AI-Native Powerby Rafael Collado

BROWSER TOOL / LOCAL PROCESSING

Train and inspect a Buck converter surrogate

Start with the reproduced synthetic case or use one compatible CSV. Training runs on this device with fixed model settings; the interface keeps validation evidence and design authority separate.

  • One CSV at a time
  • 20 MB and 100,000-row web safety cap
  • Three-minute execution limit with cancellation

01 / DATA

Choose one bounded run

The sample creates 20,000 rows from the published simplified equations. Your CSV must use the seven documented numeric columns and may contain repeated case_id values for grouped evaluation.

Which columns must the CSV contain?

Use one row per operating point. The model reads five input columns and learns two target columns. Keep the header names and SI units exactly as shown.

ColumnRoleMeaningUnitValidation
vin_vInputInput voltageVGreater than zero
vout_vInputOutput voltageVGreater than zero and lower than vin_v
iout_aInputOutput currentAGreater than zero
fsw_hzInputSwitching frequencyHzGreater than zero
inductance_hInputInductanceHGreater than zero
loss_wTargetTotal converter lossWGreater than zero
ripple_current_aTargetPeak-to-peak inductor-current rippleAGreater than zero; values above iout_a trigger a CCM warning

Optional traceability columns: case_id activates a complete group holdout when repeated cases and at least five groups are available; source records provenance but is not used for training. Other columns are ignored with a warning.

The file needs 200–100,000 rows. Required values must be numeric, finite and non-empty.

Try the reproduced sample

Seed 42 · 20,000 synthetic rows · fixed model settings.

Download sample CSV

Upload your own CSV

No file selected

Choose the sample or a compatible CSV to begin.

WHAT THE RESULT MEANS

A screening result, not design approval

The included sample demonstrates approximation of simplified CCM equations. An uploaded CSV evaluates only the rows and split represented in that file. Neither result proves performance on another simulator, component family or hardware.

The tool blocks malformed inputs, warns about weak CCM interpretation, holds out complete repeated case_id groups when available and flags physically non-positive predictions. Its training-bounds check is not complete out-of-distribution detection.

Browser capacity and dataset sufficiency are separate. If the largest learning-curve point still improves or the web limit is reached, continue with the unrestricted local repository.