Machine learning for power electronics: a practical surrogate-model workflow

Start with the engineering decision

“Use machine learning for this converter” is not a specification. First state what the model may estimate, which inputs exist before the decision, the units, the acceptable error and what happens after a warning or refusal.

• Decision: for example, which candidates deserve detailed simulation.
• Target: a measurable quantity such as loss or ripple current.
• Reference: equation, simulator or measurement used to create the label.
• Authority: the method that must recheck a shortlisted design.

Define the design space before sampling it

Record topology, component identity, operating variables, units and valid ranges. A row count has little meaning without coverage: 100,000 clustered points can contain less useful information than 10,000 points distributed across the regions that affect the decision.

Use a space-filling plan where appropriate, preserve boundary cases and keep failed solver executions separate from electrical rejection. A timeout is not a zero-loss converter and must never be encoded as one.

Build traceable data

Label every dataset as synthetic, simulated or measured. Preserve the source version, sampling plan, units, convergence state and a design or campaign identifier. Those fields determine what the eventual validation result can support.

Keep a physical baseline beside the model

Compare the surrogate with the simplest credible method on the same untouched cases. If an equation is already fast and sufficiently accurate, replacing it with a learned model adds maintenance without engineering value. The surrogate earns a role only when it improves a named cost while preserving validation and review.

Validate independent cases, not duplicated context

A random row split is useful for an independent synthetic demonstration. It can be misleading when several rows come from the same physical design, component or simulation campaign. In that case, hold out complete groups and disclose the grouping rule.

• Report every target separately: R², MAE, RMSE, MAPE and a high-percentile error.
• Inspect residuals and worst cases across voltage, load, frequency and component regions.
• Retain one final test population after method choices are frozen.
• Call a min/max check “within training bounds”, not proof that an input is in-distribution.

Ask whether more representative data still help

Train on nested subsets while keeping the test set unchanged. If error still falls at the largest size, more representative cases may help. If it plateaus above the required accuracy, inspect missing inputs, regime changes, noisy labels or model bias before generating more rows from the same distribution.

Use a deployment ladder

1. Reproduce the method locally with a versioned dataset and fixed seed.
2. Expose validation, learning-curve and failure evidence.
3. Provide a browser mode for bounded datasets and a full local mode for large campaigns.
4. Screen candidate designs with the surrogate, then recheck them in the governing source.
5. Move toward hardware only through an explicitly separate validation programme.

Apply the workflow to a Buck converter

The public Buck case uses five operating and design inputs to estimate total loss and inductor-current ripple. Its bundled equations make every step reproducible; the same CSV contract can later receive higher-fidelity simulator or laboratory data without pretending that those sources are equivalent.