Skip to content

v0.7.27

@opticsWolf opticsWolf tagged this 29 Sep 20:03
Six error channels -- thickness, n, k, grading amplitude, roughness and
interface width -- perturb a design per layer so a tolerance study answers
what a chamber will actually deposit. Every default in that subsystem was
wrong, two of its four laws were wrong, and one dispatch wrote the
neighbouring channel. None of it had been noticed, because error_mask
defaults to all-zero: nothing draws until a caller switches a channel on,
so the whole subsystem sat behind a gate no test had opened.

Behaviour changes a caller can see:

  * Every default moved. The relative spreads were a hundred times too
    wide -- 1.0 where 0.01 was meant, a 100% one-sigma tolerance. The
    absolute spreads now differ per channel, because an absolute spread
    carries the unit of the quantity it perturbs and the six do not share
    one: 0.5 nm on a thickness, 0.001 on an index, 0.0001 on k. The old
    shared nanometre-sized default put roughly half of all k draws below
    zero, which is optical gain, so a tolerance run died at the gain door
    rather than reporting a tolerance.
  * The uniform law got its centre back. U(mean +- variance) had been
    drawn as U(+-variance), so a deliberate offset silently did nothing.
  * ErrorType::Cascaded is new: the two relative stages compose as
    factors, v(1 + G)(1 + U), rather than summing. That is the right
    shape when the errors are stages in series -- a mis-calibrated rate,
    then a per-run monitor error acting on the already-wrong deposit.
    Two stages of 10% compose to 21%, not 20%, and that missing percent
    is the entire difference from Combined.
  * set_properties no longer writes the wrong channel. Its key-to-slot
    map carried a transcription error, so setting interface_error_type
    moved roughness_error_type instead. The map is now three exhaustive
    slot resolvers, where a missing arm is a compile error.
  * A config document's per-channel defaults apply to a partial block,
    not only to an omitted one. Naming a single field used to fall back
    to one shared default for all six channels; a block is now a patch
    onto its own channel's defaults.
  * An incomplete params block is called incomplete rather than
    malformed, lists every absent field at once instead of serde's first
    one, and suggests the copy idiom that fixes it. A params block still
    replaces a channel whole -- that contract is unchanged.

Anyone who had switched a channel on before this tag should re-run:
the draws they got were not the distribution they asked for.

Verification: each of the four laws now has its measured sigma pinned
from Python, so Combined and Cascaded cannot collapse into each other
unnoticed, and the thickness tolerance is checked against a real chamber.
Every fix in the series had its test watched failing first, with the
failure naming the offending channel or field.

Also in this tag: the README finally documents any of it. Across the
eleven releases above it mentioned "error" exactly once, and only about
Nevot-Croce roughness -- the same silence that let the defects survive.
The release section now says to read the feature list against what the
release added, and admits the gap it found while looking: the wheel
filename there is a seventh version site the CI gate cannot check,
because it is prose.

Full detail in CHANGELOG.md, sections 0.7.15 through 0.7.27.
Assets 2
Loading