A research-grade quantum chemistry + many-body physics engine that runs entirely in a browser tab. No install. No backend. No CUDA. Open a URL and get HF · UHF · DFT (RKS+UKS) · MP2 · CCSD · CCSD(T) · EOM-CCSD · TDDFT α(ω) · C₆ dispersion on real molecules — with GPU acceleration via WebGPU. Molden / Cube / QCSchema / XYZ exports for visualization + interop.

---

📸   See it

landing · what the engine is + run-anywhere CTAs

/molecule.html · H₂O SI report — properties, spectra, gradients

/experiments/ · research dashboard (E1–E33, JSON artifacts)

---

📊   The numbers _{^{(single source of truth — see bottom of file)}}

---

🧪   What's inside

---

⚡   How fast — honestly, both directions

_{The headline 39.3× on CCSD(T) is real. So is the fact that PySCF/NumPy is 480× faster than us on CCSD at cc-pVDZ. Both numbers come from the same comparison run, against PySCF 2.13.0 on identical inputs.}

_{↑ Single-run measurements (not warmup+trials harness). Energy agreement ≤ 10⁻⁴ Ha on all 19 comparable cells (well below chemical accuracy of 1.594 mHa). Where we win: no Python startup, HF up to medium systems, GPU CCSD(T) at cc-pVDZ. Where we lose: CPU MP2 / CCSD at production basis where NumPy / BLAS dominates. Full data: E34-comparison.md.}

---

🆚   Completeness scorecard

_{Every method PySCF 2.13 / ORCA 6.1 / Psi4 1.10 ship, our status against it, and the roadmap tier for every gap.
No "we don't do that" — every missing capability has a planned slot.}

---

🔬   Validation matrix

---

🪜   The research ladder

---

⏱️   60-second demo

git clone https://github.com/abgnydn/webgpu-q && cd webgpu-q
npm install
npm run dev          # http://localhost:5175
                     # /molecule.html → H₂O SI report
                     # /experiments/  → research dashboard

npm run test         # vitest unit/integration · CI green
npm run typecheck    # tsc --noEmit, strict + noUncheckedIndexedAccess
npm run test:e2e     # Playwright · headless WebGPU Chromium

// Or one-call from molecule to a full property report
import { molecules, quickReport, uvVisSpectrum, toMoldenString } from "./src/chemistry";

const report  = quickReport(molecules.h2o, { addD2: true, addStaticAlpha: true });
const uvvis   = uvVisSpectrum(molecules.h2o, { method: "b3lyp5" });
const molden  = toMoldenString({ atoms: molecules.h2o, /* ... */ });
// → energy, dipole, charges, bond orders, NOON, ⟨S²⟩, multireference
//   verdict, static α, D2 dispersion, UV-vis spectrum + peaks, Molden file.
//   All in a Chrome tab.

// Or piece by piece if you want control
import { runRHFSCF, runMP2, runCCSD, runCCSDT_GPU, runEOMCCSD } from "./src/chemistry";

const hf      = runRHFSCF(integrals, nElectrons);
const mp2     = runMP2(hf, integrals);
const ccsd    = runCCSD(hf, integrals);
const t       = await runCCSDT_GPU(ccsd, hf, integrals, device);  // 39× on cc-pVDZ
const excited = runEOMCCSD(ccsd, integrals, hf);

---

🧱   Architecture · URL → silicon

Research harness · experiments/lib/ runner.ts — timedRun with forced GPU sync (read-after-submit on a tiny buffer) seeds.ts — named deterministic seeds (no Math.random()) env.ts — captures adapter info, limits, SHA, UTC fidelity.ts — F = |⟨ψ_ref|ψ_test⟩|², not max|Δp| stats.ts — median, p10/p90/p99, IQR

Discipline (non-negotiable) 5 warmup + 20 trials per measurement Pass bar: F ≥ 1 − 10⁻⁵ (f32 GPU paths) std/median > 0.1 → status: "noisy" Honest negatives committed as JSON with diagnosis vitest + Playwright e2e · CI green · TS strict + noUncheckedIndexedAccess

---

📚   Method catalog

Ground-state electronic structure · HF · UHF · DFT · MP2

method	notes
RHF SCF	DIIS, frozen-core, spherical-d, f/g/h, level-shift
UHF SCF	open-shell, stacked α+β DIIS, ⟨S²⟩ check, level-shift
RKS-DFT (LDA, GGA, hybrid)	LDA · BVWN5 · BLYP · B3VWN5 · B3LYP5; Becke grid, Lebedev
UKS-DFT (LDA, GGA, hybrid)	full functional ladder, spin-polarized XC kernel, ⟨S²⟩
MP2 · DF-MP2	spin-orbital + B-tensor reformulation
Cholesky DF (CD-DF)	rank-3 B-tensor, threshold-controlled
HF / DFT analytical ∇	Pulay 1969, 8-fold ERI loop, Schwarz screening

Correlation & excited states · CCSD · CCSD(T) · EOM-CCSD · CIS · TDDFT

method	notes
CCSD (RHF)	Stanton-Bartlett, antisym spin-orbital + frozen-core
UCCSD (UHF)	shared ccsdIterate core, 3-block ERI
CCSD(T) CPU	per-triple, FCI-validated ≤ 0.25 mHa, frozen-core via Set
CCSD(T) GPU	39.3× on H₂O cc-pVDZ, f32→f64 reduce
UCCSD(T)	open-shell perturbative triples, frozen-core via Set
EE-EOM-CCSD	Stanton-Bartlett σ + stage-32c diagonal patch, Davidson
IP-EOM-CCSD	R₁ exact (brute-force); R₂ open
EA-EOM-CCSD	R₁ + R₂ patched to exact (stage 32e)
CIS · TDA · TDDFT (Casida)	full functional ladder, triplet via spin-pol, Davidson
Counterpoise / BSSE	HF / MP2 / CCSD / UHF / UCCSD / RKS / UKS + optional D2 add-on
Oscillator strengths	f = (2/3)·ω·
Spin classifier	singlet/triplet/spin-flip weight per root

Properties & spectroscopy

property	notes
Dipole μ	AO→MO transform, RHF + post-HF densities
Polarizability α(0)	finite field + analytical CPHF (RHF + UHF)
Polarizability α(ω)	TDHF + TDDFT + open-shell UHF-TDHF response
Polarizability α(iω)	imaginary-axis response for Casimir-Polder
C₆ van-der-Waals coefficients	Casimir-Polder integral; HF/UHF/DFT references
Hyperpolarizability β	3D finite-field stencil
Mulliken populations	charges + spin-density resolved (closed + open shell)
Wiberg / Mayer bond orders	+ atomic valences, Lewis-multiplicity inference
Natural orbital occupations	NOON, multi-reference diagnostic
D2 dispersion correction	Grimme JCC 2006, energy + analytical gradient
Multireference verdict	T1/D1/⟨S²⟩/NOON aggregator with cutoff flags
TRK sum rule	oscillator-strength conservation check
Foster-Boys / Pipek-Mezey	orbital localization (Boys 1960 + PM 1989)
Energy decomposition	one-electron + Coulomb + exchange + V_nn breakdown
Coordination numbers	smooth Grimme D3 CN for chemistry-aware features
Coulomb matrix descriptor	permutation-invariant ML feature
Molecular formula / graph	Hill convention, adjacency, connected components
RMSD + Kabsch alignment	optimal rotation between geometries
Standard orientation	COM + principal-axes alignment
Rotational constants	A/B/C in cm⁻¹ and GHz
Multi-frame XYZ trajectory	for geom-opt / NEB visualization
Pre-built molecule library	h2o, ch4, nh3, beh2, hf, etc.
quickReport(atoms)	one-call full property report
uvVisSpectrum(atoms)	atoms → excitations + broadened spectrum + peaks
Harmonic ω	mass-weighted Hessian by finite diff
IR intensities	dμ/dQ along normal modes
Raman activities	Placzek invariants from α(Q)
Thermo (Sackur-Tetrode + RR + HO)	H₂O entropy 45.06 vs expt 45.1
Koopmans / ΔSCF / EOM IPs	H₂O: 10.65 / 8.36 / 12.03 eV (expt 12.62)
Koopmans / EOM EAs	H₂O: −16.48 / −16.37 eV
Molden orbital export	Cartesian Gaussian basis, Jmol/Avogadro/Multiwfn
Gaussian Cube export	density + MO isosurfaces, VMD/Jmol/Multiwfn
QCSchema JSON export	MolSSI standard, QCEngine/QCFractal/cclib consumable
XYZ format I/O	parse / emit standard geometry files

Geometry & basis sets

feature	notes
BFGS geom-opt	analytical HF + DFT gradients
Lebedev angular grids	2.6× point reduction at better accuracy
STO-3G	H, He, Li, Be, C, N, O, F (full first + second period)
6-31G*	available
cc-pVDZ	H, He, Li, Be, C, N, O, F; CCSD(T) on H₂O in 5 s (GPU)
aug-cc-pVDZ	H, He, Li, Be, C, N, O, F (diffuse functions wired)
Spherical-d	sphd shell (Tier 1 bundle)
f / g / h orbitals	Cartesian integrals + transform
Schwarz integral screening	8-fold canonical loop

Many-body simulation · statevector · MPS · DMRG · kernel fusion

level	notes
L1 statevector	f32 vec2 amplitudes, N/2 threads/gate
L1 controlled-U	N/4 threads, only control=1 touched
L2 MPS	canonical form, Jacobi complex SVD
L2 TEBD	_canonicalizeBond(q) invariant before two-site
DMRG	Lanczos + MPO, ITensor cross-checked N=8
L3 fusion Tier B/C	4.18× headline (Tier C, 8×8)
L3 fusion Tier D	documented honest negative (plateau)
Phase 6 GPU MPS	χ ≤ 64

---

🔬   For researchers

📖 How to cite See CITATION.cff. For papers: Günaydın, A.B. (2026). webgpu-q v0.6.0. https://github.com/abgnydn/webgpu-q A Zenodo DOI will be minted on the next versioned release.	⚠️ Limitations Honest, single-page list of what we cannot do, what is untested, and what is known broken — system size ceilings, browser/vendor matrix, SCF failure modes, missing output formats, precision disclosures.	📊 Benchmarks queue Standardized sets we've run vs. queued: GMTKN55, Thiel/QUEST, W4-11, S66, HEAT-345, SIE4x4, wall-clock vs PySCF / gpu4pyscf, cross-vendor parity.
🛠️ Contributing · 🔁 Migration · 📐 Research standards 15 canonical principles (RESEARCH_STANDARDS.md, mirrored in sibling webgpu-dna). Validation discipline (5w + 20t, fidelity pass bar, honest negatives committed). Migration policy: hand-write only the WebGPU/WGSL/browser layer; port chemistry methods (HF, CCSD, EOM-CCSD, DFT functionals…) from PySCF / libxc with attribution. See LICENSE-PYSCF for the upstream license.	📐 Modern standards audit Every claim mapped to current literature — GMTKN55 functional rankings, EOM-CCSD literature accuracy bars, chemical accuracy bar (1 kcal/mol = 1.594 mHa), AFQMC beyond-CCSD(T), WebGPU subgroups status.	🤝 Code of conduct Contributor Covenant 2.1. Report concerns to hi@barisgunaydin.com.

---

🌐   Companion projects

kernelfusion.dev — umbrella theory site, two preprints gpubench.dev — WebGPU bench harness, 592+ devices webgpudna.com — Geant4-DNA radiobiology port (sibling repo)

zerotvm.com — Phi-3-mini in the browser, no compiler neuropulse.live — live 3.8B-param transformer visualization barisgunaydin.com — author site, project hub

---

📜   Key numbers — single source of truth

Click to expand · edit here when stages move forward

Anywhere a number appears above, it traces back to this table. Update the entry below, then rebuild the SVG hero (public/readme-hero.svg) if a top-line number changed.

symbol	value	context
TESTS	553	vitest unit + integration, all green
CHEMISTRY_TESTS	437	chemistry subset (1 skipped: opt-in cc-pVDZ CCSD(T))
E2E_SPECS	14	Playwright headless WebGPU (CCSD(T), EOM, UV-vis, wallclock-vs-PySCF, levels 1/2/3/6, smoke tests)
CCSD_T_SPEEDUP	39.3×	H₂O · cc-pVDZ · M2 Pro · vs our own CPU
CCSD_T_GPU_TIME	5.05 s	H₂O · cc-pVDZ · GPU
CCSD_T_CPU_TIME	198.6 s	H₂O · cc-pVDZ · CPU
CCSD_T_GPU_DELTA	2.4×10⁻¹⁰ Ha	H₂O · cc-pVDZ · |GPU − CPU|
WIN_HF_H2_STO3G	105×	E34 vs PySCF 2.13.0 · no-startup advantage
WIN_CCSD_LIH_STO3G	40×	E34 vs PySCF 2.13.0 · small-system advantage
LOSS_CCSD_H2O_CCPVDZ	480× slower	E34 vs PySCF 2.13.0 · BLAS gap (NumPy wins)
LOSS_MP2_H2O_CCPVDZ	136× slower	E34 vs PySCF 2.13.0 · BLAS gap
E34_ENERGY_MAX_DELTA	1.0×10⁻⁴ Ha	max |ΔE| vs PySCF over 19 cells · below chemical accuracy
E34_ENERGY_MEAN_DELTA	8.1×10⁻⁶ Ha	mean |ΔE| vs PySCF over 19 cells
EOM_CCSD_PRECISION	10⁻⁵ Ha	H₂ STO-3G · post-32c patch · 2-electron only
EOM_CCSD_LIH_TRIPLET_GAP	7 meV	E35 vs PySCF · 4-electron triplet · effectively exact
EOM_CCSD_LIH_SINGLET_GAP	~0.27 eV	E35 vs PySCF · post-32k sign-fix · within literature EOM-CCSD ↔ FCI bar (~0.1–0.2 eV)
EOM_CCSD_H2O_SINGLET_GAP	~1.9 eV	E35 vs PySCF · 10-e⁻ system · remaining missing T-dressings · PySCF port closes
IP_EOM_H2O	12.03 eV	expt 12.62
EA_EOM_H2O	−16.37 eV	STO-3G (unbound)
DF_HF_PRECISION	7×10⁻¹⁴ Ha	H₂O STO-3G
DF_MP2_PRECISION	0 Ha	H₂O STO-3G at τ=10⁻¹⁰
FUSION_HEADLINE	4.18×	Tier C · 8×8 cascade
STATEVECTOR_FIDELITY	F ≥ 0.999999	f32 GPU vs f64 CPU
MPS_N_MAX	128	TFIM/Heisenberg, χ ≤ 32, browser
MPS_CHI_MAX	64	Phase 6 GPU MPS
H2O_ENTROPY	45.06 cal/(mol·K)	expt 45.1
STAGES_SHIPPED	through v0.6.0	SVG visualization stack (α(ω) / NOON / UV-vis / dispersion / charges / basis coverage / LMO gallery / convergence / 2D field contours) wired into /molecule.html; polarizability + dispersion matrix end-to-end, UKS-DFT full functional ladder, Molden/Cube/QCSchema/XYZ exports, D2 dispersion + counterpoise, Foster-Boys + Pipek-Mezey localization, NOON + multireference verdict
LIVE_URL	webgpu-q.vercel.app	production

---

_{MIT · Built with WebGPU, TypeScript strict, vitest, Playwright
Author @abgnydn · hi@barisgunaydin.com}