Skip to content

MD algorithms — SOTA survey and Li gap analysis (md-r0-sota-survey)

Goal: md_sim_algorithms · Session: f1114f06-7079-45f3-9d88-ce5106130118 · Run: numerics_researcher-1779911656866
Agent: numerics_researcher · Mode: study-only (validity locked; no perf claims)
North star: PH-5b (proved numerics), PH-7e (math→SIMD), G-math / G-par for force loops
Preflight: benchmarks/data/latest/ecosystem-audit.json @ 2026-05-27T18:44Z · dashboard

Problem

Li’s MD vertical must reach LAMMPS/GROMACS-class neighbor lists, symplectic integrators, cutoffs, and (later) long-range electrostatics without trading checksum parity or tier-0 stability for ratio_vs_cpp. On lic main (2026-05-27):

  • md_lennard_jones — full tier-2 harness; brute O(N²) MIC forces in md_core.h.
  • md_neighbor_cell_list — WP2 catalog harness exists but #includes md_lennard_jones oracle (no cell traversal yet).
  • Registry 101–120 — composable smokes; most tier-2 rows are catalog honesty / shared oracle stubs.
  • Dashboard: all 16 md_* catalog ids are unknown (stale ingest @ 2026-05-26); org red is tier-1 horner_pure_li (0.67×) and reduce_sum (0.93×), not MD physics.

Deep implement contract for algo 105: 2026-05-25-md-r2-neighbor-list-gap.md.

Learned from (SOTA)

  1. LAMMPS — cell-linked neighbor lists (neighbor bin), pair cutoff + skin, velocity-Verlet integrators, SHAKE/RATTLE, Ewald/PPPM long-range.
  2. https://docs.lammps.org/neighbor.html · https://docs.lammps.org/Integrators.html

  3. Takeaway: O(N) forces need linked cells + skin (algo 105–106); rebuild criteria belong in params.toml before PH-7e work.

  4. GROMACS — grid neighbor search, LINCS/SETTLE constraints, PME for Coulomb; list buffer sized from max displacement.

  5. https://manual.gromacs.org/current/reference-manual/algorithms/index.html
  6. https://manual.gromacs.org/current/reference-manual/algorithms/neighbor-searching.html

  7. Takeaway: list update frequency vs energy drift is a validity axis; PME (114) deferred until cutoff+LJ path is green.

  8. OpenMM — NonbondedForce cutoff + switching; documented drift tests for thermostats/integrators.

  9. https://docs.openmm.org/latest/userguide/application.html

  10. Takeaway: thermostat/barostat rows (108–110) stay stub until NVE matrix (md-r1) is filled.

  11. Frenkel & Smit / Swope et al. 1997 — cell-linked O(N) at fixed ρ; NVE drift gates for integrator acceptance.

  12. https://www.sciencedirect.com/book/9780123872324/understanding-molecular-simulation
  13. https://doi.org/10.1006/jcph.1997.5740
  14. Takeaway: Li md_lennard_jones/PERF.md stress targets align; AT-grade NVE remains advisory until scaling table is measured.

SOTA → Li mapping (algo_registry 101–120)

ID Registry name Incumbent pattern Li package / bench Harness on main PH / proof
101 md_lj_cutoff_mic LAMMPS pair/lj/cut + MIC li-physics-particles, li-sim-scientific smoke Yes — md_lennard_jones PH-5b; LEM-PHYS-001
102 md_integrator_verlet GROMACS VV / LAMMPS verlet Shared md_core velocity-Verlet Via md_lennard_jones PH-5b
103 md_energy_drift AT NVE MSD stability.py, [conservation] in params.toml Partial Validity locked
104 md_oracle_external LAMMPS/GROMACS micro verticals.toml stub Stub Honesty only
105 md_neighbor_cell_list LAMMPS neighbor bin tier2_physics/md_neighbor_cell_list WP2 stub (shared brute oracle) PH-7e after F parity
106 md_neighbor_verlet_skin LAMMPS skin Catalog + stub path WP2 stub pattern After 105
107 md_integrator_leapfrog Leapfrog num_integ_* micro Tier-1 micro Near org red: integrator family
108–110 thermostats / barostat GROMACS/OpenMM Catalog Stub Defer post-NVE
111–112 SHAKE/RATTLE GROMACS LINCS / LAMMPS SHAKE Catalog Stub After neighbor parity
113–114 Ewald / PME GROMACS PME Catalog Stub After cutoff path
115–120 init / analysis Various md_init_fcc_mb etc. Catalog / partial Init before large-N perf

Li honesty gap (105): li_md_compute_forces remains O(N²) all-pairs; registry implemented_smoke: true on 105 is wiring only — see md-r2.

Size scaling (survey table — harness targets)

N ρ (LJ units) dt Expected dominant cost Li action
128 0.75 0.004 Force eval (small) Baseline stability (md-r1)
512 0.75 0.004 Neighbor rebuild amortization Skin sensitivity (106)
2048 0.75 0.004 Cell list + bandwidth PH-7e SIMD / G-par after parity

Repro:

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

Implementation path in lic (contracts + bench evidence)

  1. Validity (PH-5b): Keep md_core.c cross-lang oracle; verify-results checksum before any ratio_vs_cpp claim.
  2. Neighbor (105 → sim-p1-md-neighbor-cell): Half-shell cell list in md_core; gate max |F_cell − F_brute| @ N=256; see md-r2.
  3. Integrator micro (PH-7e): Tier-1 num_integ_verlet feeds rows 102/107 after harmonic tier-0 green.
  4. G-math / G-par: @vectorized / parallel for (disjoint=) on force loop only after drift lemma proves.
  5. Packages: li-sim-scientific smoke = algo 101; expand li-physics-particles with neighbor_* APIs.
  6. Ingest: Refresh dashboard so md_lennard_jones exits unknown — ecosystem visibility without weakening thresholds.

Do not: weaken threshold_ratio_cpp; ship sorry/unsafe for speed; copy harness into benchmarks repo.

Quality / improvement (survey)

Survey-only: no before/after perf claims. Improvement = clarified registry→harness map and prioritized algo 105 implement handoff. Regression: none.

Visuals: deferred to md-r1-stability-matrix (energy drift PNG) and tier-2 GIF ingest after neighbor parity.

cd benchmarks && LIC_ROOT=../lic ./scripts/render-benchmark-visuals.sh

Grade matrix

Axis Result vs prior Notes
Validity pass (survey) refreshed LJ tier-2 green; 105 harness exists but oracle-shared (honest stub)
Performance document only — MD rows unknown on dashboard; org red unrelated to MD
Memory N/A — Defer to sim-bench-memory.sh post-neighbor
Security pass — No new FFI this survey
Stability partial — md-r1 matrix defines CFL/skin; AT NVE advisory
Size scaling table attached — ≥3 N; metrics in md-r1

Tradeoffs

  • Locked: validity + stability (NVE drift, checksum vs md_core, registry honesty).
  • Improved: Corrected harness inventory (105 path exists, not missing); ingest stale called out; handoff sim-p1-md-neighbor-cell prioritized.
  • Regressed: none (survey-only).
  • Not approved: relaxing tier-0 or threshold_ratio_cpp for dashboard green.

Evidence

Type Path / command
Study docs/numerics/studies/2026-05-27-md-r0-sota-survey.md
Prior docs/numerics/studies/2026-05-26-md-r0-sota-survey.md
Neighbor contract docs/numerics/studies/2026-05-25-md-r2-neighbor-list-gap.md
Whitepaper research-findings/whitepapers/2026-05/md_sim_algorithms/md-r0-sota-survey/
Audit benchmarks/data/latest/ecosystem-audit.json
Catalog benchmarks/catalog.toml — md_lennard_jones, md_neighbor_cell_list
Harness lic/benchmarks/tier2_physics/md_lennard_jones/, md_neighbor_cell_list/
li-tests li-tests/composable/import_sim_scientific_run.li
Bench md_lennard_jones @ 48d23a7a: verify drift=0.689; li/cpp≈0.996× — python3 lic/benchmarks/harness/bench.py --tier 2 --only md_lennard_jones