Pro, on your terms.

Choose cells and pack layouts; inspect electrical, size and mass estimates.

Use the guided sizer to evaluate candidates against your duty requirements.

Road load from the machine itself — mass, drag, rolling resistance, gradient, driving mode — giving Wh/km and range, with the pack carrying its own weight.

Select one of the evidenced NTNU vessels, inspect the TwinShip-inspired component and signal contract, calculate a first-order voyage duty, grade model readiness, and compare governed measured/predicted speed, course and power samples without calling a screening model a live digital twin.

Fit a different part and see what it does. Every option on the shelf is priced BEFORE it is chosen — what it buys, what it costs, whether it breaks anything, and whether it clears a failure the design already had. A part that wins on one number and fails the audit is marked rather than ranked, and a part that does not suit the pack says so instead of quietly reporting no change.

The pack itself, in a game engine, at the size it really is — every cell where the layout put it, the cooling hardware in the space it reserved, and the audit result on the caption so a failing pack cannot be admired without being told. The renderer computes nothing: it draws a payload the design engine produced, which is what lets a face in another language exist without the tool acquiring a second opinion.

The pack shown inside the thing it powers — under the floor of a car, on the roof of a bus, in the hull of a boat, in the torso of a humanoid. Sixteen machines, each with its own shape and its own mounting, because those are four different design problems and the tool has always modelled them separately without ever showing them. The silhouette is a massing block: its cross-section comes from a measured frontal area where the application has one, its length is class-typical, and the scene says which. The PACK inside it is at true scale, so an oversized one visibly bursts out of the machine instead of the machine quietly growing to fit.

Fixed, swappable or hot-swappable as a design policy that cuts across every application. Choosing it changes four things at once: the mass stops being an outcome and becomes a requirement someone has to lift; the connector stops being a fitting and becomes a wear item with a finite mating count; you buy more packs than machines; and the pack has to survive being off the machine with no host BMS, disconnect or enclosure.

Module partition, BMS topology, the HV chain (precharge, contactors, isolation), DC-DC, and the supervisory layer above it.

Calculates and simulates the precharge resistor, screens the precharge contactor, evaluates shunt loss, accuracy and installed thermal duty, then coordinates the measured fault threshold with the R–L short-circuit trace.

What the harness must carry at cell, module, system and cooling level.

The first milliseconds of a fault: prospective current, and the race between the fuse clearing, the busbar surviving and the cells reaching runaway onset. Includes the internal short, where the parallel neighbours are the danger.

Four perspectives — mechanical, thermal, electrical, safety — plus the release checklist for your market. Every finding names the rule behind it.

The converters the pack lives between — auxiliary DC-DC, 400↔800 V boost, on-board charger, traction inverter, active rectifier — as average-value loss models: fixed + conduction (I²) + switching losses, from which the REAL efficiency curve falls out: poor at light load, peaking mid-load, sagging toward full power. The day-loss ledger runs the mission trace through the chain and reports what each converter cost; the thermal duty names the heat the cooling must move; the ripple estimate says what the switching adds to the pack's own I²R. The charge session accepts the OBC curve in place of the flat class figure. Every coefficient is a class estimate, exposed and overridable, replaced by measured data the same way a fitted cell replaces the OCV shape.

Isolation keeps fault current off the case; bonding decides what happens when isolation fails, and is asked about far less often. Every bonding path is checked three ways — continuity against the 0.1 Ω limit, touch voltage against the 60 V DC boundary, and whether the strap survives the fault current until protection clears. It catches the anodised housing that is not a bonding path however many bolts go through it, the stainless bond chosen for corrosion rather than conduction, and the machine that is conventionally ungrounded and to which none of these rules apply as written.

The pack as a connection graph rather than a number: every conductor run with its material, length and section, every joint with the two surfaces that meet there. Each run is then sized two ways — by the current-density rule of thumb and by the steady-state heat balance that says how hot it actually gets — and where they disagree the temperature answer wins. Out of the same graph come the interconnect resistance, the voltage drop and loss at continuous current, the galvanic check on every joint, and the bill of materials the customer receives.

The battery fraction of Stateflow, pre-assembled: modes, transitions guarded by battery signals (voltage, current, SoC, temperature, time in state, precharge progress), and entry actions on the five actuators a supervisor owns. The chart is typed data, not a drawing — no eval, a closed signal vocabulary, structural validation (unreachable modes refused, faults must latch), and a reference interpreter that explains every transition with the signal values that fired it. Charts run against the mission trace as observers, and generate as loop-free C held to the interpreter by an open conformance test.

Cooling loop selection, coolant flow, chiller cost, and the heater branch winter charging needs.

The machines behind the Thermal tab's loop, as the curves the flat figures hide: chiller COP against the actual lift (hot days genuinely cost more), heat pump against PTC with the crossover and the cold-weather cutoff stated, pump and fan on the affinity cube law (half the flow, an eighth of the power), and a radiator that keeps physics' own rule — never below ambient. The cooling day ledger runs the mission's own I²R heat through those curves and prices the parasitics mode by mode; preconditioning is priced hand-checkably (C·ΔT through the chosen heater) — the exact cost the mission model declines to invent when it holds the winter charge window. Every coefficient is a class estimate, stated and overridable, replaced by measured machine data the way a fitted cell replaces the OCV shape.

One cell goes — how much does each design decision help? Cell adjacency from the real layout, then conduction, radiation and the interconnect stepped in time with two thermal nodes per cell. It ranks barrier options and spacing against each other, and sizes the energy the enclosure must contain. It does NOT predict whether a pack propagates: the mechanisms that carry a real event are ejecta and flame, which are not modelled, so it under-predicts and can never clear a design.

One seeded cell, stepped through time by the Rust kernel, and watched the way an investigator watches a test: the cell map animating through the event, the fire map showing every path the event could take and the ones it took, the propagation timeline, which heat path — hot gas, ejecta, flame, conduction, radiation — carried the energy, and a closed energy ledger for trusting the run. Root cause is done the way a fire expert does it: each declared path is removed and the event re-run, and the verdict is necessary-in-model, contributing-in-model or not-decisive-in-model — never "the cause was", because physical evidence outranks every model. Every physical input is declared on screen with its meaning beside it; a catalogue cell name never chooses a runaway parameter. The event verdicts are propagation-observed-in-model or no-propagation-in-model-unproven — the screen has no way to say "safe".

A conditional pressure-relief free-area range from declared gas volume, release time, allowable gauge pressure, gas temperature and opening coefficient. A separate supplier-backed layer selects multiple vent units when required, checks market compatibility and enclosure fit, and returns provisional coordinates only on human-screened discharge faces.

Abuse data as governed evidence: ARC/abuse tables ingested with units, SOC, cell-state binding, uncertainty and rights sealed under one checksum; SOC-dependent source terms calibrated from them with provenance and conservative extrapolation warnings; the same screening kernel re-run at rung R1; chronology with energy accounting; one-change counterfactuals; and a vent-network class estimate that exists only behind pressure/flow evidence.

Explore how a pack performs over a duty cycle.

Compile typed blocks into one executable equation graph, solve finite algebraic feedback loops, recommend an explicit or small-graph implicit method from declared time scales, and integrate continuous states with adaptive error control and exact event boundaries.

A calm visual canvas over the Rust equation graph: guided, manual and automatic-draft modes; exact typed ports; deterministic graph import/export; live trace playback; and reviewable assistant recommendations and repairs that a named human must approve.

Connect blocks, save reusable components and export model files.

The parameterised model: OCV + R0 + RC branches with Arrhenius dependence, reversible entropic heat, a per-module thermal network with an ε-NTU coolant stream, and calendar plus cycle aging. Every coefficient is exposed.

The electrochemical rung: solid-state diffusion resolved inside each electrode particle (finite volumes, implicit, mass-conservative), Butler–Volmer kinetics, and optionally the SPMe electrolyte field — with SEI growth and a plating-risk indicator behind explicit capability levels, and conservation ledgers that close to round-off every run.

Explore the pack, controller and thermal system together.

Export electrical and rotational signals for saved-result animation in Blender. Live coupling is not included.

How the machine charges: on-board charger or not, connector and comms per market, charge time with the CV tail, and the depot-vs-opportunity decision.

V2L, V2H, V2G and V2V as a design policy: the parts each adds, the export budget after reserve, the wear floor that decides whether V2G pays, and the certification it drags in.

The charger CONNECTED to the battery: every second of the session is a negotiation between the charger's power, the connector's current ceiling, the pack's own charge acceptance and the terminal voltage on the cell's actual curve — fitted cells believed automatically. CC and CV fall out of the physics instead of a fixed knee, the winter charge floor and the heater branch behave exactly as the mission does, and the session names WHAT DECIDED THE CHARGE TIME with each limit's share. The AC side's closed-form estimate is cross-checked against this session by test.

Build and run test sequences in the simulation protocol bench.

One screen over one job: upload a cycler trace, and the mission model's class estimates become YOUR cell's measured curve — OCV(SoC) from the rest points, resistance by closed-form least squares under load, thermal conductance when the data can carry it. The fit is deterministic and auditable (no optimizer, no seed), judges itself against the whole trace in millivolts, and names where measurement ends and interpolation begins. Applied to a cell, the fit rides through the simulation, the verification bench and the generated C with its dataset provenance attached.

Guide parameter identification using imported measurement data.

The source/staged runner CLI importer normalizes an exactly mapped delimited or columnar-JSON trace into a versioned, checksummed canonical calibration dataset. The CLI and authenticated local API fit canonical datasets to allowlisted existing model parameters with bounded and fully counted optimizer work.

The source/staged runner CLI and authenticated local API build a versioned staged plan from separate canonical calibration and validation trials, execute only the plan's active parameter groups under exact work limits, and return diagnostic candidate parameters plus fail-closed adopted parameters with fixed full-rate holdout evidence.

Bring the cycler’s real recording next to the simulated run: one frozen plan aligns the clocks, resamples onto the instrument’s samples, and reports residuals, energy and event-time errors against bars that never bend — with every refusal named.

One vocabulary and three value layers above it — instance, experiment, runtime override — resolved with every effective value’s origin and shadowing chain shown; variants as checksummed deltas, batches counted before they run, and evidence that goes visibly stale when anything upstream changes.

Run a neighborhood study to explore nearby pack choices.

Every cell against every energy target, each one fully worked and then ranked by cost per kWh delivered, range, mass or density — across all your cores.

External high-fidelity results imported under a neutral, sealed study contract — solver identity, attribution and rights inside one checksum — and reduced models governed above them: declared train/validation/holdout partitions, explicit lineage, coverage maps, holdout-gated error reports, out-of-domain blocks and alarms, revalidation triggers, and runtime claims only from repeated governed measurements.

The mission model — OCV(SoC), terminal solve with the delivery limit, voltage-cutoff hold, CV taper, winter charge inhibit, coulomb counting, lumped thermal — generated as portable C99 for your BMS or system controller. The ABI is reentrant and fixed-step (bd_init / bd_step / bd_reset / bd_terminate), static memory only, every loop bounded, no mutable globals; the generated header carries the same honesty the browser shows: class-estimate OCV and published DCIR, to be replaced with your measured data.

Generate a reproducible FMI 2.0 source-FMU build kit containing the model description, complete C stepping source and build instructions. The release pipeline separately compiles and packages the downloadable Linux/Windows .fmu. Its FMI 2.0 interface targets import pipelines used by Twin Builder, Simulink, GT-SUITE and Dymola; proprietary host acceptance is recorded separately per product and version.

FMI 3.0 beside the preserved FMI 2.0 chain: a Co-Simulation source kit whose model core is sliced verbatim from the FMI 2 generator (tunable parameters, no invented events or arrays), a reference host proving the lifecycle, and bounded third-party FMU inspection whose trust model never executes an imported binary.

Execute the versioned software model against independent numerical acceptance ranges and deterministic test vectors. Each run checks model/checksum/solver identity, units, limits and exact repeatability instead of merely replaying a graph.

Define and validate the fixed sample period, channel units and ranges, safe output values, overrun action and required sensor/communication/power fault tests. It evaluates measured target evidence but stays Unproven when no real hardware run is supplied.

The whole designer as one JSON call, plus an MCP server so Claude or any agent can size packs, run missions and compare cells by calling the real modules — and review a design the way an engineer would, with every check read into one list ordered so that what could hurt someone comes before what costs money.

Choose tools and keep your designs in your browser.

Cycle-based cost of ownership, cost per kWh delivered over life, sensitivity to price and cycle life, and the CO₂ payback point.

Fifteen modules answer fifteen questions, each in its own shape. This reads all of them into one prioritised list — safety before cost, failures before warnings, the same problem found twice merged rather than repeated — then says what the tool is still guessing and what it did not check at all.

The whole footprint by phase — cells, conductors, enclosure, use-phase losses and recycling recovery — each carrying its own data quality, because they differ by more than an order of magnitude in how well they are known. It answers the question worth asking first: the cells are around 95% of what it costs to build a pack, so chemistry and cell count move the footprint and busbar optimisation does not. It also picks the right comparison for the energy delivered, which for a vehicle is the fuel it replaces and NOT a grid factor.

What the pack is worth at each point in its life, in the place it is actually standing. Stages are nodes and moving between them is an edge that costs money, takes time, loses some of what you had, and is sometimes not allowed. Regulation is carried as gates rather than averaged into a price — landfill is not expensive in the EU, it is prohibited — and every route reports what stands in the way, not just what it pays.