Nathan Ku

Notes

Short write-ups of findings, trade-offs, and cases where an experiment disproved an earlier assumption.

Note 01 · Multiphysics optimization

Pausing cool-down at 140 °C cuts packaging stress by a third

When silicon resonant accelerometers cool after adhesive curing, thermal expansion mismatch across substrate, epoxy, and silicon die leaves residual peeling stress and warpage that drift sensor readings.

Across 98 parameterized 3D transient viscoplastic finite-element simulations (COMSOL coupled with MATLAB LiveLink), introducing an isothermal stress-relaxation hold at 140 °C let the epoxy relax while still compliant—lowering modeled room-temperature peeling stress by 32.5% (7.70 vs 11.40 MPa linear baseline) and reducing wafer warpage by 32.4% (0.0071 vs 0.0105 µm). A 141 °C / 450 s hold kept 3-cycle thermal ratcheting drift below 1.5%.

Boundary: Simulation results under nominal constitutive parameters rather than post-fabrication interferometry; the thesis manuscript and 200.6 GB solved mesh archive remain offline.

Note 02 · Negative result & verification

Why I withdrew a 99.99% simulated airflow capture claim

Early 3D finite-element runs of a closed-loop Coandă push–pull air curtain (41,454 nodes, 219,986 tetrahedra) reported apparent bioaerosol containment near 99.99%. On paper, the result looked complete.

Cross-checking the model exposed four numerical defects: discrete pressure–velocity inconsistency and open-boundary mass leakage in the original incremental solver; negative scalar concentrations that were silently clipped; 506 underdetermined mesh cells and failed high-flow (20 CFM) convergence in an independent OpenFOAM run; and transient time-step non-independence when refining dt.

Rather than tuning around failed checks, all earlier capture percentages were explicitly downgraded to unvalidated historical outputs, solver tuning was paused, and evaluation shifted toward low-order mechanism screening and direct physical bench testing.

Boundary: No physical containment or product claim is made. Historical FEM flow fields remain viewable strictly for spatial exploration.

Note 03 · Numerical robustness

Why 16-bit fixed point breaks speech before the neural network even runs

Quantizing log-Mel features after log compression is forgiving: both 16-bit fixed point (Q15) and block floating point (BFP16) matched float32 on 30/30 test clips (2.94% CER).

Quantizing linear acoustic energy before log compression behaves completely differently. Because raw speech power spans orders of magnitude between loud vowels and quiet consonants, a global Q15 full-scale rounds quiet frames to zero—driving downstream ASR error to 41.63% (only 6/30 transcripts intact). Sharing a single exponent per frame (BFP16) keeps the same 16-bit mantissa width while holding error at 2.71% (29/30 intact).