Chemistry / QM algorithms — SOTA survey and Li gap analysis (chem-r0-qm-sota-survey)¶

Goal: chem_sim_algorithms · Session: 1bdb6322-8399-425d-9257-9b9098475e89 · Run: numerics_researcher-1779916590880
Agent: numerics_researcher · Mode: study-only (validity locked; no perf claims)
North star: PH-5b (proved numerics), PH-7e (math→SIMD on integral/ERI loops), G-math / G-par for AO contraction kernels
Preflight: benchmarks/data/latest/ecosystem-audit.json @ 2026-05-27T21:03Z · dashboard

Problem¶

Li’s chemistry vertical must reach Gaussian/ORCA/Psi4-class minimal workflows (AO integrals → Fock build → SCF → DFT energy) without trading SCF convergence honesty or basis-set correctness for ratio_vs_cpp. Today algo_registry 401–432 rows are catalog compile smokes (implemented_smoke: true) wired to the schrodinger_1d_barrier family template — not quantum-chemistry kernels. li-sim-scientific routes vertical_qm_dft() to run_algo_registry_stub for all registry QM ids except the vertical entry point, which calls algo_qm_dft_scf_energy() but still returns a stub checksum. All 32 qm_* catalog rows are unknown on the dashboard (harness pending); no red QM timing rows yet because timing is deferred.

Learned from (SOTA)¶

1. Psi4 — Python QC reference; minimal Hartree–Fock/DFT energy APIs, basis-set library, DIIS SCF.
2. Docs: https://psicode.org/ , https://psicode.org/psi4manual/master/tutorial.html

3. Takeaway for Li: v1 oracle path = Psi4 single-point on H₂/H₂O with STO-3G and 6-31G* before native RKS; use as external_binary honesty in verticals.toml.

4. PySCF — Modular AO integrals, density-fitting ERIs, RKS/UKS drivers; cost scales as O(N_basis⁴) naive ERIs, O(N²–N³) with DF.

5. Docs: https://pyscf.org/user.html , https://pyscf.org/user/scf.html

6. Takeaway: Implement 401–404 integral chain (GTO eval, overlap, kinetic, nuclear attraction) before 408 Fock or 418 qm_dft_scf_energy; density fitting (407) is a PH-7e target, not v1.

7. ORCA — Production DFT (GGA/hybrid/grid); Becke/Lebedev grids and XC families map directly to registry ids 412–417.

8. Manual: https://www.faccts.de/orca/manual/ , https://doi.org/10.1063/1.4824486 (RI-JK)

9. Takeaway: Grid/XC rows stay stub honesty until LDA (412) passes reference energies on a fixed geometry.

10. Helgaker et al. / Szabo & Ostlund — Standard AO integral recurrences, SCF stability, basis-set limit behavior.

11. Text: Molecular Electronic-Structure Theory (Helgaker, Jørgensen, Olsen); Modern Quantum Chemistry (Szabo, Ostlund)
12. Takeaway: Basis-size table (chem-r1) must show monotonic energy improvement and O(N_basis⁴) wall-time trend for naive ERIs before perf claims.

SOTA → Li mapping (algo_registry 401–432)¶

Vertical honesty (benchmarks/competitive/verticals.toml in sim worktrees): qm_dft — incumbent Gaussian/ORCA/Psi4, workload_class=stub, oracle=external_binary, notes not parity until import_chem_dft_smoke composable gate passes.

Basis-size scaling (survey targets — fill in chem-r1)¶

Reference: PySCF H₂ cc-pVDZ vs STO-3G single-point timings. Li stub must reproduce trend before absolute Ha parity.

Repro (when native stub exists):

cd lic/benchmarks/harness
python3 bench.py --tier 2 --only qm_dft_scf_energy --runs 3
cd benchmarks && LIC_ROOT=../lic ./scripts/ingest/ingest-lic.sh

Implementation path in lic (contracts + bench evidence)¶

1. Validity first (PH-5b): Add common/qm_scf_core.c (or reuse Psi4 subprocess oracle) for H₂ STO-3G total energy; verify-results checksum before wall time. No threshold_ratio_cpp relaxation (catalog default 1.2).
2. Integral microkernel chain (401–404): Port closed-form s-type overlaps then general (μν|λσ) via Obara–Saika recurrence in packages/li-physics-quantum (new) or extend std/physics/quantum.li beyond tag stub.
3. SCF driver (411, 418): DIIS + level-shifting in qm_scf_solver; li-sim-scientific must call real run_qm_dft_scf_energy_smoke instead of run_algo_registry_stub for id 418.
4. PH-7e / G-par: ERI contraction and Fock build — @vectorized / parallel for (disjoint=) only after lic build proves integral symmetry lemmas.
5. Composable gate: li-tests/composable/import_chem_dft_smoke.li + qm_dft.toml; flip verticals.toml workload_class only when smoke passes.
6. Worktree: Deep gates in lic-worktrees/sim-chem-research only — no new systemd sim loops on main.

Do not: weaken threshold_ratio_cpp; ship sorry/unsafe for speed; copy harness into benchmarks repo; claim Gaussian/ORCA parity from schrodinger_1d_barrier template timings.

Grade matrix¶

Tradeoffs¶

• Locked: validity (SCF convergence, reference energies), accuracy (basis-set monotonicity), registry honesty for template smokes.
• Improved: Clear SOTA→registry map; v1 scope = 418 + integral chain 401–404; Psi4 external oracle path documented.
• Regressed: none (survey-only).
• Explicitly not approved: trading tier-0 tolerances or threshold_ratio_cpp for green dashboard cells; production DFT accuracy claims.

Evidence¶