Vantasner Hydride–Piezo
[DOCS]
01 PITCH 02 PROGRAM 03 DETAIL 04 SOURCE
Bare-metal servers in a dark rack aisle: honeycomb front grilles, cable bundles and rows of status lights glowing amber and purple
[Illustrative image]

[03 / ENGINE]

Up to 256 agents.
Around the clock.

Run the lab, a 24/7 agent-run simulation queue and the team in parallel. Let simulation narrow the parameters before the bench. That is how ten years of research fit into twelve months.

  • [FLEET] 16×16 servers × agent slots = 256 at the M9 to M12 peak; 160 on average
  • [HOURS / MONTH] 720 the queue is built to run all of them
  • [RE-PLAN] Mon 06:00 every week, from the results database
  • [JOBS] 120k 13 job classes, 12 campaigns est.

Target start: Monday 2 November 2026 (M1)

[03.1 · THE PROMISE]

Ten years of research.
Twelve months.

That is what this engine delivers. Three loops in parallel, up to 256 agents around the clock, 160 slot-years of work in one calendar year, about 120,000 simulation jobs, and a fresh plan every Monday at 06:00.

256 agent slots at the peak, months 9 to 1216 servers × 16 slots; 160 on average
160 slot-years of round-the-clock work in one calendar year1,920 slot-months ÷ 12 · ~3.84M usable core-hours est.
120k simulation jobs in 13 classes120,020 on the Scale B basis est.
52 Monday re-plans in the year06:00 every week: a fresh parameter brief for the lab
[HOW] Three loops in parallel, not one bench in sequence
Queue §1 fixed givens Queue §2 assumptions Plan §4 compute schedule Queue §6 campaign plan

[03.2 · WHY IT COMPRESSES TIME]

Run it in parallel,
not in sequence.

Conventional research runs one hypothesis on one bench, in business hours, one step after the other. This program runs three loops at once, every month, and lets simulation narrow the parameters before the bench spends a gram of powder.

  1. [01]

    Three loops, one calendar.

    The villa lab builds and measures. The queue simulates. The team ramps. All three run in every month of the plan, side by side, instead of waiting for each other.

    Plan §3
  2. [02]

    All 720 hours of the month.

    A month is 720 hours, and the agents do not keep office hours. The queue plans for an 85 percent duty factor est. after restarts and maintenance, and re-plans itself every Monday at 06:00.

    Queue §2Queue §8
  3. [03]

    Simulate first, then build.

    By the time the first powder enters the Gen 0 cell in month 3, the plan has the queue through 14,780 jobs est.. By first deuterium in month 6: 48,830. The bench starts from a shortlist, not from the whole parameter space.

    Queue §6Plan §5

[FIG 01] The same twelve months, two ways to spend them

M1 = November 2026

[SEQUENTIAL] Conventional one hypothesis, one bench, business hours
ReadRead HypothesisHyp. BuildBuild RunRun AnalyseAnl. Next hypothesis…Next…
[24/7] Simulation queue a parameter brief every Monday: 52 in the plan year
A 4 servers · 64 slotsA4 B 8 servers · 128 slotsB8 C 16 servers · 256 slotsC16
[LAB] Villa lab 12 monthly decisions, 4 prototype generations
Orders + fit-outOrders CommissioningCommiss. First DFirst D Steady stateSteady Gen 0 Gen 1 Gen 2 Gen 3
[TEAM] Team ramp 10 full-time roles, by start month est.
5 7 9 10 10 10 10 10 10 10 10 10
Lane 1: the conventional pattern, one step at a time. Lanes 2 to 4: the program plan, with compute scale from the compute schedule, commissioning phases, prototype generations Gen 0 to Gen 3, one decision closed every month (◆) and headcount by start month. Run in parallel, this is how ten years of research fit into twelve months. Plan §1 generationsPlan §4 computePlan §5 commissioningPlan §7 team

[03.3 · THE LOOP]

One agent per slot.
Four jobs. One loop.

Each of the 16 slots on a server runs one autonomous AI agent with scoped access to the queue, the results database and its own filesystem. The four jobs of the research loop are staffed by rotation: one persona per slot per week. Humans steer by exception.

[FIG 02] The agent loop inside the fleet

01 Triage
02 Generate
03 Supervise
04 Ingest
[HEAD SERVER] Fleet coordinator re-plans weekly from the results database
  1. [01]

    Literature triage agent

    Reads the prior-art corpus and monitored new literature; extracts parameter ranges (loading ratios, drive frequencies, particle sizes, screening potentials); files job requests into the intake queue with provenance citations.

  2. [02]

    Job generation agent

    Converts job requests into validated simulation inputs through the ASE (Atomic Simulation Environment) job factories, Elmer solver-input templates, and the Python kernel library; runs one-minute smoke checks locally; submits to the class lane.

  3. [03]

    Run supervision agent

    Watches the scheduler, restarts from checkpoints on failure, enforces wall-clock budgets, and escalates repeated failures to the fleet coordinator.

  4. [04]

    Result ingestion agent

    Parses outputs, runs acceptance checks, writes results rows and artifact checksums, and drafts the weekly parameter brief for the lab campaign.

  5. [BY EXCEPTION]

    Humans on call

    Two automatic retries from checkpoint, then the supervision agent escalates to the coordinator, and the coordinator to a human on-call, within one hour.

  1. [IN] Literature the corpus the agents work from: 108 references and 84 published records, plus new papers as they appear
  2. [LOOP] Up to 256 agents triage, generate, supervise, ingest, around the clock
  3. [STORE] Results database PostgreSQL: jobs, results, artifacts, briefs; only checked results are published
  4. [MON 06:00] Parameter brief which settings the lab runs next, with a diff against last week
  5. [OUT] Lab decision a named owner closes a named parameter; from Gen 2 bring-up (month 9), lab data flows back in nightly
DETAIL Follow one job through the queue: seven steps
  1. 01

    Request

    a triage record with target observable, tool class, and citation provenance.

  2. 02

    Specification

    a JSON record with class, code, resource request, convergence tolerances, and a unique idempotency hash built from the input-factory commit, the configuration hash, and the container digest. Duplicate hashes are dropped, which makes the whole fleet safely restartable.

  3. 03

    Queue

    the job enters its class lane (K0 to K12, section 4).

  4. 04

    Run

    containerized execution under Dask; checkpoints at fixed wall-clock intervals for long MD and DFT jobs.

  5. 05

    Acceptance checks

    convergence flags, energy drift bounds, conservation checks, and physics anchors (lattice constants, bulk moduli, known rate integrals) evaluated automatically at ingest.

  6. 06

    Ingestion

    results rows plus artifact checksums into PostgreSQL; the weekly parameter brief regenerates.

  7. 07

    Failure path

    two automatic retries with checkpoint resume, then persona escalation (supervision agent to coordinator to human on-call).

Steps 01 to 06 are the happy path. Step 07 branches off steps 04 and 05: retry from checkpoint twice, then escalate. The idempotency hash in step 02 is why any server can be restarted at any time without duplicating work.

Queue §3.4 job lifecycle
DETAIL Inside one server: slot anatomy, software stack, fleet topology

Fleet topology

  • Server 1 is the head: fleet coordinator agent, PostgreSQL primary, campaign planner, observability aggregation.
  • Servers 2 to N are workers: PostgreSQL replicas, local Valkey for the server-level lanes, full toolchain containers.
  • The coordinator re-plans weekly from the state of the results database, not from human slide decks.

Slot anatomy

Each of the 16 slots runs one autonomous agent: a Python runtime with scoped tool access to the queue API, the shared results database, the input-generation repositories, and a per-slot filesystem. The four functions of the loop are staffed by persona rotation, one persona per slot per week:

  1. Literature triage agent. Reads the prior-art corpus and monitored new literature; extracts parameter ranges (loading ratios, drive frequencies, particle sizes, screening potentials); files job requests into the intake queue with provenance citations.
  2. Job generation agent. Converts job requests into validated simulation inputs through the ASE (Atomic Simulation Environment) job factories, Elmer solver-input templates, and the Python kernel library; runs one-minute smoke checks locally; submits to the class lane.
  3. Run supervision agent. Watches the scheduler, restarts from checkpoints on failure, enforces wall-clock budgets, and escalates repeated failures to the fleet coordinator.
  4. Result ingestion agent. Parses outputs, runs acceptance checks, writes results rows and artifact checksums, and drafts the weekly parameter brief for the lab campaign.

Software stack

Queue broker
ValkeyPriority lanes per job class, agent-to-agent busBSD-3-Clause
Scheduler
Dask distributedOne distributed scheduler per server over the 48-core pool; fleet-aware work stealing across serversBSD-3-Clause
Results database
PostgreSQLShared, replicated job and results store with JavaScript Object Notation (JSON) columnsPostgreSQL License
Input generation
ASE job factories, Elmer templates, Python kernelsReproducible inputs keyed by commit plus configuration hashLGPL / GPL / BSD
Simulation codes
Quantum ESPRESSO, GPAW, CP2K, LAMMPS, PRISMS-PF, MOOSE, FEniCSx, Elmer, LIGGGHTS-PUBLIC, phonopy, MACESee TOOLCHAIN.mdGPL / LGPL / BSD / Apache-2.0
Execution isolation
Pinned containers executed with the Linux container runtime available on the hostIdentical job environment on every servern/a (distribution method)
Observability
Prometheus and GrafanaFleet health, per-class throughput, cost metersApache-2.0 / AGPL-3.0
Configuration
Ansible playbooksSingle-command server rebuildGPL-3.0
Queue §3.1Queue §3.2Queue §3.3
DETAIL The assumptions behind every throughput number
Marbell server shape
64 physical cores, 512 GB random access memory, 8 TB non-volatile memory express (NVMe) storage, 10 GbE network (estimate)The CHF 1500 covers amortized hardware, colocation, power, and bandwidth (estimate)
Agent reasoning model
Hosted application programming interface (API) call, one large language model (LLM) per slotToken costs sit outside the CHF 1500 server rate and are budgeted separately in the program plan (assumption)
Agent runtime footprint
1 reserved core and 16 to 32 GB per slot (estimate)Agents are input/output bound: they mostly wait on LLM responses, so their cores free up for compute
Server compute pool
48 of 64 cores shared by the 16 slots under weighted fair scheduling (estimate)Per-slot baseline entitlement 3.0 core-equivalents; bursty effective use about 3.2 (estimate)
Month
720 hours; duty factor 85 percent after restarts, queue overhead, maintenance (estimate)About 2000 usable compute core-hours per slot-month (estimate)
Accelerators
No graphics processing units (GPUs) assumed; machine-learning potential training sized for central processing units (estimate)A GPU option would raise class K12 throughput (estimate)
Queue §2 assumptions (all estimates)
DETAIL The results database: four tables, one source of truth

Sketch of the shared store:

  • jobs: id, class, specification (JSON), provenance (JSON), state, timestamps.
  • results: job id, observable, value, unit, uncertainty, context (JSON: composition, strain, drive, carrier).
  • artifacts: path, checksum, size.
  • briefs: week, parameter brief document, diff against previous week.

The database is the single source of truth for the monthly decision link to the lab campaign (section 7).

Queue §3.5 results database

[03.4 · TWELVE CAMPAIGNS]

Twelve campaigns.
Twelve months.

Every month the queue runs one campaign aimed at the lab's next decisions. The fleet grows with the program: 4 servers while the villa is fitted out, 8 through the screening and carrier months, all 16 from Gen 2 bring-up. Press play, or pick a month.

  • Triage
  • Generate
  • Supervise
  • Ingest
  • Not yet in fleet
[FIG 03] The fleet through the year: server and slot counts from the plan's compute schedule; persona colours rotate every Monday; each flash is a finished job.
or pick a month below

[M01 · NOVEMBER 2026]

Scale A · 4 servers · 64 slots · CHF 6,000 / month est.

Queue commissioning; DFT foundations for PdD0.7, TiD2, ZrD2, NiHx; screening kernels v1; first horn model

6,120 jobs est.

[NARROWS IN THE LAB]

  • Hydride shortlist for stage 1
  • DFT anchor set
  • horn first resonance
Month 1 in the 12-month plan

[M02 · DECEMBER 2026]

Scale A · 4 servers · 64 slots · CHF 6,000 / month est.

Strain scans and phonons around the hydride transitions; potential benchmark phase; screening fit v1

8,660 jobs est.

[NARROWS IN THE LAB]

  • Temperature window
  • thermal expansion mismatch
  • kernel benchmark against measured metallic screening potentials
Month 2 in the 12-month plan

[M03 · JANUARY 2027]

Scale B · 8 servers · 128 slots · CHF 12,000 / month est.

Transient MD ensembles; DEM force chains v1; FEniCSx piezo grain kernel v1

1,350 jobs est.

[NARROWS IN THE LAB]

  • Drive amplitude ceiling
  • particle size window
  • contact-stress pulse distributions
Month 3 in the 12-month plan

[M04 · FEBRUARY 2027]

Scale B · 8 servers · 128 slots · CHF 12,000 / month est.

Concept matrix scans: 4 hydrides x 4 piezoelectrics x carrier shortlist

20,900 jobs est.

[NARROWS IN THE LAB]

  • Piezoelectric composition
  • volume fraction
  • carrier shortlist (dry bed versus deuterated oil versus electrolyte)
Month 4 in the 12-month plan

[M05 · MARCH 2027]

Scale B · 8 servers · 128 slots · CHF 12,000 / month est.

Carrier decision support: field-survival arithmetic, Debye screening, loaded acoustic models

11,800 jobs est.

[NARROWS IN THE LAB]

  • Carrier choice
  • gap-field budget feasibility
  • coating and fit tolerances
Month 5 in the 12-month plan

[M06 · APRIL 2027]

Scale B · 8 servers · 128 slots · CHF 12,000 / month est.

MLIP dataset generation and first training; 2D phase-field fronts

480 jobs est.

[NARROWS IN THE LAB]

  • Interface mobilities
  • ReaxFF replacement for shock cells
Month 6 in the 12-month plan

[M07 · MAY 2027]

Scale B · 8 servers · 128 slots · CHF 12,000 / month est.

3D phase field under oscillatory strain; full-cell coupled FEM

500 jobs est.

[NARROWS IN THE LAB]

  • Loading ramp protocol
  • duty cycle
  • resonance tracking versus temperature
Month 7 in the 12-month plan

[M08 · JUNE 2027]

Scale B · 8 servers · 128 slots · CHF 12,000 / month est.

Drive frequency and duty-cycle sweeps

780 jobs est.

[NARROWS IN THE LAB]

  • Drive frequency final choice
  • power ceiling
  • duty cycle
Month 8 in the 12-month plan

[M09 · JULY 2027]

Scale C · 16 servers · 256 slots · CHF 24,000 / month est.

Lab bring-up support: as-built cell models from measured geometry

7,600 jobs est.

[NARROWS IN THE LAB]

  • As-built resonance prediction
  • expected contact charge
  • thermal balance
Month 9 in the 12-month plan

[M10 · AUGUST 2027]

Scale C · 16 servers · 256 slots · CHF 24,000 / month est.

Signal-class predictions: spectra kernels over the surviving matrix

30,070 jobs est.

[NARROWS IN THE LAB]

  • Detector exposure plan
  • proton line versus continuum expectations
  • minimum detectable yield targets
Month 10 in the 12-month plan

[M11 · SEPTEMBER 2027]

Scale C · 16 servers · 256 slots · CHF 24,000 / month est.

Design lock: sensitivity sweeps over the 12 design parameters; margin list

20,560 jobs est.

[NARROWS IN THE LAB]

  • Final parameter brief version 1.0
  • margin list for the lab
Month 11 in the 12-month plan

[M12 · OCTOBER 2027]

Scale C · 16 servers · 256 slots · CHF 24,000 / month est.

Campaign support: assimilate first device data, correct contact-stress and field models

11,200 jobs est.

[NARROWS IN THE LAB]

  • Live parameter corrections feeding the next build cycle
Month 12 in the 12-month plan
Slots in the fleet (plan) Heavier jobs (K1 to K12) Kernel scan points (K0) Job counts: queue document, Scale B basis est.

Kernel scan points (K0, about half a centime each est.) dominate the count. Months 6 to 8 run few jobs, but heavy ones: ab initio trajectories, machine-learned potential training, 3D phase field. The queue document writes the month plan on the Scale B basis (120,020 jobs in all); the program plan applies its Scale A rules in months 1 to 2 and Scale C rules in months 9 to 12. Queue §6 campaign planPlan §4 compute scale

[03.5 · DECISION LINK]

Simulation decides.
The lab confirms.

No campaign runs for its own sake. Each one exists to narrow a named lab parameter, by a named month, for a named owner. The coordinator agent regenerates this map every week from the results database; this is the month-1 plan.

  1. DFT relaxations and strain scans K3
    Hydride choice per stage: deuteron density versus price versus mobility (PdD0.7 versus TiD2 versus ZrD2 versus NiH)
    Months 1 to 2
    Lab lead
  2. Phonons and quasi-harmonic expansion K4
    Temperature window; thermal expansion mismatch tolerances for particle coating and cell fit
    Month 2
    Lab lead
  3. Screening and tunneling kernels K0
    Expected signal class; minimum detectable yield; detector exposure plan
    Months 2 and 10
    Diagnostics lead
  4. Transient MD K1K2K6
    Maximum drive amplitude; particle size window (1 to 10 micrometer baseline)
    Month 3
    Lab lead
  5. DEM force chains K11
    Piezoelectric volume fraction (about 10 percent baseline); packing fraction; hydride-to-piezo size ratio
    Month 4
    Formulation lead
  6. FEniCSx piezo kernels K10
    Carrier fluid choice (field survival); gap budget against the 10⁸ to 10⁹ V/m design target
    Month 5
    Device lead
  7. Elmer horn and cell acoustics K9
    Horn geometry; drive frequency (20 versus 40 kHz class); power ceiling below cavitation onset
    Months 1, 8, 9
    Device lead
  8. Phase field K7K8
    Loading ramp protocol; duty cycle (drive-on fraction)
    Months 7 to 8
    Lab lead
  9. Assimilation lane K10K0
    Live corrections of contact-stress and field models from first device data
    Month 12 onward
    Device lead

Month numbers run M1 (November 2026) to M12 (October 2027). The lab side of each decision is scheduled month by month in the plan.

Queue §7 decision linkPlan §3 month by monthThe 12-month plan

[03.6 · TOOLCHAIN]

From electrons
to the horn.

Open-source codes, each chosen for one question about this device. Each hands its numbers up to the next scale. Lattice strain, hydride phase transitions and sharp mechanical transients are explicit inputs everywhere, never afterthoughts.

  1. Q1

    Which hydride, which piezoelectric, which carrier?

    Tools: DFT (Quantum ESPRESSO, GPAW, CP2K), screening kernels, DEM

    Feeds: Stage-1 material matrix

  2. Q2

    What contact stress, charge, and gap field does the force-chain network actually deliver?

    Tools: LAMMPS, LIGGGHTS, Elmer, FEniCSx kernels

    Feeds: Drive design, carrier choice, duty cycle

  3. Q3

    What deuteron energies and yields does a given configuration imply?

    Tools: Screening and tunneling kernels fed by DFT densities and FEM field maps

    Feeds: Diagnostic planning, go/no-decision briefs for the lab

[01]ELECTRONS + ATOMSÅ · fs Density functional theory Quantum calculations of how the atoms of a hydride bond, stretch and hold deuterium. Every parameter in the chain starts here. K3K4K5K6

[CODES]Quantum ESPRESSO · GPAW · CP2K · phonopy · what is DFT?

[CONSTRAINS]
Hydride choice per stage; lattice constants and stiffness under strain; thermal-expansion mismatch between grains; electron densities at deuterium sites for the screening kernels; the field across a piezo-to-hydride contact gap.
[HANDS UP]
Formation energies to phase field, reference data to the MLIP, site densities to the kernels

Job classes: K3K4K5K6 Toolchain §2 in full

[02]ATOMSnm · ps Molecular dynamics Up to millions of atoms moving through a single contact-stress pulse, the device's native event. K1K2

[CODES]LAMMPS (EAM, ReaxFF) · ASE as workflow glue · what is MD?

[CONSTRAINS]
Contact-pulse amplitude and duration; the maximum safe drive amplitude; particle-size effects on strain transfer; deuterium diffusivity under strain.
[HANDS UP]
Per-pulse constitutive data to the FEM codes, which carry the 20 kHz drive

Job classes: K1K2 Toolchain §3 in full

[03]ATOMS, FASTERnm · ns Machine-learned potentials A model trained on DFT results so that molecular dynamics runs near DFT accuracy at force-field cost. K12

[CODES]MACE, trained on CP2K and Quantum ESPRESSO trajectories · what is MLIP?

[CONSTRAINS]
Closes the accuracy gap between reactive force fields and ab initio dynamics for shock-like contact pulses.
[HANDS UP]
Better potentials back into MD

Job classes: K12 Toolchain §7 in full

[04]GRAINSµm Phase field How the hydride phase boundary moves through a grain as it loads, with the drive on or off. K7K8

[CODES]PRISMS-PF · MOOSE · FEniCSx kernel if needed · what is Phase field?

[CONSTRAINS]
The loading ramp protocol; the duty cycle (drive-on fraction); phase-front timescales against the drive period.
[HANDS UP]
Loading and duty-cycle settings to the lab

Job classes: K7K8 Toolchain §4 in full

[05]POWDERµm to mm Discrete element method Every grain of the powder as it packs, jams and forms force chains. K11

[CODES]LIGGGHTS-PUBLIC · what is DEM?

[CONSTRAINS]
Force-chain statistics; contact-stress histograms against packing fraction and piezo volume fraction (about 10 percent baseline); hydride-to-piezo size ratio.
[HANDS UP]
Contact-stress distributions to FEM and the kernels

Job classes: K11 Toolchain §7 in full

[06]HORN + CELLmm to cm · µs Finite element method Sound, stress and electric fields in the ultrasonic horn, the cell and a single piezo grain between two hydride grains. K9K10

[CODES]Elmer · FEniCSx / DOLFINx · what is FEM?

[CONSTRAINS]
Drive frequency (20 versus 40 kHz class); horn and cell geometry; the power ceiling below cavitation onset; how much gap field survives in each carrier.
[HANDS UP]
Field maps and contact charge to the kernels

Job classes: K9K10 Toolchain §6 in full

[07]SIGNALeV to keV Screening and tunneling kernels Small, fast calculations of how often two deuterons fuse at a given energy and electron screening, run millions of times. K0

[CODES]Custom Python, benchmarked against published compilations

[CONSTRAINS]
Expected deuteron energy spectra; the minimum detectable yield; the detector exposure plan before the first run.
[HANDS UP]
Signal-class predictions to the diagnostics lead

Job classes: K0 Toolchain §5 in full

DETAIL The validation ladder: nothing runs production until it clears its rung

Every tool is checked against physical anchors before it touches a production job: known lattice constants, bulk moduli, tabulated shock states, published rate compilations and, once hardware exists, the physical horn on a network analyser.

Toolchain §9 validation ladder
  1. [RUNG 01]

    DFT anchors

    Reproduce face-centered-cubic palladium lattice constant near 3.89 angstrom, the beta-PdD lattice expansion on the order of 3 to 4 percent on full loading (reported experimental range), TiH2 and ZrH2 fluorite stability, and bulk moduli against literature values.

  2. [RUNG 02]

    MD anchors

    Elastic constants and H solution energies versus the DFT results of rung 1; Hugoniot states against tabulated data.

  3. [RUNG 03]

    Phase-field anchors

    Dilute-limit chemical potential against DFT; interface width and mobility against MD.

  4. [RUNG 04]

    Kernel anchors

    Bare Gamow and S-factor arithmetic against standard compilations; enhancement curves against the Huke et al. 2008 consolidation.

  5. [RUNG 05]

    FEM anchors

    Horn impedance spectrum against network-analyzer measurements of the physical horn once hardware exists (lab link, months 1 to 3 of the build plan).

SOURCE The toolchain document, section by section

[03.7 · JOB CLASSES]

Thirteen job classes.
Priced per job.

Cost per job is core-hours ÷ 2,000 usable core-hours per slot-month × CHF 93.75. A kernel scan point costs half a centime; a thousand-atom quantum trajectory costs CHF 216 est..

DETAIL All 13 job classes: tool, typical job, cores, wall-clock, core-hours, CHF per job, throughput
K0 Python kernelsGamow/screening scan point, S-factor integral, spectrum kernel 1 core 5 to 30 min 0.1 core-h CHF 0.005 3,000* 48,000

Throughput ceiling est.3,000* jobs per slot-month · 48,000 per server-month

[RUNS IN THE CAMPAIGN PLAN · SCALE B BASIS]

  1. M15,000
  2. M28,000
  3. M420,000
  4. M510,000
  5. M95,000
  6. M1030,000
  7. M1120,000
  8. M1210,000

108,000 jobs across 8 months est.

[FEEDS LAB DECISIONS]

  • Diagnostics lead, months 2 and 10: Expected signal class; minimum detectable yield; detector exposure plan
  • Device lead, month 12 onward: Live corrections of contact-stress and field models from first device data
K1 LAMMPS (EAM)10⁶ atom contact-pulse transient 16 cores 6 h 96 core-h CHF 4.50 20 333

Throughput ceiling est.20 jobs per slot-month · 333 per server-month

[RUNS IN THE CAMPAIGN PLAN · SCALE B BASIS]

  1. M260
  2. M3300
  3. M8200

560 jobs across 3 months est.

[FEEDS LAB DECISIONS]

  • Lab lead, month 3: Maximum drive amplitude; particle size window (1 to 10 micrometer baseline)
K2 LAMMPS (ReaxFF)Ti-H/Zr-H interface or fracture cell 32 cores 24 h 768 core-h CHF 36.00 2.6 41

Throughput ceiling est.2.6 jobs per slot-month · 41 per server-month

[RUNS IN THE CAMPAIGN PLAN · SCALE B BASIS]

  1. M3100
  2. M1050

150 jobs across 2 months est.

[FEEDS LAB DECISIONS]

  • Lab lead, month 3: Maximum drive amplitude; particle size window (1 to 10 micrometer baseline)
K3 Quantum ESPRESSO64-atom relaxation or strain point 32 cores 8 h 256 core-h CHF 12.00 7.8 125

Throughput ceiling est.7.8 jobs per slot-month · 125 per server-month

[RUNS IN THE CAMPAIGN PLAN · SCALE B BASIS]

  1. M1800
  2. M2400
  3. M4300
  4. M11200

1,700 jobs across 4 months est.

[FEEDS LAB DECISIONS]

  • Lab lead, months 1 to 2: Hydride choice per stage: deuteron density versus price versus mobility (PdD0.7 versus TiD2 versus ZrD2 versus NiH)
K4 Quantum ESPRESSO + phonopyPhonon + quasi-harmonic volume sweep 32 cores 48 h 1,536 core-h CHF 72.00 1.3 21

Throughput ceiling est.1.3 jobs per slot-month · 21 per server-month

[RUNS IN THE CAMPAIGN PLAN · SCALE B BASIS]

  1. M1120
  2. M2200

320 jobs across 2 months est.

[FEEDS LAB DECISIONS]

  • Lab lead, month 2: Temperature window; thermal expansion mismatch tolerances for particle coating and cell fit
K5 GPAWHybrid-functional interface slab 48 cores 72 h 3,456 core-h CHF 162.00 0.6 9

Throughput ceiling est.0.6 jobs per slot-month · 9 per server-month

[RUNS IN THE CAMPAIGN PLAN · SCALE B BASIS]

Not listed as a dominant class in any month of the Scale B plan; it runs inside the weekly mix the coordinator re-balances.

[FEEDS LAB DECISIONS]

Supports the campaigns above rather than closing a decision directly.

K6 CP2K1000-atom ab initio MD trajectory (5 to 20 ps) 48 cores 96 h 4,608 core-h CHF 216.00 0.4 7

Throughput ceiling est.0.4 jobs per slot-month · 7 per server-month

[RUNS IN THE CAMPAIGN PLAN · SCALE B BASIS]

  1. M660
  2. M1020

80 jobs across 2 months est.

[FEEDS LAB DECISIONS]

  • Lab lead, month 3: Maximum drive amplitude; particle size window (1 to 10 micrometer baseline)
K7 PRISMS-PF / MOOSE2D mechanics-coupled precipitation front 16 cores 24 h 384 core-h CHF 18.00 5.2 83

Throughput ceiling est.5.2 jobs per slot-month · 83 per server-month

[RUNS IN THE CAMPAIGN PLAN · SCALE B BASIS]

  1. M6400
  2. M7200

600 jobs across 2 months est.

[FEEDS LAB DECISIONS]

  • Lab lead, months 7 to 8: Loading ramp protocol; duty cycle (drive-on fraction)
K8 PRISMS-PF / MOOSE3D precipitation under oscillatory strain 48 cores 48 h 2,304 core-h CHF 108.00 0.9 14

Throughput ceiling est.0.9 jobs per slot-month · 14 per server-month

[RUNS IN THE CAMPAIGN PLAN · SCALE B BASIS]

  1. M7100
  2. M880
  3. M1160

240 jobs across 3 months est.

[FEEDS LAB DECISIONS]

  • Lab lead, months 7 to 8: Loading ramp protocol; duty cycle (drive-on fraction)
K9 ElmerHorn-plus-cell eigenfrequency or driven acoustics 16 cores 6 h 96 core-h CHF 4.50 21 333

Throughput ceiling est.21 jobs per slot-month · 333 per server-month

[RUNS IN THE CAMPAIGN PLAN · SCALE B BASIS]

  1. M1200
  2. M4400
  3. M5300
  4. M7200
  5. M8500
  6. M9600
  7. M11300
  8. M12200

2,700 jobs across 8 months est.

[FEEDS LAB DECISIONS]

  • Device lead, months 1, 8, 9: Horn geometry; drive frequency (20 versus 40 kHz class); power ceiling below cavitation onset
K10 FEniCSxPiezo grain time-domain elastodynamics + electrostatics 8 cores 4 h 32 core-h CHF 1.50 62 1,000

Throughput ceiling est.62 jobs per slot-month · 1,000 per server-month

[RUNS IN THE CAMPAIGN PLAN · SCALE B BASIS]

  1. M3800
  2. M51,500
  3. M92,000
  4. M121,000

5,300 jobs across 4 months est.

[FEEDS LAB DECISIONS]

  • Device lead, month 5: Carrier fluid choice (field survival); gap budget against the 10⁸ to 10⁹ V/m design target
  • Device lead, month 12 onward: Live corrections of contact-stress and field models from first device data
K11 LIGGGHTS-PUBLICDEM force-chain packing sweep 32 cores 12 h 384 core-h CHF 18.00 5.2 83

Throughput ceiling est.5.2 jobs per slot-month · 83 per server-month

[RUNS IN THE CAMPAIGN PLAN · SCALE B BASIS]

  1. M3150
  2. M4200

350 jobs across 2 months est.

[FEEDS LAB DECISIONS]

  • Formulation lead, month 4: Piezoelectric volume fraction (about 10 percent baseline); packing fraction; hydride-to-piezo size ratio
K12 MACEPotential training run (CPU) 48 cores 24 h 1,152 core-h CHF 54.00 1.7 28

Throughput ceiling est.1.7 jobs per slot-month · 28 per server-month

[RUNS IN THE CAMPAIGN PLAN · SCALE B BASIS]

  1. M620

20 jobs across 1 month est.

[FEEDS LAB DECISIONS]

Supports the campaigns above rather than closing a decision directly.

* K0 is capped at 3,000 per slot-month by orchestration overhead. Per server-month = 16 slots. Classes share the same compute pool, so the campaign mix, not the ceiling, sets realised counts.

DETAIL The throughput model and a balanced month

Formula: jobs per slot-month per class = 2000 usable core-hours divided by core-hours per job, with K0 additionally capped at 3000 per slot-month by orchestration overhead (all estimates). Per-server figures multiply by 16 slots. Classes compete for the same pool; the campaign mix, not the ceiling, sets realized counts.

Illustrative mixed portfolio per server-month under a balanced campaign plan (estimates): 1000 K0, 60 K1, 15 K2, 80 K3, 12 K4, 4 K5, 4 K6, 40 K7, 6 K8, 120 K9, 250 K10, 40 K11, 4 K12. The mix is re-balanced weekly by the coordinator.

Queue §4 taxonomyQueue §5 throughput

[03.8 · COST AND SCALE]

CHF 93.75
buys an agent
slot for a month.

CHF 1,500 per server-month ÷ 16 slots = CHF 93.75. A fixed rate. On top sits the agents' reasoning, paid in language-model tokens, and the fleet grows with the program.

  • [SERVERS] CHF 180k 1,920 slot-months × CHF 93.75 = CHF 180,000 est.
  • [LLM TOKENS] CHF 384k 1,920 slot-months × CHF 200 planning figure = CHF 384,000 est.. On the plan's band of CHF 100 to 350 per slot-month: CHF 192,000 to 672,000.
  • [THE WHOLE ENGINE] CHF 564k about 20 percent of the CHF 2,839,023 program est., for 160 slot-years of round-the-clock work

[FIG 04] Engine cost per month, CHF est.

18.8k M1
18.8k M2
37.6k M3
37.6k M4
37.6k M5
37.6k M6
37.6k M7
37.6k M8
75.2k M9
75.2k M10
75.2k M11
75.2k M12
Servers (CHF 1,500 × servers) LLM tokens (CHF 200 × slots)

Why the plan steps instead of picking one size.

Foundations run on 4 servers while the villa is fitted out, the standard plan on 8 through the screening and carrier months, and the full 16 with a nightly lane that assimilates lab data from Gen 2 bring-up onward. Servers bill monthly, so the fleet can step up one scale on one month's notice.

Queue §9 cost summaryPlan §4Plan §8 buffers
  1. Scale A all year · 4 servers CHF 72,000 768 slot-months · Foundations: DFT, kernels, MD, 2D phase field, horn acoustics; 3D phase field quarterly; one MLIP generation
  2. Scale B all year · 8 servers CHF 144,000 1,536 slot-months · Standard plan: weekly 3D phase field, full-cell FEM, four MLIP generations, full matrix sweeps
  3. The plan A → B → C CHF 180,000 1,920 slot-months · 2 months at A, 6 at B, 4 at C est.
  4. Scale C all year · 16 servers CHF 288,000 3,072 slot-months · Everything in B plus replicate ensembles, 20 to 1000 kHz frequency sweeps, and live assimilation of lab data with weekly re-planning

Server figures exclude LLM tokens (shown separately above), laboratory diagnostics and storage beyond the servers' own disks.

[03.9 · PROVENANCE]

Every number that reaches the lab carries its provenance.

  • [PUBLISHED]

    Only results that pass the automatic acceptance checks at ingest reach the brief.

  • [FAILURES KEPT]

    Failed jobs stay in the database with their failure class, for the weekly report. Nothing is quietly dropped.

  • [RESTARTABLE]

    Duplicate idempotency hashes are dropped, so any server can be restarted at any time without double-counting.

  • [REBUILDABLE]

    Nightly filesystem snapshots, offsite backup of the database and briefs, and a single-command rebuild of any server.

[FIRST CHECK · 2026-11-30]

Seven long-lead purchase orders signed; Marbell Scale A queue running; the results database holds accepted DFT rows with the palladium lattice constant within 1 percent of 3.89 angstrom and the beta-PdD expansion inside the 3 to 4 percent reported band.

Plan §8 milestonesQueue §8 operating notesToolchain §9
DETAIL One results row, field by field (schema sketch)
results · one row[SCHEMA SKETCH]
job
‹id› · class K3 · Quantum ESPRESSO
observable
‹observable› ‹value› ‹unit› ± ‹uncertainty›
context
composition · strain · drive · carrier
acceptance
convergence · energy drift · conservation · physical anchor ✓
code version
‹release›
container digest
sha256:‹digest›
input commit
‹input-factory commit›
config hash
‹configuration hash›
checkpoint
‹path›
artifacts
‹path› · checksum · size
estimate tag
set wherever the value is an estimate
Field names from the queue document (§3.5, §8) and the toolchain validation ladder (§9); values shown as placeholders.
DETAIL Operating notes, as written
  • Weekly re-planning: the coordinator agent publishes a report every Monday 06:00 from database state; humans steer by exception.
  • Escalation: repeated job failure moves supervision agent to coordinator to human on-call within one hour.
  • Every number that reaches the lab carries provenance (code version, container digest, input hash) and an estimate tag where it is an estimate.
  • The queue publishes only results that passed the automatic acceptance checks at ingest; failures stay in the database with their failure class for the weekly report.
  • Data hygiene: nightly snapshots of the filesystem, offsite backup of the database and briefs, and a documented single-command rebuild of any server from the Ansible playbooks.
SOURCE The simulation queue document, section by section

[GO DEEPER]

SOURCE The simulation queue, in full Architecture, taxonomy, throughput, campaign plan, decision link, cost: 2,400 words SOURCE The simulation toolchain, in full Every code, what it computes for this device, runtimes, licences, validation ladder SOURCE Campaign plan, month by month Queue §6: scales A, B and C, with the adjustment rules PROGRAM The 12-month plan Where every lab decision lands, with budget, team and milestones PROGRAM The lab The villa laboratory that receives the Monday brief