ALD · Process inference
Separate the delivery from the reaction.
Connect inlet history, surface response and prediction on unseen pulse conditions.
Fitted second rate multiplier
Values & comparison details
Difference = comparison − reference. Rounded for display; original data retained.
- True model value
- 1
- Fit at upper bound
- 3
The fit reaches its imposed upper bound. This difference does not validate a physical reaction rate.
Open full figure THE QUESTION
When does an inferred reaction rate absorb an incorrect delivery or observation model?
THE APPROACH
A dimensionless water half-reaction model compares a single-pole inlet with finite storage. Four training-selected fits use a single-pole delivery model with either acquisition averaging or a deliberately misspecified point observation. Two unseen pulse histories test the predictions.
WHAT THE COMPARISON SHOWS
Three things to take away.
- 01
Storage changes both the inlet time history and delivered amount under the same command. The larger conversion cannot be attributed to timing alone.
- 02
With buffer truth, the fitted second rate reaches the imposed upper bound of 3. A constrained solution does not make the fitted mechanism physically correct.
- 03
The worst held-out scaled RMSE is about 0.1111 for buffer truth with averaged observations. The matched, noiseless recovery control is about 9.83 × 10⁻¹². These are numerical comparisons, not sensor precision.
Evidence, assumptions & the next question
What was numerically checked
All 909 retained chemical records passed an independent balance and ledger audit. Separate formulations and refined references supplied 69 pairwise checks across 23 control, truth and prediction trios.
How far the result goes
Dimensionless, noiseless and well-mixed half-reaction model with fixed wall and sensor assumptions. Conversion is not film thickness or growth per cycle. Scaled RMSE is not a percentage or a noise estimate.
What a physical study would need next
Measure the species-specific inlet history and acquisition response independently, then preserve matched run identities before interpreting physical reaction constants.
SA-PROC-ALD-TRANSIENT-002 · Delivery-buffer comparison, ALD016 · 2026-09-18
The download contains selected plot data, captions and a redraw script. Source records and hashes keep the public figures traceable to the completed study.
RESEARCH SUPPORT · FIGURES · MANUSCRIPTS
Bring your question. Build the full story.
We can help organize your literature and numerical results, build scientific diagrams and figure sets, develop original cover concepts, and refine the manuscript around them.