Blog
Technical Notes from the Lab
Honest engineering writing about glass interposers, semiconductor yield, physics solvers, and the things we got wrong before we got them right.
v3.6 Release: A1 Surrogate R²=0.9520, Palace FEM at 12.3%, 28 Filing Packages
A1 neural surrogate converges on 15.9M-row corpus. Palace FEM independently reproduces BEM Z₀ at 12.3% median offset. Portfolio triaged to 29 filing packages.
Why Coaxial Approximation is 55% Wrong for Glass TGVs
Everyone uses the coaxial formula for TGV impedance. It's wrong by 55%. Here's why multiconductor BEM matters.
How We Found and Fixed 25 Inflated Claims in Our Own Portfolio
Radical transparency: we retracted 25 claims, documented every correction, and made the portfolio stronger.
10 Wafers: The Minimum KLA Archer Scans for Yield Calibration
Our 10,000-campaign Bayesian DOE proves 10 wafers is the minimum for CI<20µm on correlation_length.
No Commercial EDA Tool Supports Glass Interposers. We Built One.
Cadence, Synopsys, and Ansys don't support glass. Here's the platform we built and why it matters.
From R²=0.8725 to R²=0.07: What We Learned About FNO Overfitting
Our FNO trained on narrow data looked great. On realistic data, it collapsed. The honest story of how we found and fixed it.