RT

Article ·

Thinking in power electronics

A design is a walk through a graph of choices. Each step has a price. Write the path down, and another engineer can review the result.

The answer is not in the formula

Most students expect one correct design. Take the specification, apply the formula, read the component value. A real converter is not built that way.

Two plus two is still four. But every design is the end of a path of choices. A different path ends at a different converter, and that converter can be equally valid. The optimal design for one customer is not the optimal design for the next one.

The graph

Think of a design as a graph. A node is the state of the design: the specification plus every value chosen so far. An edge is one choice. Every edge has a price in a named currency: efficiency, money, volume, time or risk.

The space of solutions is the set of nodes you can reach. The formulas give the price of each edge. They do not tell you which edge to take. That part is engineering.

One walk, with numbers

Lesson 1 of my HTWG course follows TI’s PMP31349 buck: a 30–60 V bus to 12 V at 20 A, switching at 350 kHz.

The first edge is the inductor ripple. At 10 % ripple and 60 V input, the inductor needs 13.71 µH and the output capacitor carries 0.577 A RMS. At 100 % ripple, the inductor needs 1.371 µH and the capacitor carries 5.774 A RMS. You pay a factor of ten in inductance or a factor of ten in capacitor current. We chose 30 % and continued with TI’s 4.7 µH. Both neighbours are valid paths. They end at different converters.

The second edge is the output capacitor. The ripple limit alone asks for 17.37 µF. A 10 A load step at a 35 kHz crossover asks for 126.3 µF. The ripple-only choice fails by more than a factor of seven. That branch is a dead end. The lesson keeps it on the page, marks it No!, and walks back one node.

Four rules for the walk

  1. Label every limit before the first step. Is it given by the customer, or chosen by you? A missing limit is a decision, not a free parameter.
  2. Price every step. Name the choice, one benefit, one cost and one credible alternative.
  3. Name the measurement that would change your mind. Without it, nobody can review the decision.
  4. Keep the dead ends. A failed branch tells the next reader why the path turns. Delete it, and the next engineer walks into it again.

The price of an edge is not always paid at once. I have seen devices fail in the field after a dead-time change from 300 ns to 200 ns. The step looked small on the bench. Its price arrived months later.

The same walk outside the converter

Power electronics is where I learned this method. It is not the only place where I use it. A team grows through choices about who owns which seam between disciplines. An AI agent walks the same graph faster than a person, and it forgets the path faster too.

FrameShift is my attempt to record the walk itself: frame the problem, explain the causes, generate structurally different options, stop at a human checkpoint, and write the decision record. Today it is a specification and a set of contracts, not a finished application. pyPLECS makes one kind of edge cheap: it turns a simulation into a repeatable evaluation of a choice.

Where to follow the walk