Program · computing

Compute, algorithms & the AI-native plant.

The program runs on one versioned engine and a disciplined compute stack — from the in-browser WebAssembly runtime on these pages to the HPC codes (GENE, CGYRO, OpenMC) queued in the team Lab.

Compute · Breeder · Hyperion

Compute · HPC & Quantum

The compute backbone: what runs now (the in-browser solver, scans, Monte-Carlo) and the declared ports to HPC and quantum resources that feed the heavy analysis (FEA, CFD, 3-D neutronics, VQE) when run backend. Grounded in the AI+Quantum simulation deposit.

⚠ Ports + provenance, not live cluster jobs here. Heavy compute runs on the record hardware (56-core / 768 GiB / CPU-only) and in the Lovable backend. This surface declares the interfaces and shows what has been reproduced.

Compute registry
CapabilityEngineWhere it runsState
Breeder gain / powerdt_evaluator.py (NumPy)in-browser (Pyodide)COMPUTES NOW
Design-space scan25,201-config dt_scanprecomputedCOMPUTES NOW
UQ Monte-Carlo2,000-sample MC engineprecomputed · HPC for scaleCOMPUTES NOW
Burner solvermirror.py (scipy/pandas)backend / HPCPORT DECLARED
Structural FEAANSYS / COMSOLHPCPORT DECLARED
Thermal CFD / edgeconjugate-heat · SOLPSHPCPORT DECLARED
3-D neutronicsOpenMC / MCNP + FISPACT-IIHPCPORT DECLARED
Quantum chemistryVQE (qiskit / pyscf)quantum sim / HWPORT DECLARED
AI / ML surrogates (advisory — bounded by the deterministic safety nets)
ModelPurposeTrackRole / result
PINNFast equilibrium / transport surrogatetrack2advisory
GNNTwin forward modeltrack3advisory
Disruption predictorProximity-to-limit, actionabletrack4advisory
RL controlClosed-loop actuationtrack5clamp: 0 escapes / 150k
GP / BO calibrationModel-to-data calibrationtrack6spine 1691× / 36.8× / 7475×
Quantum sensingSensing → disruptiontrack8badvisory
Quantum
VQE proof-of-concept
statevector run
Tier-1 byte-reproducible
deposit track8
Quantum resource estimate
fault-tolerant cost model
no crossover this decade
deposit track7
Sensitivity / advantage
resource sensitivity
honest — classical wins now
deposit track7
Compute provenance & reproducibility
AI+Quantum depositDOI 10.5281/zenodo.21842371 · reserved, deposit pending publication
Integrity182 / 182 artifacts verified (md5)
Tier-1 reproducibilitybyte-identical — MC engine · quantum resources · VQE · calibration
Tier-2 reproducibilitywithin 5-seed tolerance — PINN / GNN / disruption / RL
Hardware of record56-core · 768 GiB · CPU-only

Acceptance criteria that must reproduce: RL clamp 0 escapes / 150k steps; calibration spine 1691× / 36.8× / 7475×; quantum no crossover this decade; CBF certificate 3/4 faces + calibration-through-proof 10/10.

Live solvers run in-browser; FEA/CFD/MCNP/quantum are declared ports resolved backend/HPC. Deposit DOI is reserved — not yet a live link. PORTS + PROVENANCE
🔒 Compute · HPC & Quantum is Kronos-only. The overview above is public — the full data, tables, live calculations and engineering detail on this page require a Kronos team sign-in.
Open methods

Algorithm library

Every calculation in this model, documented in the open — what it does, the method, its inputs and outputs, where it runs, and how confident we are. This is the logic behind the numbers; the exact code lives in the open deposits. Pairs with Governing Physics ↗ and Methodology ↗.

For DOE / national labsEach entry names its numerical method and basis (validated vs. reduced-order vs. design-assumption), so a reviewer can judge fidelity per calculation and map it onto a lab code.
21 of 21 methods
0-D power-balance evaluator (breeder)PLASMA & SYSTEMSVALIDATEDIn-browser (Pyodide)
Solves the coupled confinement + fusion-reactivity + power balance for the breeder design point.
METHOD · Zero-dimensional systems model (dt_evaluator.py + kronos_clean.py, pure NumPy). Fusion power from ⟨σv⟩ reactivity; heating/loss balance sets τ_E and Q.
R₀, a, κ, B₀, Ti₀, nₑ, H₉₈, TBR  →  Q, P_fus, P_n, I_p, β_N, τ_E, T (kg/yr)
10.5281/zenodo.21746157
Confinement scaling (IPB98(y,2) · H₉₈)PLASMA & SYSTEMSSOURCEDIn-browser (live)
Sets energy confinement time from the standard tokamak scaling law, scaled by the H₉₈ factor.
METHOD · τ_E = H₉₈ · τ_IPB98(I_p, B, n, P, R, a, κ, A). H₉₈ is the design's largest single uncertainty.
I_p, B, n, P_heat, geometry  →  τ_E
10.5281/zenodo.21746157
Bootstrap current fractionPLASMA & SYSTEMSREDUCED-ORDERPrecomputed (frozen)
Splits plasma current into self-driven (bootstrap) and externally driven parts.
METHOD · I_p = I_bootstrap + I_driven (frozen: 1.50 + 8.36 MA, f_bs ≈ 0.152) from the pressure profile.
pressure profile, β  →  f_bs, I_driven
10.5281/zenodo.21746157
Stability / β gatePLASMA & SYSTEMSSOURCEDIn-browser (live)
Feasibility gate on plasma pressure.
METHOD · β_N < 4.2 (no-wall NSTX limit) combined with damage-life and coil limits as the "closes" criterion.
β_N, B_peak, T (dpa-life)  →  closes / does-not-close
10.5281/zenodo.21746157
Burner closure / operating windowPLASMA & SYSTEMSREDUCED-ORDERPrecomputed (frozen)
Determines whether the D–³He burner produces net power and how much.
METHOD · Power ledger with DEC recovery over the fuel trajectory; Q_E and f_n vs x(³He), Ti, nₑ. Closure crosses ~x0.44.
x(³He), Ti, nₑ, B_throat  →  Q_E, f_n, net MWe
10.5281/zenodo.21746479
Plug-density requirement (Frank Eq. 3.4)PLASMA & SYSTEMSREQUIREMENTPrecomputed (frozen)
The binding condition for burner net power.
METHOD · φ_i = T_e · ln(n_p/n_c); net power requires plug-to-central density ratio n_p/n_c ≈ 16. A requirement, not a solved result.
T_e, target confining potential  →  required n_p/n_c
10.5281/zenodo.21746479
Design-space scanDESIGN EXPLORATIONVALIDATEDPrecomputed (frozen)
Maps performance and feasibility across the whole design space.
METHOD · 25,201-configuration grid evaluated through the real 0-D evaluator; each point tagged by gate closure and fuel.
grid over I_p, B, Ti, n, …  →  Q / P_fus / closes per config
10.5281/zenodo.21746157
Monte-Carlo uncertainty (UQ)DESIGN EXPLORATIONVALIDATEDPrecomputed (frozen)
Propagates input uncertainty to the headline outputs.
METHOD · 2,000-sample Monte-Carlo over H₉₈, Ti₀, f_He4, f_G, TBR through the real evaluator → distributions and P(Q ≥ target).
input distributions  →  output PDFs, feasibility probabilities
10.5281/zenodo.21746157
Sensitivity / tornado rankingDESIGN EXPLORATIONREDUCED-ORDERIn-browser (live)
Ranks which inputs move the answer most.
METHOD · One-at-a-time swing analysis of |ΔQ| across the input ranges; H₉₈ dominates the spread.
input ranges  →  ranked sensitivities
10.5281/zenodo.21746157
Agentic what-if searchDESIGN EXPLORATIONREDUCED-ORDERIn-browser (live)
Finds the smallest change that hits a performance goal while staying feasible.
METHOD · Search over levers (H₉₈, B₀, Ti₀) calling the real evaluator; returns the minimal gate-preserving change to hit e.g. "+20% Q".
goal + allowed levers  →  lever, value, recomputed result, gate
10.5281/zenodo.21746157
REBCO J_c(B,T) scalingMAGNETS & MATERIALSPUBLIC-DATAPrecomputed (frozen)
Critical current density of the superconductor vs field.
METHOD · Power-law B-scaling anchored to public 20 T tape data, with layer-geometry conversion to superconductor J_c (17.59→6.66 MA/cm² over 5→20 T at 4.2 K).
B, T, tape geometry  →  J_c, engineering J_e
10.5281/zenodo.21842514
Winding-pack rule-of-mixtures (λ)MAGNETS & MATERIALSDESIGNPrecomputed (frozen)
Effective structural/electrical properties of the conductor stack.
METHOD · Volume-fraction weighting across REBCO (1.34%), Cu, Hastelloy, MP35N co-wind and insulation.
layer fractions & properties  →  effective allowable, λ_sc
10.5281/zenodo.21842514
Bore-resolved hoop stressMAGNETS & MATERIALSCOMPUTED (refutes field-only)Precomputed (frozen)
Structural stress in the high-field plug coil.
METHOD · σ(bore) from magnetic pressure with a reinforced rule-of-mixtures allowable. 87–350 MPa across the swept bore — under the 1091 MPa allowable, refuting the naïve field-only scaling (~1047 MPa).
B, J_wp, bore, reinforcement  →  σ_bore vs allowable
10.5281/zenodo.21842514
Quench protection (adiabatic MIITs)MAGNETS & MATERIALSDESIGN (benchmarked)Precomputed (frozen)
Peak temperature during a magnet quench.
METHOD · Current-divider adiabatic MIITs (∫J²dt) with contact-resistance retention fraction; self-protecting, benchmarked to Suetomi 2021 (reproduces 330 K).
stored energy, ρ_ct, geometry  →  T_max, V_terminal, τ_dump
10.5281/zenodo.21842514
Fatigue life (Basquin + Goodman)MAGNETS & MATERIALSINFEASIBLEPrecomputed (frozen)
Cyclic survival of the plug winding pack.
METHOD · S-N fatigue with Goodman mean-stress correction on layer-resolved FEA hoop stress. At operating field N_cycles ≪ design cycles → INFEASIBLE (the binding magnet gate).
σ_max, R, S-N constants  →  N_cycles vs design cycles
10.5281/zenodo.21842514
Magnet neutron lifetimeMAGNETS & MATERIALSDESIGN-STUDYPrecomputed (frozen)
Coil life under fusion-neutron fluence vs shielding.
METHOD · life = life₀ · exp(Σ_R · t) as a function of tungsten shield thickness; 15 cm W → ~0.56 FPY (below the 2 FPY target).
shield thickness, Σ_R, fluence  →  lifetime (FPY)
10.5281/zenodo.21842514
Direct energy conversion (ion-only)ENERGY CONVERSIONSOURCEDPrecomputed (frozen)
Converts charged-particle energy to electricity without a steam cycle.
METHOD · Ion-only DEC recovery efficiency (η ≈ 0.70) applied to the escaping charged-particle power — key to the low-neutron burner.
charged-particle spectrum  →  recovered electric power
10.5281/zenodo.21842864
Grounded retrieval (Ask-the-Twin)INTELLIGENCEREDUCED-ORDERIn-browser (live)
Answers questions only from the cited corpus.
METHOD · Lexical token-overlap scoring over machine-tagged chunks; answers above a ≥2 threshold with a citation, otherwise declines. Honest: lexical, not neural embeddings.
question, machine  →  answer + DOI citation (or decline)
10.5281/zenodo.21842371
Anomaly detection (digital shadow)INTELLIGENCEREDUCED-ORDERIn-browser (live)
Flags off-nominal behaviour in the telemetry stream.
METHOD · Rolling mean ± 3.2σ over a windowed baseline (excludes the recent window); step faults trigger an anomaly flag.
telemetry stream  →  σ-score, anomaly flag
10.5281/zenodo.21842371
Two-tier verificationREPRODUCIBILITYVALIDATEDHPC / local (declared)
Proves every result regenerates from the open record.
METHOD · Tier-1 byte-identical (md5) for deterministic outputs + Tier-2 tolerance (relative-diff) for floating-point; 182/182 checks pass.
deposit code + data  →  pass/fail per check
10.5281/zenodo.21842371
Quantum proof-of-concept + resource estimateINTELLIGENCEREDUCED-ORDERHPC port (declared)
Honest assessment of quantum advantage for the workload.
METHOD · Reproducible quantum PoC with a classical-vs-quantum resource estimate. Verdict: no crossover this decade — classical methods win today.
problem instance  →  resource estimate, verdict
10.5281/zenodo.21842371
Method descriptions are public; proprietary tuning and high-fidelity variants stay in the deposits and internal HPC. Basis tags: VALIDATED / SOURCED · REDUCED-ORDER / DESIGN · REQUIREMENT · REFUTED / INFEASIBLE.
KRONOS FUSION ENERGY · Kronos Model Conceptual design and simulation study; no machine has been built