Designing RF/mmWave passive components is a time-consuming process that relies heavily on expert knowledge and iterative EM simulations. Although recent ML surrogate models have shown great promise in accelerating passive design, they remain task-specific and lack generalization. This paper presents FoundaRF, a foundation model framework that unifies passive layout modeling across diverse layout structures and technologies. It combines a self-supervised encoder with adaptive layout processing and physics-aware decoders to enable few-shot transfer. Experiments across four semiconductor technologies show strong generalization to unseen structures, outperforming baselines with up to a 3.3x reduction in prediction error, reducing labeled-data requirements by up to fourfold, and substantially shortening design turnaround. FoundaRF establishes a scalable and data-efficient solution for RF/mmWave passive design automation.
Figures & tables
Figure 1. Different ML modeling paradigms: (a) task-specific fully supervised vs. (b) proposed FoundaRF framework.
PulseRF ( 31 )
MOTIF-RF ( 12 )
T-Coil
Coil
Tech Node
InP
SOI
SiGe
GaN
Topology
Transformer
Transformer
Coupled Inductors
Spiral Coil
Key Features
Non-intuitive
Multi-template
Center-tapped
Standard
Frequency Range
0–300 GHz
0–200 GHz
0–100 GHz
0–40 GHz
Number of Ports
4
4
3
2
Simulation Tool
Ansys HFSS
Cadence EMX
Cadence EMX
ADS Momentum
Table 1. Summary of Datasets Used in This Work
Figure 2. (a) two-port inductors in Qorvo 150-nm GaN process and (b) three-port T-coils in GF 130-nm SiGe process.
Figure 3. FoundaRF overview. The framework consists of four stages: adaptive layout processing (Sec. 3.1 ), a universal encoder (Sec. 3.2 ), a FiLM-conditioned decoder (Sec. 3.3 ), and a task-specific adapter that outputs the S-parameter matrix (Sec. 3.4 ).
Figure 4. DF-Based Pretraining Overview.
Technology
SOI
InP
SiGe
GaN
Structure
1:1
M:N
8-shaped
Parallel
PulseRF
T-Coil
Coil
Metric
MAE
R2
MAE
R2
MAE
R2
MAE
R2
MAE
R2
MAE
R2
MAE
R2
Tabular ( 12 )
0.0214
0.978
0.0638
0.886
0.0108
0.992
0.0099
0.995
0.0302
0.919
0.0138
0.996
0.0041
0.997
Image ( 31 )
0.0193
0.982
0.0513
0.911
0.0125
0.990
0.0146
0.990
0.0250
0.921
0.0217
0.990
0.0074
0.997
Graph Transformer ( 12 )
0.0164
0.987
0.0614
0.890
0.0312
0.925
0.0583
0.871
n/a ‡
0.0139
0.995
n/a ‡
Concat(VarImg, Scaling)
0.0089
0.996
0.0465
0.927
0.0072
0.996
0.0072
0.995
0.0183
0.950
0.0127
0.996
0.0028
0.999
Table 2. Model evaluation for adaptive layout processing and FiLM-conditioned decoder across datasets.
Technology
SOI
InP
Label Frac.
1%
5%
1%
5%
Metric
MAE
R2
MAE
R2
MAE
R2
MAE
R2
Fully Supervised
0.0833
0.787
0.0347
0.953
0.0906
0.546
0.0547
0.808
Frozen
0.0519
0.917
0.0309
0.965
0.0898
0.496
0.0562
0.791
Fine-tuned
0.0460
0.936
0.0263
0.975
0.0811
0.587
0.0500
0.819
Table 4. Few-Shot Evaluation: (SOI+InP) Training Scheme
Figure 5. Model performance vs. sample size for transfer learning from SOI+InP to other technologies (SiGe/GaN).
Figure 6. Schematic models for all three component families.
Metric [-1pt] ( ∗ : average)
Direct pred.
S-param pred. (ours)
Tech
MedPE
R2
MedPE
R2
Improvement
L∗
SOI
6.39 %
0.187
0.45 %
0.999
14.2×
SiGe
5.33 %
0.856
0.56 %
0.996
9.5×
GaN
2.09 %
0.978
0.35 %
0.999
6.0×
Rs∗
SOI
23.53 %
0.256
3.03 %
0.995
7.8×
SiGe
28.20 %
0.626
7.83 %
0.912
3.6×
Table 5. S-param prediction (ours) vs. direct equivalent circuit parameter prediction. MedPE = median percentage error.
Figure 7. T-coils inverse design for (a) driver and (b) RX input.
Figure 8. T-coil inverse design results for (a) driver and (b) RX input.