Skip to content

Research goal digest — chem_sim_algorithms

Session: 1bdb6322-8399-425d-9257-9b9098475e89 · Cycle: 1 · Agent: numerics_researcher
Run: numerics_researcher-1779940539765 · Generated: 2026-05-28T03:56:40Z
North star fit: Chemistry / QM — PH-5b (proved numerics), PH-7e (integral/Fock SIMD), domains: scientific_computing, hpc

Executive summary

  • Mode A SOTA survey complete (chem-r0): Psi4, PySCF, ORCA, Helgaker/Szabo mapped to algo_registry 401–432; implement 401–404 (integrals) before 408/418 (Fock/SCF energy).
  • Li gap: All 32 qm_* catalog rows remain unknown on the dashboard (preflight 2026-05-28); tier-2 harnesses still use the schrodinger_1d_barrier family template, not QC kernels.
  • li-sim-scientific: vertical_qm_dft() → algo_qm_dft_scf_energy() → run_algo_registry_stub (checksum 1.001); registry ids 401–432 route through the same stub path.
  • std/physics/chem.li: tag-only (physics_chem_std_tag); no GTO/SCF surface on main yet.
  • Composable gate missing: li-tests/composable/import_chem_dft_smoke.li not on main; verticals.toml honesty stays stub / external_binary until it passes.
  • v1 implement target: chem-r2-dft-scf-gap → qm_dft_scf_energy (418) with Psi4 subprocess oracle (H₂ STO-3G) before native RKS perf; validity/stability/accuracy axes locked.
  • Evidence pack: chem-r0 whitepaper · study · catalog qm_dft_scf_energy (size_label=harness pending).
  • No perf claims this cycle; threshold_ratio_cpp (1.2) and tier-0 tolerances unchanged.

Deliverable / findings

Completed artifacts (cycle 1, session 1bdb6322)

Step Run Artifact
survey_sota-1 numerics_researcher-1779916590880 Study · Whitepaper
digest numerics_researcher-1779940539765 This file · Session log

Learned from (SOTA, 2–4)

  1. Psi4 tutorials — minimal HF/DFT energy, basis sets, DIIS SCF → v1 external oracle for H₂/H₂O STO-3G.
  2. PySCF user guide — AO integrals, density fitting, RKS drivers → 401–404 before 408/418.
  3. ORCA manual — GGA/hybrid/grid (registry 412–417) → stub honesty until LDA (412) reference energy on fixed geometry.
  4. Helgaker, Jørgensen, Olsen — Molecular Electronic-Structure Theory → basis monotonicity required in chem-r1 scaling table.

Li mapping (PH / proof)

Surface PH / group Status
SCF convergence + Ha refs PH-5b Documented; Psi4 oracle path proposed
ERI/Fock contraction loops PH-7e Deferred until integral parity
Gaussian recurrence, symmetry G-math chem.li / future li-physics-quantum
AO block parallelism G-par Post-proof @vectorized / parallel for

In-repo gap evidence (verified 2026-05-28)

```167:195:packages/li-sim-scientific/src/lib.li def run_algo_registry_stub(algo_id: int, detail: int) -> SimRunResult ... if algo_id >= 401: if algo_id <= 432: vert = vertical_qm_dft() ... r.checksum = 1.001 

```246:248:packages/li-sim-scientific/src/lib.li
  if vertical_id == vertical_qm_dft():
    var d3: int = detail
    return run_algo(algo_qm_dft_scf_energy(), d3)

benchmarks/catalog.toml — qm_dft_scf_energy: size_label = "harness pending", threshold_ratio_cpp = 1.2, validity_required = true.

Grade matrix

Axis Li today Target Locked?
Validity registry smokes; vertical stub Psi4 Ha oracle on 418 yes
Stability SCF max-iter undocumented DIIS + damping table yes
Accuracy template only basis monotonicity (chem-r1) yes
Performance all qm_* unknown after validity no (deferred)
Memory N/A ERI profile post-407 no

Tradeoffs

Validity, stability (SCF convergence), and accuracy (basis-set monotonicity) remain locked. Speed on ERI/Fock rows is explicitly deferred until checksum/Ha parity. External Gaussian/ORCA binaries are not required in CI; Psi4 subprocess oracle is optional and documented. No axis may regress to green dashboard cells without human approval of locked-axis tradeoffs.

Mandatory evidence (this run)

Type Path
Numerics study docs/numerics/studies/2026-05-27-chem-r0-qm-sota-survey.md
Whitepaper research-findings/whitepapers/2026-05/chem_sim_algorithms/chem-r0-qm-sota-survey/
Bench catalog benchmarks/catalog.toml → qm_dft_scf_energy
Preflight benchmarks/data/latest/ecosystem-audit.json (32 unknown qm_*)

Implementation path (handoff → bench_improver / sim worktree)

  1. Add common/qm_scf_core.c or Psi4 subprocess oracle for H₂ STO-3G; wire qm_dft_scf_energy harness off schrodinger_1d_barrier template.
  2. Implement integral microkernels 401–404 before Fock/SCF.
  3. Replace run_algo_registry_stub for id 418 with real run_qm_dft_scf_energy_smoke.
  4. Add li-tests/composable/import_chem_dft_smoke.li; flip verticals.toml only when green.
  5. Deep gates in lic-worktrees/sim-chem-research only — no new systemd sim loops on main.
Title Repo Labels Owner agent
chem-r2: qm_dft_scf_energy harness + Psi4 oracle (418) lic numerics-research, bench bench_improver / sim-p2-qm-dft-scf
chem-r1: fill STO-3G / 6-31G / cc-pVDZ scaling table with Psi4 refs lic numerics-research numerics_researcher
Add import_chem_dft_smoke composable gate lic numerics-research, li-tests code_implementer
Honesty: verticals.toml qm_dft stub → oracle after smoke benchmarks numerics-research bench_improver
Integral chain 401–404 — s-type overlap smoke lic numerics-research, PH-5b bench_improver

PR template reminder (when opening lic PR):

<!-- li-agent -->
## Agent deliverable
- [x] Tests added or updated — cite `docs/numerics/studies/2026-05-27-chem-r0-qm-sota-survey.md`
- [x] Bench evidence — catalog `qm_dft_scf_energy` (harness pending; dashboard unknown)
- [x] No merge-approved until human review

Deferred

  • chem-r3-package-placement — package layout (li-physics-quantum vs std/physics/quantum.li) after r2 smoke.
  • Post-HF rows 422–425 (MP2/CCSD/TDDFT) — after 418 green.
  • Grid/XC 412–417 — after LDA (412) reference energies on fixed geometry.
  • Native RKS perf / PH-7e SIMD — after chem-r1 basis scaling table filled with measured data.
  • Production DFT accuracy — out of v1 scope per scaffold.
  • Autoresearch — novel integral/SCF splittings only if SOTA path insufficient (numerics-autoresearch gates).