WHAT THIS CASE PROVES
The recorded run proves that the published code can approximate its bundled simplified equations on a random held-out sample from the same synthetic distribution. It does not establish a simulator speed-up, accuracy on another data source or a hardware-ready design.
The engineering question
Given a synchronous Buck operating point and inductance, can a supervised model estimate total converter loss and peak-to-peak inductor-current ripple inside a documented design domain? The purpose is to demonstrate the complete surrogate workflow, not to replace equations that are already inexpensive and available.
The synthetic design domain
| Input | CSV column | Sampled range | Unit |
|---|---|---|---|
| Input voltage | vin_v | 12–60 | V |
| Output voltage | vout_v | 3.3 to min(24, 0.85 Vin) | V |
| Output current | iout_a | 0.5–15 | A |
| Switching frequency | fsw_hz | 50,000–500,000 | Hz |
| Inductance | inductance_h | 10–150 | µH |
Sampling is uniform for voltage and current and log-uniform for frequency and inductance. Points with ripple greater than output current are rejected to keep this teaching example conservatively inside CCM.
The targets are loss_w and ripple_current_a. The fixed, generic loss parameters are not tied to a commercial part number.
The reproducible pipeline
- Define the input domain and the quantities to estimate.
- Generate synthetic reference rows with a fixed random seed.
- Validate column names, numeric values, signs and the Buck voltage relationship.
- Reserve 20% of the rows and train separate histogram gradient-boosting regressors.
- Report every output separately and retain held-out predictions.
- Measure a learning curve against one unchanged test set.
- Screen candidates inside the recorded training bounds, then return them to the source method.
Held-out results
The recorded run uses seed 42, 16,000 training rows and 4,000 held-out rows. Values below come from the versioned report rather than from a marketing estimate.
| Target | R² | MAE | MAPE | 95th-percentile APE |
|---|---|---|---|---|
| Total loss | 0.9996 | 0.0348 W | 1.52% | 4.77% |
| Ripple current | 0.9943 | 0.0644 A | 5.58% | 15.55% |
Ripple is visibly the weaker target. Reporting both targets separately prevents the stronger loss estimate from hiding that difference.
The learning curve answers the row-count question
| Target | MAPE at 10,000 training rows | MAPE at 16,000 training rows |
|---|---|---|
| Total loss | 1.60% | 1.53% |
| Ripple current | 5.96% | 5.64% |
Additional rows still help, but with diminishing returns. This supports neither a universal 10,000-row minimum nor a universal 20,000-row guarantee. A new simulator or laboratory dataset needs its own curve, coverage analysis and grouped tests.
Inference timing without an invented speed-up
A native Python run predicted 10,000 rows in a median 28.7 ms across 15 repetitions, approximately 2.87 µs per row. The measurement excludes CSV input/output and was not compared with a simulator. Browser execution has a different runtime and must be measured separately before making a browser-performance claim.
Replace the generator with SPICE, PLECS or experimental data
Export one valid steady-state operating point per row using the same seven numeric columns. Preserve case_id and source as traceability fields. Failed simulations belong in a separate failure record, not as zero-valued labels.
- Convert values to the documented SI units before export.
- Retain the simulation deck, extraction code, solver status and measurement window.
- Group repeated operating points by the physical design or campaign that must remain unseen.
- Compare shortlisted designs again in the original simulator or laboratory.
Limitations and the next evidence
- The generator omits thermal dynamics, magnetic saturation, layout parasitics and control-loop dynamics.
- Independent regressors do not enforce every physical relationship between outputs.
- Training min/max bounds are a useful warning, not complete out-of-distribution detection.
- No simulator or hardware comparison is included in the reproduced reference run.
The next valuable release is a versioned simulator dataset with grouped design identities, extraction evidence and at least one region where the surrogate changes the simulation priority.