What AI Is Made Of

Run 03 battery material forensics

The selected battery cell still cannot be identified. This work establishes a physically derived minimum active inventory, three explicit cell-recipe scenarios, and a failed collector-model transfer. It does not qualify a selected Li5 copper, aluminum, steel or lithium base case.

All quantities use calculate_balance(..., snapshot="SNAP-RUN02") and the canonical BOM for PLANT-GB300-1GW-2026Q4-A. The architecture and equipment records are unchanged. Files in this directory are additive research records only.

Preserved plant and physical nesting

Selected classCabinetsModulesPhysical cells, 9S3P/moduleGross cabinet mass, t
IT batteries7,704138,6723,744,1444,545.360
Facility batteries64211,556312,012378.780
Combined B2 incremental8,346150,2284,056,1564,924.140

The exact current OEM manual specifies 60 Ah nominal cabinet capacity, eighteen series modules, 9S3P/module, LFP chemistry and convective cooling without fans. The exact-family OEM datasheet retained in Run 01.5 specifies LFP cylindrical cells (R15-CLM-E-ENERGYCORE-LI5). Three parallel cells imply 20 Ah nominal per physical cell. This is a capacity interpretation, not a cell SKU. Canonical 31.1 kWh/cabinet remains preserved; 518.4 V × 60 Ah = 31.104 kWh is only a rounding check. The unresolved facility UPS interface qualification from Run 01.5 remains.

Cabinet includes modules; module includes cells, busbars, monitoring/balancing electronics, supports, wiring and thermal/insulating materials. Gross assembly, active material, constituent elements and residual are alternate nested views. They are not additive stock categories.

Physical lower bounds

For the one-electron LFP reaction, one mole of reversible lithium and one mole of FePO4 framework support at most F/3600 Ah, where F = e × N_A = 96485.33212331001 C/mol from exact SI constants. Thus:

n_min = physical_cell_count × 20 Ah × 3600 C/Ah / F

m_min = n_min × molar_mass

Using conventional Li 6.94, Fe 55.845, P 30.973761998 and O 15.999 g/mol gives LiFePO4 molar mass 157.754761998 g/mol and theoretical specific capacity 169.893326 Ah/kg. At fixed nominal capacity:

Material viewCombined battery lower bound, t/GW ITBaseHigh
LiFePO4 fully lithiated active-equivalent477.494nullnull
Reversible elemental lithium inventory21.006nullnull
Iron in the LFP framework169.032nullnull
Phosphorus in the LFP framework93.752nullnull
Oxygen in the LFP framework193.704nullnull

The element rows partition the LFP equivalent; do not sum them with their parent. Lithium consumed during formation, excess active reserve and electrolyte lithium can raise stock above the minimum. State of charge moves lithium between electrodes: the LFP equivalent describes the fully lithiated accounting basis, not a measured installed LiFePO4 phase inventory. These bounds are conditional on ordinary terrestrial conventional atomic masses and the specified LFP redox mechanism. They are not isotope assays or an exact OEM formulation.

The cathode/framework and mobile-lithium budget must be non-copper, non-aluminum and non-steel. Consequently each individual battery upper bound is:

material_max = 4,924.140 t gross − 477.494 t minimum active-equivalent = 4,446.646 t

Copper, aluminum and steel individually have 0 / null / 4,446.646 t outer bounds. All share the same gross budget and cannot be summed. The zero lower bound records lack of a positive quantified coefficient, not material absence. This exposure is too large to dismiss battery uncertainty as immaterial. Nothing allocates the rest to steel.

Total elemental lithium has a separate conservative whole-equipment ceiling: gross − minimum non-lithium FePO4 framework. Thus its outer bound is 21.006 / null / 4,467.652 t/GW IT. This permits lithium in electrolyte, excess active inventory and unresolved components. It does not multiply cabinet gross by the LFP lithium fraction, and its large upper endpoint is a physical exposure ceiling rather than a plausible composition estimate.

The engine-ready proposed_forensics.json contains two COMMON records, one per exact IT/facility battery BOM class, and six conditional recipe records. COMMON retains 590 kg exact cabinet gross, Faraday minima and explicit whole-equipment bounds with null central estimates. The active-equivalent parent and its elemental views describe one mass budget, including mobile lithium independent of state of charge. Each recipe names the corresponding replaces_common_id, so a coherent scenario selects one decomposition per equipment class. Every recipe carries the independent active-equivalent and Li/Fe/P/O Faraday minima and whole-equipment ceilings forward on its existing components, with a null steel bound component; substitution cannot erase those constraints. Active and salt lithium are never added as separate whole-class lower bounds. All eight records pass the current engine validator and physical/nested conservation. COMMON is ENGINEERED / scenario for bounds only; no uncalibrated material coefficient enters primary or expanded forensic subtotals.

Cell-composition scenarios

A new primary Discover manufacturer SDS, Rev T, July 10, 2025, specifies multiple exact reference-cell recipes. The complete IFR32650 cylindrical recipe is:

Declared reference-cell categoryMass fraction
LiFePO427.04%
Graphite12.78%
Aluminum6.44%
Copper9.22%
Iron, type unspecified23.52%
Nickel1.18%
Polyethylene4.37%
LiPF62.01%
Organic electrolyte solvents13.44%

These sum to 100% for this reference recipe. The HELIOS ESS GSP71173204F and AES Rackmount GSP50160119F declared recipes sum to 95.05% and 91.22%; missing mass remains residual. The source's statement that metallic lithium is zero does not set elemental lithium in LFP or electrolyte to zero.

The cell reconstruction uses electrical capacity and chemistry before applying the recipe:

m_active = Q / q_effective

m_cell = m_active / declared_reference_LFP_fraction

m_material = m_active × declared_reference_material_fraction / declared_reference_LFP_fraction

Three coherent capacity-utilization cases are retained. Low is ideal one-electron theoretical utilization; base uses the explicit 160 Ah/kg LFP input in Lechner et al. 2024; high uses that same input with the paper's 7.29% initial irreversible capacity-loss reserve, q_effective = 160 × (1 − 0.0729). These are conditional engineering cases, not confidence limits and not a guaranteed envelope for every possible Li5 cell. Additional reserve could exceed the high case.

Conditional recipe transferred to selected cell countCu low/base/high, tAl low/base/high, tGraphite low/base/high, tCell mass low/base/high, kg
Cylindrical IFR32650162.8 / 172.9 / 186.5113.7 / 120.8 / 130.2225.7 / 239.6 / 258.50.435 / 0.462 / 0.499
HELIOS ESS GSP71173204F81.3 / 86.3 / 93.152.5 / 55.8 / 60.1239.0 / 253.8 / 273.70.292 / 0.310 / 0.335
AES Rackmount GSP50160119F56.4 / 59.9 / 64.653.3 / 56.6 / 61.1263.2 / 279.4 / 301.40.346 / 0.367 / 0.396

All three have the same explicit active-LFP cases, 477.5 / 507.0 / 546.9 t. Elemental lithium in this active-equivalent parent is 21.0 / 22.3 / 24.1 t. In the cylindrical recipe, an additional 1.6 / 1.7 / 1.9 t lithium view is calculated from the separately booked LiPF6 salt; it is not added to salt gross in the physical sum. The exact electrolyte recipe is not identified for Li5, so none is admitted as selected total lithium.

The cylindrical case retains a 415.3 / 441.0 / 475.7 t iron category, with the source's unspecified metallurgical identity. It is not steel. The cell recipe alone does not cover cabinet/module busbars, frame, controls or wiring. Subtracting the conditional cell projection from cabinet gross leaves 2,901.6–3,158.3 t, with 3,049.1 t in the base case. This is a source-calculated residual from the conditional cell recipe; it is not a fully known physical decomposition. The engine separately records known projected cell masses and leaves full physical sum/residual endpoints null where the steel bound-only component has no composition coefficient. The alternatives leave still more unresolved mass.

Every reference/target mismatch is recorded: manufacturer, family, function/rate, capacity, voltage, cylindrical versus unresolved/energy formats, dimensions, gross mass, cooling interface and vintage. IFR32650 is only 5 Ah and has no qualified Li5 supplier lineage; the two ESS recipes are further mismatched. These are PROXY / scenario records, excluded from declared/design-qualified copper and expanded forensic copper. Arithmetic precision does not strengthen transfer.

Validation results

Exact-family gross partition: The current Vertiv manual lists 400, 448, 495, 543 and 590 kg at 10, 12, 14, 16 and 18 modules. The endpoint model gross = 162.5 + 23.75 × modules kg, trained only at 10/18 modules, predicts withheld 12/14/16 population masses within 0.5 kg, maximum 0.112% relative error. Reported adjacent increments are 23.5–24 kg/module. Applied diagnostically at the selected 18 modules this gives 423 / 427.5 / 432 kg module-associated population and 158 / 162.5 / 167 kg nonpopulation residual. The slope can include shelves and interconnects; the intercept is not measured empty steel. This validates a gross assembly relationship within the exact published family, not an elemental coefficient.

Measured large cylindrical LFP object: Liu et al. 2025's BYD 4680 teardown measures cathode foil 5,370 × 69 mm, single-sided areal capacity 2.09 mAh/cm², and full-cell 15.41 Ah. 2 × area × areal_capacity predicts 15.488154 Ah: 0.507% error. This validates area/capacity bookkeeping. Measured 10 μm foil geometry reconstructs 35.071 g Cu collector and 10.004 g Al collector, but no independent weighed metal masses validate these coefficients. The source cells' maximum continuous characterization is 2C and their identity is not Li5; target transfer remains unadmitted.

Failed high-power collector transfer: The previously captured DOI 10.3390/batteries11010038 paper is Wätzold, Schlösser, Leistikow and Kowal, published January 20, 2025, titled *Design Analysis of 26650 and 18650 LFP Cells for High Power and Low Temperature Use Cases*. Historical Run 02 prose misattributes it to Probst; the retained original establishes the correct new metadata. Literal density × foil_area × reported_currentcollector_thickness is tested for all five physical objects. Three exceed measured cell gross after adding only the minimum LFP cathode, before graphite, casing and electrolyte. The ANR26650M1B case predicts 65.043 g Cu + 14.631 g Al = 79.674 g collectors against 75.2 g measured gross. The literal model is rejected. This records a data/interpretation conflict, without asserting a corrected thickness or the authors' intended interpretation.

The active-recipe scenarios have no independent measured Li5 composition calibration. Passing source-fraction and cabinet-gross conservation does not qualify them for the base balance.

Admission, priority and next evidence

Battery-active material is high priority: the exact design has 259.5606 MWh nominal DC storage and nearly 5 kt of cabinet gross. Battery Cu/Al/steel are material-priority classes because several kilotonnes of physical exposure remain unresolved. Manufacturer cell/module recycling instructions, current module SKU and matched cell SDS or teardown would materially tighten these bounds. Cabinet drawings with actual sheet thickness and busbar geometry could quantify separate structure/conductors without waiting for cell identity.

Strong-evidence battery Cu/Al/steel subtotals: null. Expanded forensic battery Cu/Al/steel subtotals: null. Outer bounds: available, too wide for provisional closure. Per-material battery status remains OPEN. Bulk module and cabinet residual is explicit; semiconductor, civil, extraction and upstream manufacturing scopes are not expanded.

Reproduce offline with:

python research/run03/battery/derive_battery_forensics.py

evidence.json includes canonical count lineage, COMMON bound records and recipe scenarios compatible with the proposed Run 03 relation, explicit equations and source claims, nested element views, validation cases, residuals, bounds and admission decisions. proposed_forensics.json is the eight-record integration candidate; proposed_conflicts.json uses the existing canonical conflict schema for the failed literal collector interpretation. sources.json, claims.json, source_manifest.json, search_log.json and captured original PDFs/HTML/text preserve the evidence trail. Raw connected-search captures are in web_capture.json. No canonical tables were edited by this workstream.

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/battery/HANDOFF.md
SHA-256: 8651c0b2a9ce601c89273bf6c4cbd29f85f648336fc3f15a816a6184ec4ca26f
Download research file

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