{
  "abstract": "Laboratories increasingly base formulation, protocol and safety decisions on battery models, yet models rarely receive the controls that laboratory methods do: versions, validation on unseen data, reproducible records and verification of the numerical solution. We report a suite of 9 registered models that run inside a laboratory platform (a cell simulator with equivalent-circuit, single-particle, electrolyte-corrected and transient-electrolyte rungs, electrode balancing, pulse and impedance predictors, a plating-limited charging optimiser, a receding-horizon charge controller, an unscented Kalman filter with a stability review, and a mechanistic ageing model), together with the results they produce for a published cell parameterisation and the evidence for trusting those results. Half-way through a 1C discharge, solid diffusion, electrolyte transport and contact resistance cost 85, 49 and 50 mV; the electrolyte-corrected and transient-electrolyte rungs agree within 25 mV at 1C but diverge by 1,051 mV at 3C, where the latter depletes its electrolyte. A plating-limited current staircase reaches 80% state of charge in 22.9 min at 25 ◦C and 83 min at 0 ◦C, and the closed-loop controller needs 21.5 min without a constraint violation, against 28.0 min for the fastest constant current that respects the constraints. Verification against analytic solutions finds the particle solver second-order accurate in its concentration profile but first-order in the surface value that sets the voltage, a reported state of charge that exceeds the charge balance by up to 19.8 points because shells are averaged without volume weights, and a reversed sign in a kinetic term that is numerically negligible. Varying 9 shipped parameters together by a factor of 1.5 gives a 90% voltage band of ±50 mV at 1C, dominated by the positive particle radius and larger than both the discretisation error and the difference between the two electrolyte rungs; at 3C the choice of electrolyte closure outweighs every other source. That budget assumes its parameters reach the decisions a laboratory makes, and they do not: moving the solid diffusivity over two decades changes the delivered capacity at 3C by 34.8 points while leaving the plating limit, the staircase time, the controlled charge time and the least anode potential unchanged in every digit reported, because the charging model reads 2 cell parameters and the controller 2. Where the discharge rung does respond its capacity collapses onto the diffusion time R2/D to within 0.7 points, and because the particle equation is stepped explicitly, the regime of small particles and fast solid diffusion is the most expensive one to simulate. A grid convergence index shows the simulated voltage converging at first order, leaving the default particle grid 30.7 mV from the extrapolated voltage at 3C; polynomial chaos confirms the linear apportioning of variance, and Gaussian-process emulators predict the voltage to 0.8 mV from 100 runs, with no gain from choosing runs by predictive spread. Drawing the same design from a Sobol sequence rather than a Latin hypercube cuts the largest index error at 256 runs from 0.082 to 0.00022, and a reduced two-parameter problem mapped in full on 625 runs puts the sampled estimator within 0.004 of indices computed exactly. Gaussian splatting of simulated fields, implemented for this study, did not compress regular two-dimensional grids better than proper orthogonal decomposition but fitted scattered samples with an error of 9.97 mV against 14.55 mV for triangulation; in three dimensions, across rate and temperature, it beat both decomposition and interpolation. An adaptation of a recent densification strategy from rendering failed at the largest budget; splitting along the axis of largest field change and band-limiting the primitives to the sample spacing made it stable, with a median held-out error over 10 fits of 13.2 mV against 14.4 mV for residual-driven control. Extended to the vehicle, battery power determines only 5 of 6 longitudinal parameters: a physics-informed network fitted it from 8 starting points with the identifiable combinations within 1.9% of the truth, while the masses it returned ranged from 0.79 to 0.98 times the true value. Re-executing the 10 recorded runs from their stored inputs reproduced 9 bit for bit; the exception showed that a run’s code reference names a month, not a version. A physics-informed extreme learning machine, fitted for this study, solves the particle-diffusion problem to a surface-stoichiometry error of 1.2× 10−4 and passes an admission tolerance that the platform’s 14-shell solver fails; a gradient-trained network of the same form improves 21-fold on least squares with the same few fixed features but still fails that tolerance. Toward a digital twin, the suite’s estimator tolerates the drift of an ageing cell, but calibrating the particle diffusivity through the default grid misses it by −17.4%, against −0.5% through the verified surrogate. Governance made each finding checkable; verification made them visible.",
  "abstract_shown": true,
  "affiliations": [
    "Department of Chemical Engineering & Applied Chemistry, University of Toronto, 200 College Street, Toronto, Canada, M5S 3E5",
    "Leslie Dan Faculty of Pharmacy, University of Toronto, 144 College Street, Toronto, Canada, M5S 3M2",
    "Department of Mechanical and Industrial Engineering, University of Toronto, 5 King’s College Road, Toronto, Canada, M5S 3G8"
  ],
  "article_type": null,
  "arxiv": null,
  "author_total": 2,
  "authors": [
    {
      "affs": [
        0
      ],
      "name": "Siddhartha Pahari",
      "orcid": "0000-0002-1586-5950"
    },
    {
      "affs": [
        1,
        2
      ],
      "name": "Jainish Shailesh Solanki"
    }
  ],
  "bibtex": "https://research.draconitic.com/paper/2xvzfjx6dmy9y.bib",
  "cited_by": 0,
  "doi": "10.26434/chemrxiv.15009462/v1",
  "fulltext_url": null,
  "id": "2xvzfjx6dmy9y",
  "issns": [],
  "landing_url": "https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009462/v1",
  "language": null,
  "licence": "CC-BY",
  "licence_label": "CC BY",
  "licence_source": "journal",
  "licence_url": "https://creativecommons.org/licenses/by/4.0/",
  "metadata_licence": null,
  "native_subjects": [
    [
      "openalex",
      "2203|Advanced Battery Technologies Research"
    ]
  ],
  "openalex": "W7214300808",
  "pdf_url": "https://chemrxiv.org/doi/pdf/10.26434/chemrxiv.15009462/v1",
  "pmcid": null,
  "pmid": null,
  "published": "2026-09-24",
  "publisher": "American Chemical Society (ACS)",
  "ris": "https://research.draconitic.com/paper/2xvzfjx6dmy9y.ris",
  "sources": [
    {
      "first_seen": "2026-10-30T12:17:21Z",
      "native_id": "10.26434/chemrxiv.15009462",
      "source": "chemrxiv",
      "url": "https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009462/v1"
    }
  ],
  "subjects": [
    {
      "field": "chemistry",
      "origin": "source:chemrxiv",
      "subfield": "general"
    }
  ],
  "title": "Battery models as controlled laboratory methods: simulated results across the operating envelope, code verification and a physics-informed surrogate",
  "title_text": "Battery models as controlled laboratory methods: simulated results across the operating envelope, code verification and a physics-informed surrogate",
  "type": "preprint",
  "updated_at": "2026-10-30T12:17:21Z",
  "url": "https://research.draconitic.com/paper/2xvzfjx6dmy9y",
  "venue": "ChemRxiv",
  "versions": []
}