webgpu-q

Real quantum chemistry, in a browser tab. Open a URL and run the actual thing — Hartree–Fock, DFT, MP2, CCSD, CCSD(T), EOM-CCSD — computed on your own machine, in the tab. No install. No backend. No CUDA. Every number cross-checked against PySCF.

   

What this is A browser-native electronic-structure engine. Methods ported from PySCF and brute-force verified to ≤ 1e-10 Ha element-wise. A WebGPU systems demo — a ~110-line WGSL kernel makes CCSD(T)/cc-pVDZ browser-feasible (~14× median over CPU TypeScript; noisy). A distributed-compute substrate — a browser-tab "swarm" that splits a Hartree–Fock build across tabs/machines (N=2 cross-machine HF verified, energy matches single-machine to 5.7e-14). A teaching / reproducibility platform — every calculation is a shareable URL; drag-import XYZ/PDB/MOL/SDF; in-browser "Compare to PySCF" REPL.

What this isn't Not a PySCF replacement. PySCF is 10–100× faster, scales 100× larger, has 25 years of validation. If you do production chemistry, use it. Not for big systems. Benchmarked up to benzene/naphthalene HF and H₂O cc-pVDZ CCSD(T) — nothing at porphyrin / metal-complex / enzyme scale. Not a CUDA competitor. WebGPU has no f64 and is ~2–5× behind tuned CUDA today. Not novel chemistry. Every method was in PySCF first. The contribution is the substrate, not the algorithms.

Who it's for Educators — every intermediate of the HF→MP2→CCSD pipeline is inspectable, in real time, with zero install. WebGPU / systems researchers — a real O(n⁷) kernel benchmarked end-to-end in the browser. Reproducibility — calculations are URLs; results shareable via auto-run links + IndexedDB history.

---

See it

/screening.html · rank a library by HOMO–LUMO gap, live HF per molecule

/molecule.html · full chemistry-paper SI for a small molecule

/experiments/ · the research ladder, live runner + JSON artifacts

---

How fast — honestly, both directions

All vs PySCF 2.13.0 on identical inputs (E34 run); energy agreement ≤ 10⁻⁴ Ha on all 19 comparable cells (well below the 1.594 mHa chemical-accuracy bar).

	system	result
🟢 win	HF · H₂ · STO-3G	105× faster (no Python startup)
🟢 win	CCSD · LiH · STO-3G	40× faster (small-system advantage)
🟢 win	CCSD(T) · H₂O · cc-pVDZ · GPU	~14× median vs our CPU TypeScript (39× best run; std/median ≈ 42%, noisy — LIMITATIONS)
🔴 loss	CCSD · H₂O · cc-pVDZ	480× slower (NumPy/BLAS dominates)
🔴 loss	MP2 · H₂O · cc-pVDZ	136× slower (BLAS gap)

The honest takeaway: we win on no-startup + small systems + the GPU (T) kernel; we lose badly at production basis where BLAS rules. The speedups are vs our own CPU baseline, not vs PySCF wall-clock or GPU4PySCF — that apples-to-apples comparison is open work.

---

Validated against ground truth

layer	cross-checked against	residual
HF (RHF/UHF, DIIS, spherical-d)	PySCF	≤ 0.5 mHa (≤ 0.1 mHa cc-pVDZ sphd)
MP2 · FCI	PySCF · analytic	≤ 0.76 mHa (CH₄ FCI)
CCSD(T)	FCI	≤ 0.25 mHa
EE / IP / EA-EOM-CCSD	explicit H̄ = e⁻ᵀHeᵀ projection	< 1e-10 Ha element-wise (EA: < 5e-13 on multi-electron LiH)
DF-HF · DF-MP2	direct 4-index ERI	~7×10⁻¹⁴ Ha (engineering assertion)
statevector (GPU f32)	f64 CPU reference	F ≥ 0.999999
MPS / DMRG	ITensor · Pfeuty/Bethe	f64 at N=8 · 1/N limits
swarm HF (cross-machine)	single-machine	5.7e-14 Ha

Reading the precision numbers: anything tighter than 1.6 mHa (= 1 kcal/mol) is a software regression assertion (does the GPU/DF port match the CPU/direct path?), not a chemistry result — basis-set incompleteness dwarfs it by 6+ orders. The EOM ports are faithful to PySCF's eom_gccsd / eaccsd_matvec, not re-derived; the brute-force diffs prove the implementation, not method accuracy on real systems.

---

60-second demo

git clone https://github.com/abgnydn/webgpu-q && cd webgpu-q
npm install
npm run dev          # http://localhost:5175  ·  /screening.html  ·  /molecule.html  ·  /experiments/
npm run test         # vitest · CI green        npm run typecheck   # tsc strict

// one call: molecule → full property report (energy, dipole, charges, bond orders,
// NOON, ⟨S²⟩, multireference verdict, α, D2, UV-vis, Molden) — all in a tab
import { molecules, quickReport, uvVisSpectrum } from "./src/chemistry";
const report = quickReport(molecules.h2o, { addD2: true, addStaticAlpha: true });
const uvvis  = uvVisSpectrum(molecules.h2o, { method: "b3lyp5" });

// or piece by piece
import { runRHFSCF, runMP2, runCCSD, runCCSDT_GPU, runEOMCCSD } from "./src/chemistry";
const hf  = runRHFSCF(integrals, nElectrons);
const t   = await runCCSDT_GPU(runCCSD(hf, integrals), hf, integrals, device); // ~14× median (noisy)
const eom = runEOMCCSD(runCCSD(hf, integrals), integrals, hf);

---

What's inside

Ground state HF · UHF · RKS/UKS-DFT (LDA/GGA/hybrid ladder) · MP2 · DF-MP2 · aux-basis f64 density fitting · analytical HF/DFT gradients · BFGS geom-opt Correlation & excited states CCSD · UCCSD · CCSD(T) (CPU + WebGPU) · EE/IP/EA-EOM-CCSD · CIS · TDA · TDDFT (singlet + triplet) · oscillator strengths Properties dipole · α(0)/α(ω)/α(iω) · C₆ · β · Mulliken · Wiberg/Mayer · NOON + multireference verdict · D2 dispersion · IR · Raman · UV-vis · thermochemistry · IPs/EAs I/O Molden · Gaussian Cube · QCSchema · XYZ · drag-import PDB/MOL/SDF Many-body GPU statevector · MPS/DMRG (ITensor-checked) · kernel fusion (4.22×)

Full method catalog (click)

ground state	notes
RHF / UHF SCF	DIIS, frozen-core, spherical-d, f/g/h, level-shift, ⟨S²⟩
RKS / UKS-DFT	LDA · BVWN5 · BLYP · B3VWN5 · B3LYP5; Becke + Lebedev grids
MP2 · DF-MP2	spin-orbital + B-tensor; aux-basis f64 DF, no 4-index ERI
HF/DFT gradients	analytical Pulay 1969, Schwarz screening, BFGS / L-BFGS

correlation / excited	notes
CCSD · UCCSD	Stanton-Bartlett, antisymmetrized spin-orbital, frozen-core
CCSD(T) CPU + GPU	FCI-validated ≤ 0.25 mHa; GPU ~14× median (noisy), f32→f64 reduce
EE / IP / EA-EOM-CCSD	PySCF eom_gccsd / eaccsd_matvec ports; brute-force < 1e-10 Ha
CIS · TDA · TDDFT	Casida, full functional ladder, triplet via spin-pol, Davidson

properties	notes
α(0)/α(ω)/α(iω) · C₆ · β	CPHF + TDHF + TDDFT; Casimir-Polder; finite-field β
Mulliken · Wiberg/Mayer · NOON	charges/spin, bond orders, multireference verdict (T1/D1/⟨S²⟩)
IR · Raman · thermo	mass-weighted Hessian; Placzek; Sackur-Tetrode (H₂O S = 45.06 vs expt 45.1)
IPs / EAs	Koopmans / ΔSCF / EOM (H₂O EOM IP 12.03 eV vs expt 12.62)

basis / many-body	notes
STO-3G · 6-31G* · cc-pVDZ · aug-cc-pVDZ	first + second period; spherical-d; f/g/h
statevector · MPS · DMRG	GPU f32 (F ≥ 0.999999); Jacobi SVD + canonical TEBD; Lanczos + MPO
kernel fusion	4.22× (Tier C, 8×8 cascade); Tier D 3.78× plateau = documented honest negative

---

Research discipline

Harness · experiments/lib/ runner.ts — timedRun with forced GPU sync (read-after-submit) seeds.ts — named deterministic seeds, no Math.random() fidelity.ts — F = |⟨ψ_ref|ψ_test⟩|², not max|Δp| env.ts / stats.ts — adapter/SHA capture · median, p10/p90/p99

Non-negotiables 5 warmup + 20 trials per measurement pass bar F ≥ 1 − 10⁻⁵; std/median > 0.1 → status: "noisy" honest negatives committed as JSON with a diagnosis vitest + Playwright e2e · CI green · TS strict + noUncheckedIndexedAccess

---

For researchers

• 📖 Cite — CITATION.cff. Concept DOI 10.5281/zenodo.20494382 (resolves to latest; each release also gets a version DOI).
• ⚠️ Limitations — what we cannot do, what's untested, what's known broken: size ceilings, vendor matrix, SCF failure modes, precision disclosures.
• 📊 Benchmarks queue — run vs queued: GMTKN55, QUEST, W4-11, S66, wall-clock vs PySCF/gpu4pyscf.
• 🛠️ Contributing · 🔁 Migration · 📐 Research standards — port chemistry from PySCF/libxc with attribution (LICENSE-PYSCF); hand-write only the WebGPU/WGSL/browser layer.
• 🤝 Code of conduct — Contributor Covenant 2.1.

Key numbers — single source of truth (click)

symbol	value	context
CCSD_T_SPEEDUP_MEDIAN	13.8×	H₂O cc-pVDZ · 5w+20t · vs our CPU TS · NOISY (std/median 42%)
CCSD_T_SPEEDUP_P10 / P90	28.4× / 10.1×	best / worst across 20 trials (39× once-headlined was a single lucky run, retired)
CCSD_T_GPU / CPU_TIME	8.4 s / 116.4 s	median GPU per-call vs single CPU run, M2 Pro
WIN_HF_H2_STO3G	105×	E34 vs PySCF 2.13.0 (no-startup)
WIN_CCSD_LIH_STO3G	40×	E34 vs PySCF 2.13.0 (small system)
LOSS_CCSD_H2O_CCPVDZ	480× slower	E34 vs PySCF 2.13.0 (BLAS gap)
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 (sub-chemical)
EOM_BRUTEFORCE_DIFF_LIH	< 1e-10 Ha	EE/IP σ vs explicit H̄, element-wise; PySCF-ported, no patches
EA_EOM_BRUTEFORCE_DIFF_LIH	< 5e-13 Ha	EA-EOM σ vs exact H̄ (multi-electron LiH); eaccsd_matvec port
DF_HF_PRECISION	7×10⁻¹⁴ Ha	DF-HF vs direct (engineering assertion)
FUSION_HEADLINE	4.22×	Tier C · 8×8 cascade
STATEVECTOR_FIDELITY	F ≥ 0.999999	f32 GPU vs f64 CPU
MPS_N_MAX / CHI_MAX	128 / 64	TFIM/Heisenberg in browser · Phase 6 GPU MPS
H2O_ENTROPY	45.06 cal/(mol·K)	vs expt 45.1
STAGES_SHIPPED	v0.12.0	full site/landing redesign (live wavefunction hero, cohesive spectral system across all pages); both manuscripts corrected + every citation verified

---

Companion projects

kernelfusion.dev (umbrella · kernel fusion) · gpubench.dev (WebGPU bench, 592+ devices) · webgpudna.com (Geant4-DNA in the browser) · zerotvm.com (Phi-3 in the browser) · neuropulse.live (live transformer activations) · barisgunaydin.com (author)

_{MIT · WebGPU · TypeScript strict · vitest · Playwright · by @abgnydn · hi@barisgunaydin.com}