins with schematic-level optimization, which selects the topology and passive values to meet matching targets under ideal, lossless interconnect assumptions.
Prior work has explored schematic-level topology search and passive-value optimization [3, 4]. To improve physical design accuracy, space mapping and surrogate-assisted EM optimization [5, 6, 7], automated passive synthesis and layout generation [8], and Gaussian-process/Bayesian optimization of matching network layouts [9] have been investigated. Learned EM surrogates have also been developed for passive structures and matching networks [10, 11, 12, 13, 14, 15, 16]. While these methods improve the design of passive and matching network modules, they do not explicitly account for the impact of the external routing interconnects connecting these modules to the rest of the system. Perfect matching at the schematic stage may translate to significant losses in a placed and routed design.
Figure 1 illustrates this limitation using a schematic-level optimized matching network. After layout generation using COmPOSER [8], unmodeled routing parasitics degrade input/output reflection () and transmission (), violating matching targets. Recovering these targets requires post-layout passive retuning, which can alter passive dimensions, terminal locations, and routing responses. Evaluating these changes requires costly EM simulations [17, 18], potentially repeated across retuning iterations. Efficient retuning therefore requires a generalizable EM model that captures varying route geometries and terminal connectivities (e.g., point-to-point (P2P) routes (NET1) and branched multiport routes (NET2)), as shown in Figure 1 without repeated full EM simulations.
Figure 1. Performance degradation due to layout-oblivious matching network design and unmodeled routing EM effects.
To address these requirements, this work proposes an automated matching network optimization framework using a route-level EM surrogate model. We train a multilayer perceptron (MLP) surrogate to predict the frequency-dependent complex admittance of P2P routes from their geometries. Passivity-aware training enforces passivity of the predicted admittances at the sampled frequency [19, 20]. Unlike complete-network models, the route-level surrogate can be reused across matching topologies and layouts within its trained geometry, metal-stack, and technology domain. Starting from the matching requirements, we use differential evolution (DE) [21] to optimize passive values across candidate LC topologies to build the matching network at the schematic level, followed by physical layout generation using the open-source netlist-to-GDSII tool COmPOSER [8]. We decompose branched multiport routes into P2P segments and reconstruct their admittances at the original terminals using the Kron reduction [22]. These reconstructed route models guide post-layout passive retuning that updates the values in the matching network based on routing parasitics, preserving the selected topology and layout footprint to ensure convergence. We then generate the updated layout and verify its performance using full EM simulation. We evaluate the proposed framework across nine LC matching network designs with varying operating frequencies and bandwidths up to 100 GHz in commercial 65 nm CMOS technology. Post-layout retuning achieves up to 83.2% power transfer and as low as −27.5 dB within 70 s, with all nine designs meeting matching targets under full EM verification.
Proposed Framework
Schematic-level sizing of matching networks
The proposed flow (Figure 2) begins with schematic-level sizing of candidate LC matching networks. Inputs include the matching requirements (as illustrated by the targets in Figure 1): , , , LC topology library, and technology. Fig. 3(a) illustrates the source-to-load matching problem. Maximum power transfer ideally requires conjugate matching, and , through optimized LC values. Fig. 3(b) shows the desired S-parameter response over the target band, with and below the dotted specification line to minimize reflected power. Meanwhile, should be maximized to improve source-to-load power delivery.

Figure 2. Proposed LC matching network design flow.
For each topology, we employ a DE [21]-based approach to optimize LC values using population-based candidate generation and cost-based selection. Each candidate is evaluated by stamping ideal passive admittances, and , into a nodal matrix. Two nodal solves, one for an excitation at each port, provide the power-wave S-parameters referenced to and . The matching cost, , is evaluated over frequency samples as
These terms penalize input reflection, undelivered available power, and output reflection, respectively. For single-frequency matching, at ; otherwise, samples span the target band (Fig. 3(b)). Prescribed S-parameter limits must be obeyed at all sampled frequencies. The optimized topologies are ranked by cost, with fewer components preferred for ties. The selected design then proceeds to initial layout generation.
Initial layout generation
As shown in Figure 2, the selected topology and optimized LC values are passed to COmPOSER [8, 23]. COmPOSER maps passive values to PDK-realizable layouts, optimized for high quality factor and low loss, and performs placement and routing within the available layout region while maintaining fixed source/load interfaces. The resulting GDSII provides the interconnect geometries, passive footprints, and terminal locations for subsequent extraction and route-aware retuning.
ML surrogate-based EM estimation and retuning
Accurate estimation of interconnect parasitics is essential for layout-aware matching network optimization, but repeated EM simulations during retuning are computationally expensive. We therefore employ an MLP-based surrogate to predict the frequency-dependent admittance of routed interconnects from their geometry. Depending on their connectivity, interconnects may connect two or more terminals (Figure 1). P2P routes are modeled directly as 2-port networks, while multiport routes are decomposed into P2P segments whose predicted admittances are assembled and reconstructed using Kron reduction. This allows the same surrogate to model different terminal counts and routing configurations.
Point-to-point routes. The surrogate is trained on 50,000 EMX-characterized, PDK-valid multilayer Manhattan P2P routes, generated in about 6 hours. In a commercial 65 nm CMOS, the upper metal layers (M8–M9) are used due to their low resistance and suitability for RF/mm-wave design. This choice is technology-specific; other technologies require appropriate low-resistivity layer selection and retraining. The training routes contain 0–3 bends and segment lengths of up to 100 m: such lengths are observed even in compact matching network layouts within LNA/PA designs, even though COmPOSER optimizes routes to favor short interconnects. Each route uses 594 features describing terminal locations, metal polygons, segment dimensions, bends, and vias. Hyperparameter tuning selects a 3-hidden-layer MLP (128/56/128 neurons) to predict the complex 2-port admittance matrix
at 200 frequency samples up to 100 GHz, yielding 1600 outputs from the real/imaginary parts of four admittance entries.
Non-passive surrogate predictions can introduce spurious power generation and physically invalid retuning solutions. We ensure passivity at each sampled frequency through passivity-aware training, followed by correction of remaining violations before optimization. For the Hermitian part , passivity requires , i.e., all eigenvalues of must be nonnegative [19, 20]. The training loss includes a penalty:
where , scales the penalty, and after a five-epoch warmup. This regularization reduces passivity violations in the ML model from 51.07% to 0.26% and MAE from 15.39 mS to 4.23 mS. Remaining violations are corrected by interpolating between adjacent passive frequency samples. Thus, all admittances are guaranteed to satisfy passivity.
Multiport routes. For multiport routes such as NET2 in Figure 1, the P2P surrogate is extended through decomposition and reconstruction. Figure 4 illustrates decomposition of a branched route into P2P segments at internal junctions and reconstruction using Kron reduction. In the three-terminal example, the route connecting , , and through junction is decomposed into –, –, and –. Each segment is evaluated using the same MLP surrogate, and its predicted 2-port admittance is stamped into a common nodal admittance matrix according to the connectivity [24]. With external nodes and internal node , the assembled admittance matrix is partitioned into external (), internal (), and cross-coupling (, ) blocks. Kron reduction [22] eliminates the internal junction, yielding the reconstructed admittance:
Figure 4. Point-to-point decomposition and Kron reconstruction of a branched multiport interconnect.
at the original terminals. This avoids training separate surrogate models for different port count/branched interconnects.
Interconnect-aware retuning. With the interconnect admittance models, the next stage incorporates these models into the matching network to compensate for the routing-induced performance degradation illustrated in Figure 1. The objective is to recover the optimized schematic-level response by adjusting the values while preserving the selected topology. The target admittance is obtained by assembling the optimized schematic-level passive admittances into a nodal matrix and eliminating internal nodes to obtain the two-port response at the source/load terminals. During retuning, remains fixed, while the routed network response is evaluated using current passive values and reconstructed interconnect models. The deviation from the target is expressed as
stacking the real () and imaginary () errors at frequency . To avoid repeating global optimization, a first-order Taylor expansion approximates the error variation with passive values. The Jacobian captures the sensitivity of the error to each value, giving . A bounded least-squares step determines the adjustments that bring the routed response closer to . The empirically chosen bounds limit passive-value changes to of the schematic solution to help preserve the initial layout footprint and routing, avoiding repeated interconnect model updates due to changes in placement and routing.
Algorithm 1 summarizes this iterative procedure. Retuning is initialized with the route models, target response, and schematic values (Lines 1–2). Each iteration evaluates the routed response (Line 4) and its error (Line 5). Finite-difference perturbations of each estimate the Jacobian columns (Lines 7–8). Bounded least-squares updates of passive values continue until convergence (Lines 10–11), yielding the retuned passive values (Line 13).
Algorithm 1 Route-Aware Passive Retuning
Input: route models , target , initial values
while not converged do
Stamp current values and to obtain
Compute using (4)
for each passive value do
Perturb by and compute the resulting error
end for
s.t.
end while
return retuned values
The returned values are mapped to realizable devices while preserving the existing layout footprint and source/load interfaces. Unchanged routes reuse surrogate predictions, while modified routes are regenerated and reevaluated. COMPOSEER generates the updated layout, followed by full EMX verification of the complete matching network (Figure 2).
Experimental Setup and Results
EM surrogate accuracy
P2P models. We first evaluate the 2-port EM surrogate using 5,000 unseen routes (10% of the dataset). The surrogate is trained for 50 epochs using AdamW with an adaptive learning rate initialized at . Table 1 reports mean absolute and percentage error (MAE and MAPE) for the real and imaginary parts of all four admittance parameters, averaged over test routes and frequency samples. The average MAE is 4.23 mS across all eight outputs, with comparable accuracy for the transfer (, ) and self-admittance (, ).
| Metric | Real | Imag. | Real | Imag. | Real | Imag. | Real | Imag. |
| MAE (mS) | 3.69 | 4.74 | 3.76 | 4.72 | 3.71 | 4.70 | 3.70 | 4.79 |
| MAPE (%) | 3.09 | 2.07 | 2.99 | 1.94 | 2.99 | 1.94 | 3.14 | 2.09 |
Table 1. 2-port EM surrogate accuracy over 5,000 test routes.
Multiport route model. Next, we evaluate multiport reconstruction using the 2-port surrogate and Kron reduction (Section II-C). We evaluate 50 unseen layouts each for three- to five-ports, covering the maximum port count in our testcases. Full EMX simulation of each intact multiport net serves as the reference. Each net is reconstructed twice using the same Kron-reduction procedure: first from individually simulated EMX P2P segments and then from surrogate-predicted segments. The EMX-segment path isolates decomposition errors, primarily from neglected mutual EM coupling, while the surrogate path evaluates the complete reconstruction.
Table II reports complex-admittance MAE and MAPE against EMX, averaged over layouts, admittance entries, and frequencies. Both reconstruction approaches exhibit comparable errors, with MAPE ranging from 3.45% to 4.78% for EMX segments and 4.13% to 4.91% for surrogate segments. The lower surrogate MAE for the 4-port case results from partial cancellation of prediction and decomposition errors.
| Reconstruction | 3-port MAE (mS) | 3-port MAPE (%) | 4-port MAE (mS) | 4-port MAPE (%) | 5-port MAE (mS) | 5-port MAPE (%) |
| EMX segments + Kron | 12.43 | 3.45 | 19.73 | 4.06 | 21.15 | 4.78 |
| Surrogate segments + Kron | 13.01 | 4.83 | 18.92 | 4.13 | 22.03 | 4.91 |
Table II. Multiport reconstruction accuracy compared to EMX.
Comparing the two reconstructed admittances helps isolate the additional error from surrogate substitution. Figure 5 further illustrates that the surrogate-based reconstruction closely tracks the EMX responses up to 100 GHz across the admittance matrix. These results support surrogate-based multiport reconstruction for route-aware retuning.

Figure 5. Full EMX vs. surrogate-based Kron-reconstructed admittance response for the 3-port route in Figure 4.
End-to-end EM-aware matching network synthesis
We evaluate the complete flow on nine matching problems: three single-frequency and six banded cases with bandwidths of 2–13 GHz (Table III). The framework and testcases are open-sourced at [25] and cover operating frequencies of 5–95 GHz, varying complex source/load impedances (, ), and networks with 2–7 passive elements. Single-frequency cases use compact L- and -type networks at target frequencies, while banded cases employ -, T-, and higher-order LC networks with additional tunable passive values to maintain matching across frequency bands.
| Topology | (GHz) | (Ohm) | (Ohm) | (GHz) | (dB) | (dB) | (dB) | (%) | Time (s) | |
| HP-L | 2 | 15 | 14.0/15.0 | -7.1/-21.7 | -3.0/-0.9 | -2.1/-20.6 | 50.1/81.2 | 32.0 | ||
| LP-L | 2 | 40 | 35.7/40.0 | -2.5/-22.2 | -4.7/-0.8 | -3.6/-20.2 | 33.9/83.2 | 33.5 | ||
| LP- | 3 | 65 | 54.9/65.0 | -6.7/-19.7 | -3.1/-1.1 | -8.8/-18.9 | 48.9/77.6 | 36.9 | ||
| HP- | 3 | 5–7 | 6.0/6.0 | -0.6/-19.9 | -5.1/-1.0 | -7.1/-23.8 | 30.9/79.4 | 41.1 | ||
| BP-T | 6 | 38–42 | 36.5/40.0 | -9.1/-24.5 | -5.8/-1.0 | -7.4/-24.1 | 26.3/75.8 | 51.8 | ||
| BP- | 6 | 89–95 | 79.7/91.0 | -7.1/-27.5 | -4.8/-0.9 | -7.4/-24.6 | 33.1/81.3 | 61.4 | ||
| LP-5 | 5 | 35–42 | 34.0/38.5 | -4.0/-21.4 | -7.8/-1.1 | -2.8/-20.9 | 16.5/77.6 | 55.7 | ||
| HP-5 | 5 | 67–80 | 68.0/73.5 | -9.1/-22.0 | -5.5/-0.9 | -5.5/-20.7 | 28.1/81.2 | 64.1 | ||
| LP-7 | 7 | 80–92 | 74.0/86.0 | -5.7/-21.5 | -4.6/-1.3 | -4.2/-19.8 | 34.6/74.1 | 68.9 |
Table III. Evaluated testcases and their performance metrics before and after route-aware retuning.
For each test case, the matching-performance targets are dB and dB. The reflection limits correspond to at most 3.2% reflected power at either port. The percent forward power transfer [1], [26], with the target requiring at least 70% transfer. The achieved center frequency, , is also tracked against the single target frequency, , or the band midpoint. Single-frequency designs are evaluated at ; for banded designs, worst-case and , and minimum and are reported over the target band, with final design compliance confirmed by EMX.
Table III focuses on the performance recovered through route-aware retuning relative to the degraded initial physical layout, since schematic-level DE optimization ideally targets zero reflection and 100% power transfer (Section II-A). Each A/B entry for , , , , and denotes the surrogate-estimated initial-layout / full-EMX-verified post-retuning final-layout performance. Thus, A captures the degradation after physical realization, while B shows the recovery achieved through route-aware retuning and layout regeneration. In the table, and Time denote the passive-element count and optimization runtime, respectively.
Notation: HP/LP/BP: high-/low-/band-pass; L, T, : network configurations; : passive-element count. A/B: surrogate-estimated initial-layout / full-EMX-verified post-retuning final-layout performance.
The initial layouts based on schematic optimization exhibit substantial matching degradation due to routing parasitics. Across nine cases, falls to 16.5–50.1% (: −7.8 to −3.0 dB), while and degrade to −9.1 to −2.5 dB and −9.1 to −2.1 dB, respectively. No initial design satisfies the prescribed reflection or transmission targets; routing parasitics also shift from the target. The degradation is consistent with concurrent losses and distributed EM effects of interconnect routes, ignored in the ideal schematic representation.
Route-aware retuning recovers the required performance in all nine cases, pulling the shifted center frequency back to its target value. Full-EMX verification of the regenerated layouts yields from −27.5 to −19.7 dB, from −24.6 to −18.9 dB, and from −1.3 to −0.8 dB, corresponding to of 74.1–83.2%. Even the worst final designs retain margins of 4.7 dB for , 3.9 dB for , and 0.2 dB for , while even the minimum exceeds the 70.8% requirement by 3.3 percentage points. For banded cases, these values represent worst-case performance across the entire frequency interval.
The optimization runtime of 32.0–68.9 s includes schematic-level DE optimization, ComPOSER layout generation, and route-aware post-layout retuning. On average, schematic-level DE optimization accounts for 3.5% of the execution time, ComPOSER layout generation consumes 64.4%, and route-aware post-layout retuning takes the remaining 32.1%. Offline surrogate training and final EMX verification are excluded from this runtime (Section II-C). Overall, surrogate-guided retuning recovers the required matching performance across the evaluated topologies and operating frequencies, with final compliance independently verified using full-layout EMX.
Conclusion
We present a route-aware LC matching network optimization framework using EM surrogates to model frequency-dependent interconnect parasitics. Post-layout retuning recovers matching performance lost during schematic-to-layout translation. The surrogate achieves 2.53% average MAPE. Across nine matching network topologies spanning frequencies and bandwidths, retuning improves average power transfer by 45.4 percentage points and by 15.9 dB, achieving 83.2% power transfer and as low as −27.5 dB within 70 s. All meet targets under full-layout EM verification.
References
References
- [1]D. M. Pozar, Microwave Engineering, 4th ed. New Jersey, USA: John Wiley & Sons, 2011.
- [2]B. Razavi, RF Microelectronics, 2nd ed. New Jersey, USA: Prentice Hall Press, 2011.
- [3]A. Chen, T. Jiang, Z. Chen, and Y. Zhang, “A genetic and simulated annealing combined algorithm for optimization of wideband antenna matching networks,” International Journal of Antennas and Propagation, vol. 2012, no. 1, Apr 2012, 6 pages.
- [4]B. S. Yarman, M. Sengul, and A. Kilinc, “Design of practical matching networks with lumped elements via modeling,” IEEE Transactions on Circuits and Systems I: Regular Papers, vol. 54, no. 8, pp. 1829–1837, 2007.DOI
- [5]J. W. Bandler, R. M. Biernacki, S. H. Chen, P. A. Grobelny, and R. H. Hemmers, “Space mapping technique for electromagnetic optimization,” IEEE Transactions on Microwave Theory and Techniques, vol. 42, no. 12, pp. 2536–2544, Dec. 1994.DOI
- [6]B. Liu, D. Zhao, P. Reynaert, and G. G. E. Gielen, “GASPAD: A general circuit synthesis method based on surrogate model assisted evolutionary algorithm,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, vol. 33, no. 2, pp. 169–182, Jan. 2014.DOI
- [7]B. Liu, L. Xue, H. Fan, Y. Ding, M. Imran, and T. Wu, “An efficient and general automated power amplifier design method based on surrogate model assisted hybrid optimization technique,” IEEE Transactions on Microwave Theory and Techniques, vol. 73, no. 2, pp. 926–937, Feb. 2025.DOI
- [8]S. Ghosh, S. S. Peri, S. A. Berhan, E. Y. Gebru, R. Harjani, and S. S. Sapatnekar, “ComPOSER: Circuit optimization of mm-wave/RF circuits with performance-oriented synthesis for efficient realizations,” in Proceedings of the ACM/IEEE Design Automation Conference, Jul. 2026, 7 pages.arxiv.org/abs/2603.20486
- [9]G. Chen, Y. Wang, Y. Song, H. Liu, Q. Wu, G. Lu, W. Hong, and H. Wang, “Fully automated schematic and layout synthesis of broadband LC matching networks with flexible topologies for RFIC designs,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, Sep. 2026, 14 pages.DOI
- [10]F. S. N. Lourenço, E. Roca, R. Martins, R. Castro-López, N. Horta, and F. V. Fernández, “PACOSYT: A passive component synthesis tool based on machine learning and tailored modeling strategies towards optimal RF and mm-wave circuit designs,” IEEE Journal of Microwaves, vol. 3, no. 2, pp. 599–613, Apr. 2023.DOI
- [11]S. Er, E. Liu, M. Chen, Y. Li, Y. Liu, T. Zhao, and H. Wang, “Deep learning assisted end-to-end synthesis of mm-wave passive networks with 3D EM structures: A study on a transformer-based matching network,” in Proceedings of the IEEE MTT-S International Microwave Symposium, Jun 2021, pp. 66–69.arxiv.org/abs/2201.02141
- [12]D. Lee, G. Shin, S. Lee, K. Kim, T.-H. Oh, and H.-J. Song, “Neural-network-based automated synthesis of transformer matching circuits for RF amplifier design,” IEEE Transactions on Microwave Theory and Techniques, vol. 70, no. 11, pp. 4726–4739, Nov. 2022.DOI
- [13]H. He, Y. Xu, L. Xia, Y. Hu, F. Cai, and T. Chi, “MOTIF-RF: Multi-template on-chip transformer synthesis incorporating frequency-domain self-transfer learning for RFIC design automation,” in Proceedings of the Asia-South Pacific Design Automation Conference, Jan. 2026, pp. 1138–1144.arxiv.org/abs/2511.21970
- [14]E. A. Karahan, Z. Liu, and K. Sengupta, “Deep-learning-based inverse-designed millimeter-wave passives and power amplifiers,” IEEE Journal of Solid-State Circuits, vol. 58, no. 11, pp. 3074–3088, Nov. 2023.DOI
- [15]H. Chae, H. Yu, S. Li, and D. Z. Pan, “PulserRF: Physics-augmented ML modeling and synthesis for high-frequency RFIC design,” in Proceedings of the IEEE/ACM International Conference on Computer-Aided Design, Aug. 2024, 9 pages.DOI
- [16]X. Tian, Z. Li, J. Huang, and S. Gan, “An efficient RF circuit optimization method based on pixelated microstrip matching networks,” Microelectronics Journal, vol. 175, Jun. 2026.DOI
- [17]Q. Xia, D. Li, J. Huang, J. Li, H. Chang, B. Sun, and H. Liu, “A 28 GHz linear power amplifier based on CPW matching networks with series-connected DC-blocking capacitors,” Electronics, vol. 9, no. 4, Apr. 2020.DOI
- [18]J. P. Aikio, A. Sethi, M. Hietan en, J. Rusanen, T. Rahkonen, and A. Pärssinen, “Ka-band stacked power amplifier supporting 3GPP new radio FR2 band n258 implemented using 45 nm CMOS SOI,” Applied Sciences, vol. 11, no. 15, Jul. 2021.DOI
- [19]B. Gustavsen and A. Semlyen, “Enforcing passivity for admittance matrices approximated by rational functions,” IEEE Transactions on Power Systems, vol. 16, no. 1, pp. 97–104, Feb. 2001.
- [20]P. Triverio, S. Grivet-Talocia, M. S. Nakhla, F. G. Canavero, and R. Achar, “Stability, causality, and passivity in electrical interconnect models,” IEEE Transactions on Advanced Packaging, vol. 30, no. 4, pp. 795–808, Nov. 2007.DOI
- [21]R. Storn and K. Price, “Differential evolution – a simple and efficient heuristic for global optimization over continuous spaces,” Journal of Global Optimization, vol. 11, no. 4, pp. 341–359, 1997.DOI
- [22]F. Dörfler and F. Bullo, “Kron reduction of graphs with applications to electrical networks,” IEEE Transactions on Circuits and Systems I: Regular Papers, vol. 60, no. 1, pp. 150–163, Jan. 2013.arxiv.org/abs/1102.2950
- [23]S. Ghosh, S. S. Peri, S. A. Berhan, E. Y. Gebru, R. Harjani, and S. S. Sapatnekar, “Circuit Optimization of mm-wave/RF circuit Optimization of mm-wave/RF circuits with Performance-Oriented Synthesis for Efficient Realizations,” https://github.com/UMN-EDA/ComPOSER, 2026, accessed on Aug. 28, 2026.arxiv.org/abs/2603.20486
- [24]C.-W. Ho, A. E. Ruehli, and P. A. Brennan, “The modified nodal approach to network analysis,” IEEE Transactions on Circuits and Systems, vol. 22, no. 6, pp. 504–509, Jun. 1975.DOI
- [25]S. Ghosh, S. S. Peri, R. Harjani, and S. S. Sapatnekar, “Route-aware EM surrogate for automated LC matching network optimization,” https://github.com/UMN-EDA/ComPOSER/tree/main/MN_synthesis, 2026, accessed on Oct. 1, 2026.arxiv.org/abs/2611.00409
- [26]K. Kurokawa, “Power waves and the scattering matrix,” IEEE Transactions on Microwave Theory and Techniques, vol. 13, no. 2, pp. 194–202, Mar. 1965.DOI

