Research problem

Inverse CPW antenna design without rerunning CST for every idea.

This system turns a full CST sweep campaign into a local design terminal. A reviewer can trace the data path, inspect coverage, test forward predictions, synthesize candidate antenna dimensions, and verify the result against trained final-data models.

Final CST rows
6072
Complete lp/wp sweeps
Antennas
12
Four geometry families
S11 samples
151
1-7 GHz response curve
Serving mode
Trained
trained
Live system state
API okDataset 1.0.0-finalScalar 0.2.0-xgboostSignature 0.2.0-torchInverse 0.2.2-tandem-scaled
1
2
3
4
5
6
7
8
9
10
11
12

Each cell is one antenna template. A complete final build has 506 normalized runs per antenna.

Workflow

From raw CST curves to an interactive synthesis loop.

01

CST exports

Raw curves from 12 CPW antenna templates.

02

Normalization

S11, gain, efficiency, and sweep metadata become one master table.

03

Forward models

Geometry predicts scalar metrics and full S11 signatures.

04

Inverse tandem

Target behavior proposes antenna ID, Lp, and Wp.

05

Verification

Forward model checks the synthesized design against the requested target.

Model evidence
Scalar R20.866
S11 curve R20.875
Inverse accuracy0.894
Inverse success0.078

The inverse model is intentionally shown with both classification and strict target-match success so the research limitation is visible.

Sweep surface preview

Heatmaps expose how resonance and gain move across the lp/wp design grid.

Forward labChange antenna geometry, render it instantly, then run trained prediction.Best for explaining geometry-to-performance behavior.Inverse labEnter a target response and inspect the proposed antenna plus verification errors.Best for demonstrating the research objective.EvaluationReview model metrics, heatmaps, leaderboard entries, and dataset evidence.Best for thesis review and reproducibility checks.