{"id":"t1-wander-mechanism","title":"T1 Wander Mechanism — Metallic Saddles and SCF Fragility (2026-07-22)","subtitle":"Path-0 diagnostic: the GPAW↔VASP offset wander is intrinsic to metallic, near-zero-gap saddles rather than a settings artifact; the adopted settings are validated as converged on the worst-behaved path.","category":"references","tags":["t1","wander","gpaw","scf","metallic-saddle"],"source":"articles/docs/analysis/t1-wander-mechanism.md","lang":"en","words":407,"readMinutes":2,"toc":[],"html":"<h1 id=\"t1-wander-mechanism-metallic-saddles-and-scf-fragility-2026-07-22\">T1 wander mechanism — metallic saddles and SCF fragility (2026-07-22)</h1><p><strong>Question:</strong> why does the GPAW↔VASP offset wander (T1) vary 30× across the Z1 panel — 135 meV (path-7) to 4212 meV (path-0)?</p>\n<p><strong>Evidence (path-0, Ag-F-Li, <code>mp-761269_2_1_1_-1_0</code>):</strong></p>\n<ul>\n<li>Adopted settings (h=0.20, Gamma): img3 GPAW <strong>converged</strong> at −492.718 eV (matches campaign checkpoint).</li>\n<li>h=0.18, Gamma: img3 <strong>did not converge</strong> — 454 iterations, residual exploding to 12.4. The −471 eV figure from that run is meaningless.</li>\n<li>Electronic structure at the saddle: <strong>Gap = 0.018 eV (metallic), Fermi level ≈ 0.67 eV</strong>; endpoints are well-behaved.</li>\n<li>Both engines agree the saddle is img3 and the basin is img0 — same extrema — but GPAW puts the saddle 5.79 eV above the basin where VASP (LiTraj nebDFT2k, CI-NEB PBE) puts it 1.58 eV.</li>\n</ul>\n<p><strong>Finding:</strong> the wander is <strong>intrinsic to the system class</strong>, not a settings artifact. Metallic / near-zero-gap transition states in GPAW fd mode (default smearing) vs VASP CI-NEB settings converge to different electronic descriptions on the saddle image specifically; the offset wander concentrates on exactly that image (per the same-extrema identity, <code>barrier_sub_eq_offset_sub</code>, the entire barrier error is the saddle-vs-basin offset difference).</p>\n<p><strong>Consequences:</strong></p>\n<ol>\n<li>SCF convergence on this cell class is fragile across the settings we tested: h=0.18 at Gamma did not converge (454 iterations, residual exploding), and the full frozen profile (h=0.18, (2,2,2)) — not separately rerun here — already showed the pathology at production scale: path-16 consumed 16.3 h / 62 CPU-hours at frozen settings without producing a receipt. The adopted loosened profile (h=0.20, Gamma) is the only setting so far observed to converge reliably on this class. A full frozen-mesh rerun on this cell remains a separate (expensive) confirmation, not a requirement for the conclusion that matters: the adopted anchors are converged data.</li>\n<li>The union campaign&#39;s adopted settings are validated as <em>converged</em> on the worst-behaved path; anchors are usable data.</li>\n<li>The same-engine basis (amendment 01) is the only meaningful verdict basis for metallic-saddle paths; VASP-referenced numbers for those paths are T1-contaminated <strong>with a mechanism</strong>, not just a flag.</li>\n<li>The T1 law (<code>abs_barrier_sub_le_wander</code>) prices the failure: measured VASP-referenced MAE 1246 meV vs mean wander 1269 meV over the first 7 complete paths.</li>\n<li>Future engine-equivalence work on LiTraj-class panels must address smearing/occupation policy at metallic saddles explicitly (e.g., wider smearing, occupation control, or spin-polarized restarts) — a candidate theorem line: saddle-metallicity detector (gap &lt; ε at the model-predicted saddle ⇒ require occupation audit before any cross-engine verdict).</li>\n</ol>\n<p><strong>Artifacts:</strong> <code>/tmp/z1-diagnose/img3-adopted.txt</code>, <code>/tmp/z1-diagnose/img3-h018.txt</code> (GPAW outputs); campaign receipts <code>/tmp/z1-union-local/anchors/path-0/</code>.</p>\n"}