▤DataCenterTwinLAB
ENGINEERING / COMMISSIONING RELEASE 2

Inspect the assumptions.

Validation status

Software validation and numerical reference cases are tested. Actual CPU/WebGPU parity is tested on available hardware. Real-facility calibration is not yet established. Local CSV comparison, bounded parameter fitting and chronological holdout are available in Analyze → Validation Lab; synthetic test success is not measured accuracy.

Parameter ensembles report conditional planning intervals. They do not include unquantified model-form discrepancy or measurement uncertainty. Completed validation studies retain the baseline model, dataset hash, engine version, bounds, seed and assumptions locally.

These are planning calculations, not sensor readings, certified designs or validated industrial CFD. Numerical precision does not establish physical accuracy. All facility documents and simulations remain local to the browser.

Workload and electrical network

Piecewise-linear workload profiles scale rated equipment demand. A directed acyclic graph routes that demand through sources, ATS, UPS and distribution. Dual feeds split load; a surviving feed carries the full request. Failed routes remove served load and its heat.

S=P/PF; balanced three-phase current is I=1000P/(√3 VLL PF). Optional rack phase shares, node power factors, apparent-power limits and current limits extend the basic calculation. This is approximate phase allocation, not a phasor load-flow or fault-current study. Optional generic thermal trip accumulation and priority shedding add dynamic consequences; they are not manufacturer protection coordination.

Optional efficiency curves are iterated with a bounded solver. UPS battery depletion includes discharge efficiency and minimum SOC. Optional ATS detection/start/transfer/retransfer delays respond to source state. Optional user fuel curves constrain generator operation. Recharge includes charge efficiency and upstream demand. Optional user pickup/TCC data causes current-based trips and supports indicative coordination samples. Battery aging, engine ramp, short-circuit and certified selectivity are not modeled.

Air heat and spatial field

The coupled timeline uses conservative supply/return and rack inlet/exhaust control volumes. Reference air density is 1.2 kg/m³ and specific heat is 1.006 kJ/(kg·K). Effective rack thermal capacitance is configurable and uncalibrated. Internal transport conserves energy; sinks are capacity/setpoint bounded. Initial preconditioning is not a guarantee of steady state.

The separate deep solve constructs a 3D Cartesian field with obstacles, heat sources and pressure-projected velocity. Fan speed, pressure and resistance affect initial forcing through a reduced-order sensitivity model. Compatibility retains frozen flow; opt-in unsteady Stokes-Boussinesq updates couple temperature, buoyancy and velocity with separate solid/air storage. Relative pressure is a correction diagnostic, not validated absolute static pressure. An additional opt-in mode adds first-order semi-Lagrangian momentum advection and GPU pressure projection. Turbulence, indoor humidity transport and calibrated porous-rack models remain unsupported.

Real WebGPU compute can advance spatial thermal transport when adapter/device initialization succeeds; thermal batches remain GPU-resident between readbacks, with an optional resident WGSL pressure solve. Momentum and solid exchange remain CPU-side; this is not full GPU residency. The CPU fallback is explicitly labeled. Grid dimensions, iterations, elapsed time, divergence and energy residuals are reported. The spatial solve captures one operating state rather than continuously replacing the event timeline.

Hybrid liquid cooling

Configured hybrid racks split served IT heat between room air and cold-plate storage. Facility water, a primary pump, a capacity-limited CDU exchanger, secondary duty/standby pumps and parallel rack branches form each lumped loop. Darcy-Weisbach resistance intersects a quadratic pump curve. Heat transfer uses Q=ṁ Cp ΔT and counterflow effectiveness-NTU.

A temperature PID changes pump speed, actual flow and heat transfer. CDU, FWS, pump and branch failures alter those equations. Lost-flow heat accumulates in cold plates; failure-temperature timing is calculated, not scripted. Fixed water properties, a mixed coolant mass and an externally regulated FWS boundary remain simplifying assumptions. Pump motor heat, coolant chemistry and detailed chiller energy are excluded.

Optional secondary hydraulic graphs solve relative node pressures and flows with Reynolds-dependent Darcy friction, valves and parallel pumps. Eight-cell supply pipes can delay temperature transport. Constant reference-water properties remain; glycol, leaks and cavitation are not supported. An optional lumped plant now updates the primary facility-water boundary.

Opt-in equipment RC response reduces workload at configured throttle limits and latches shutdown. BMS sensors include bias, noise, drift, sample timing, latency and stuck/offline failures. Temperature PID outputs change real fan airflow/power or pump speed. Trend Lab compares actual and measured signals with actuator output.

Capacity, resilience and search

Current capacity findings, automated N−1 contingencies, selected/capped N−2 cases and sampled worst-case timeline summaries have distinct scopes. Finite batteries do not establish sustained redundancy. Seeded reliability requires supplied failure parameters. Bounded multi-rack/hybrid placement reports electrical, thermal, liquid and service constraints and nondominated tested alternatives. The separate fixed-position load search remains a restricted screen. Neither establishes a global maximum safe MW. Two cords do not certify 2N, Tier compliance or uptime.

An explicit energy boundary gates modeled PUE; incomplete boundaries show a partial energy ratio. Inlet limits are configurable; the default 27 °C warning does not establish equipment-class compliance. Physical Validation can run 24/40/64 grid refinement and bounded spatial airflow calibration with training/holdout probes. Measured facility accuracy remains unestablished.

Digital commissioning and evidence

Design intent, topology-generated failures and editable acceptance rules compare simulation with optional local measured CSV. Missing required observations remain conditional. Timing uses integer-second threshold crossings. Immutable local records retain model revisions, inputs, results and assumptions. Software tests, numerical convergence and measured physical validation are distinct evidence.

Plant, building and long operations

Generic or user-entered chiller performance, ambient derating, dry-cooler economizer and wet-bulb tower approximation drive facility-water storage and routed electrical demand. Weather CSV stays local. ASHRAE psychrometric relationships calculate outdoor humidity ratio, dew point, enthalpy and wet bulb; no indoor moisture or latent-coil model is claimed. Modeled PUE requires complete declared auxiliaries, reconciled components and a storage-neutral interval; otherwise a partial energy ratio is shown.

Rectangular rooms, adiabatic walls and doors affect reduced-order heat exchange and spatial obstacles. Perforated tiles use a pressure-bounded uniform plenum/orifice model; there is no resolved underfloor CFD or IFC import. Operational workers support up to seven days with event-aligned exchange and fast thermal substeps. Operational workers support in-memory pause/resume; persistent checkpoints are outside V1.

Robust design screening evaluates captured alternatives on the same seeded, independent uniform scenarios and explicit constraints. Pass fractions are sample outcomes, not calibrated reliability probabilities. User costs are compared only in a common unit. No global optimum or maximum safe MW is claimed.

Verification and references

Tests cover unit roundtrips, power routing, phase/current constraints, thermal conservation, pressure projection, liquid equations/failures, workload coupling, control response, resilience and persistence. Software tests are distinct from empirical engineering validation. Rendering quality does not change physical values or solver resolution.