What AI Is Made Of

Run 03 transformer material forensics

No numerical transformer material coefficient is admitted to the fixed plant's primary balance. The investigation produces reproducible physical models, explicit validation errors, conductor alternatives, a closer PDU gross-mass sensitivity and specific unresolved inputs. UNKNOWN remains distinct from zero.

Canonical SNAP-RUN02 calculation supplies 24 campus 84 MVA 230/34.5 kV transformers in B3; 428 IT 5 MVA and 214 facility 3.5 MVA 34.5/0.48 kV transformers in B2; and 1,712 isolated-transformer 1 MVA 480-to-415 V PDUs in B2. The latter are an additional transformer-bearing class. No counts, topology or equipment selections were changed.

All four classes merit high Cu/Al/electrical-steel research priority. The large population of MV/LV and PDU transformers makes even moderate per-unit winding/core mass potentially material. This is a priority inference, not a coefficient. An exact transformer catalogue model, cooling stage and conductor are still absent. Rating alone cannot identify physical mass.

Physical reconstruction and validation

The winding model uses rho × bare conductor area × physical turns × mean turn length × phases. A parameter interface relates MVA and voltage/connection to phase current, current density to bare area, and frequency/flux/net core section to nominal turns. It requires physical tapped turns and winding geometry as additional inputs; cooling and impedance do not map uniquely to mass.

Three independent design-team cases from the CLASP 2011 dry-transformer report provide geometry and engineering BOMs. Across six HV/LV windings, reconstructed copper differs from published design mass by at most 0.209%. Physical maximum tap turns are necessary: using nominal turns omits the extra copper. Bare wire/foil area excludes insulation. These are ENGINEERED design checks, not measured teardowns or declared selected equipment.

A symmetric three-limb core approximation uses rho × net iron section × [3 × (clear-window height + yoke depth) + 4 × limb pitch]. Diagram dimensions and equal-section assumptions are explicit. Three core checks differ by at most 1.222%. Joints, nonuniform yokes and detailed core construction remain unresolved. Net area already excludes voids, so a second stacking factor would double-remove them.

Good agreement only establishes conditional plausibility with complete geometry. It does not validate transfer to the selected 84/5/3.5/1 MVA classes. Zero measured physical-object validation cases were retrieved. Selected conductor, cooling/type, winding connection, flux, bare area, turns, geometry, loss specification and gross mass remain absent. Target low/base/high therefore remain null.

The tempting shortcut m = density × load loss / [electrical resistivity(T) × mean current density squared] was tested against all 18 CLASP 2013 large-transformer designs. Its mean absolute error is 21.348%, with errors from −18.001% to +70.746%. Treating aggregate load loss as DC winding loss and reported mean current density as effective per-winding density fails. This shortcut is rejected, including for the base case. No kg/MVA regression or linear scaling was used.

New physical evidence and source-only scenarios

The Vertiv PPC brochure and installation manual provide a close PDU family: 480 V, 60 Hz input, 415 V output, isolation transformer and Cu/Al options. However 1,000 kVA is engineer-to-order; the weight table stops at 750 kVA. Scaling that weight would invent a relationship.

Schneider's Galaxy PMM1000-CUB has the target 1,000 kVA rating, 480-to-415 V function and a copper transformer. Whole-PDU installed gross is 4,550 kg, with 4,700 kg shipping mass excluded. Applying canonical count yields a 7,789.6 t source-product gross sensitivity. Manufacturer, selected conductor, options, dimensions and selected gross are unqualified. This is a separate QUALIFIED PROXY scenario, not a Cu/Al/steel coefficient or a finite physical bound on selected equipment. The PDU transformer is already nested inside this gross.

A Larson manufacturer catalogue offers a near-voltage-matched 5 MVA, 34.5/0.48 kV, 60 Hz, Al-wound KNAN product. It supplies a useful aluminum-conductor lead, but fluid volume conflicts (700 versus 190 gallons), weights leave a 666 lb discrepancy, and listed losses conflict with listed efficiency. Material transfer and gross bounds are rejected.

Public NYISO procurement/interconnection evidence describes an 84/112/135 MVA two-winding 230/34.5 kV transformer with ONAN/ONAF/ONAF ratings, alongside a separate three-winding product. It reinforces the need to identify cooling stage and winding topology before matching an “84 MVA” object. Neither product is selected for this plant.

Conservation, bounds and terminology

The CLASP 2011 priced BOMs including cabinet exceed the reported transformer weight by 245, 79 and 85 kg. Cabinet exclusion from bare-transformer weight may explain the scope mismatch. The combined BOM is inadmissible against that gross until clarified; no source typo is asserted and no negative residual is assigned to steel. Core-plus-winding subtotals independently remain below gross and leave unresolved residuals.

Core electrical steel, structural steel, cast iron, generic iron, copper conductors, aluminum conductors, oil and insulation remain separate. Neither a whole transformer nor a residual is relabeled steel. Material-specific finite upper bounds on selected transformers remain unavailable because exact gross mass is unknown. Source-product gross masses cannot bound an unselected object. Nonnegativity supplies only a trivial zero lower bound, with high null.

Cu/Al alternatives preserve unresolved selected conductors without changing equipment topology. Their material masses are null. B2 and B3 transformer material closure remains OPEN. Isolation-transformer composition must replace the corresponding PDU parent residual when it becomes available; it must never be added to a complete parent mass.

Files and checks

Run:

python research/run03/transformers/build_handoff.py python research/run03/transformers/build_registers.py python -m unittest discover -s research/run03/transformers -p 'test_*.py' -v

Run02 source conflicts are preserved: CLASP 63 MVA case4 detailed core428,500 kg versus summary42,850 kg, with cost arithmetic supporting42,850; 25 MVA case2 load loss98 versus98.5 kW. SNAP-RUN02 remains unchanged.

Source details

Archived Run 01-03 research. These statements retain their time-of-writing context; current material estimates use Run 04. Web prose uses current site terminology; downloadable records preserve original wording. Hashes identify the downloadable records.

Source: research/run03/transformers/HANDOFF.md
SHA-256: d66584c06de5070fbef497b507720b34f885f6af47f41087d1fadc6c692b9aa1
Download research file

Optional: tell Sid you’re reading
The document is open; sharing your name/email is optional.