Protect the constraint.
Measure blocking, starvation and recovery before buying speed. A larger buffer buys time but consumes space and inventory.
VERSION 02 / EV MANUFACTURING SYSTEMS
One vehicle.
Three production streams.
One connected system.
An engineering study by Sachin Cherukuri. Follow the material, inspect the release gates, and see how one station’s problem becomes the whole plant’s problem.
Follow the 3D buildRead the 48 page engineering reportProcess architecture · automation decisions · constraint recoveryEV / SYSTEM STUDYCONCEPT VEHICLE / 3D STUDY
100,000vehicles / year · target
60 kWhpack · design assumption
75 groupsstation-level concept register
18 sourcesevidence and access register
01 / SEE THE COMPLETE SYSTEM
The factory, connected.
Start at marriage, where body, battery and drive must arrive together. Select any operation to trace what enters, what changes, and what allows it to leave.
Part flow is shown by directed connections. This is a process topology, not a scaled factory layout. Some assembly work occurs before marriage.
INSIDE THE SELECTED PROCESS
WHERE AI COULD HELP
02 / FOLLOW A FAILURE THROUGH THE LINE
What happens next?
Illustrative causal sequence. The slider is not elapsed time and does not calculate queue quantities. Use the production experiment for numerical scenarios.
DETECT
ASSIGN
RECOVER AND VERIFY
03 / MAKE THE ENGINEERING TRADE OFF
A faster station.
A better factory?
Only if the improvement reaches accepted output. Move the pack or test cycle below and watch the nominal constraint change.
Assumed selective automation cyclesVertical line: 126 s demand takt
Test buffers, downtime and rejectsIdeal bound = 50,400 seconds / slowest cycle. This simplified balance view excludes flow losses and does not update the separate simulation.
Keep the release gate.
A prediction can prioritise inspection. It cannot replace unvalidated test coverage or turn missing evidence into a pass.
Design the recovery.
Count interventions, concurrent calls and travel. Robot count alone does not establish the number of people needed.
PROJECT CHARTER / SYSTEM OPTIMISATION
Minimise cost per good vehicle.
Reach stable output sooner.
Improve the whole system while preserving product quality, required controls and recoverability. The objective is total labour, material and time loss per accepted vehicle, including maintenance and exceptions.
126 sdemand takt at 400 vehicles/day
14 hscheduled production / day
3 streamsbody, battery and drive converge
Sheet ≠ foil. Body panels use suitable metal sheet or coil. Battery collectors use thin foil. The body design may combine materials and joining methods.
Cycle ≠ lead time. A faster robot does not necessarily reduce waiting, long process residence, rework or time to stable production.
Automation ≠ AI. Identification, interlocks and machine sequences often use deterministic controls. Learning is proposed only for a defined prediction or perception problem.
Independent concept study, not years of measured factory operation. Research findings, design proposals and synthetic model results are identified separately. Equipment-level design, supplier trials and plant validation remain required.
THE FACTORY JOURNEY / OPERATION VIEW01 / 14
AI CANDIDATE / HUMAN-VERIFIED
AUTOMATION APPROACH
CHALLENGE
PROPOSED RESPONSE
Control checks & human responsibility
0%
ProcessConstraint appearsResponse appliedScroll / drag timeline / use stage selector
HOW THE FACTORY CONNECTS
Three streams. One vehicle.
BODY STREAM
Forming · joining · paint
ENERGY STREAM
Qualified cells · pack · test
DRIVE STREAM
Motor · gearbox · drive test
Vehicle marriage → Assembly → Quality release → Delivery
The walkthrough follows these branches one at a time. The real factory operates them in parallel. Cell manufacturing is an optional integrated workstream; extraction and refining remain supplier interfaces.
UNDERSTAND THE PRODUCT
What the battery
actually does.
A cell stores electrochemical energy. The pack adds mechanical protection, thermal management, monitoring and electrical interfaces. The inverter converts the pack’s DC electrical output into controlled power for the motor.
NEGATIVE ELECTRODEGraphite anode
Lithium ions through electrolyte and separator→ → →
POSITIVE ELECTRODELFP cathode
Electrons through the external circuit: anode → load → cathode
Cell / pack DC⇄Inverter or charger⇄Motor or supply
Simplified lithium-ion operating principle. Charge reverses the ion/electron transfer directions. The separator prevents direct electrode contact while supporting ionic transport; electrons use the external circuit. Concept sizing: 120 cells in series × 3.2 V nominal = 384 V nominal; 156.25 Ah × 384 V = 60 kWh. Neither geometry nor the animated arrows represent qualified cell design or control settings. Process source S02.
75 PROCESS GROUPS / AN ENGINEERING DECISION AT EACH STEP
Inside every station.
Explore the input and output, the real constraint, automation, AI trade-offs and the people needed for normal work and recovery. These are proposed station groups; a production line may split or parallelise them.
Staffing is calculated, not guessed from robot count. Every station example uses assumed workload and a 7-hour productive shift at 80% planning occupancy. Shared-team fractions indicate workload demand, not a final roster. Simultaneous calls, travel, skills, relief, two-person tasks and legal/site requirements need a separate staffing study. Do not add rounded standalone figures to obtain plant headcount.
THREE-STREAM PRODUCTION EXPERIMENT
Does the improvement
survive the whole line?
Body, pack and drive production feed a common marriage station, assembly, end testing and a repair bay. Finite buffers, one interruption at each modelled operation, rejects and a retest outcome are represented. The assumptions are deliberately exposed.
Body streamPack streamDrive stream
all three requiredMarriage
Assembly
End test
Repair / release
Read the model assumptions and boundaries
| Capacity-equivalent operation | Baseline cycle | Selective automation | AI case |
|---|---|---|---|
| Body stream | 150 s | 110 s | 110 s |
| Pack stream | 145 s | 110 s | 110 s |
| Drive stream | 130 s | 105 s | 105 s |
| Marriage | 105 s | 95 s | 95 s |
| Assembly | 140 s | 110 s | 110 s |
| End test | 120 s | 115 s | 115 s |
All numbers are synthetic inputs. Each upstream stream is a capacity-equivalent operation, not a complete physical sub-line. There is one body, one pack and one drive per joined vehicle. The cell factory is excluded from this capacity model; qualified cell supply is assumed. Paint curing and other long residence times are not represented by the body-stream cycle. Model lead time is measured only from marriage entry to final good exit.
One-second steps, a 50,400-second day, empty start, no demand cap or planned breaks within the 14 hours. Each of the six operations has one interruption starting at hour 4 plus 25 minutes × its zero-based index. Completed units block when an output buffer is full. Baseline first-test reject probability is 4%; automation 2%; AI multiplies 2% by one minus the selected relative reduction. A single repair bay takes 600 / 480 / 480 seconds per case, holds six waiting units and has 90% repair success. Failed repairs are scrapped. All common settings and keyed random draws are paired across 10 seeds. Displayed minimum–maximum is run spread, not a statistical confidence interval.
The AI case keeps the automation cycle times. It changes only assumed outage duration and reject probability. No trained AI is running here. Conservation checks reconcile each source stream and joined vehicles. No additional OEE factor is applied. Supplier variation, changeovers, shared maintenance crews, inspection escapes and most detailed process physics remain unmodelled.
INCREMENTAL AI BUSINESS CASE
Capacity is useful.
Cash savings need evidence.
Uses the last completed production run. Compare the AI case with selective automation at the same annual volume. Extra output capacity is shown separately and assigned no sales value. Equipment automation investment needs its own business case.
Illustrative euros, not supplier quotations or a financial forecast. Labour input must already be net of added review/recovery effort; do not count the same hours again as downtime savings. Repair cost excludes labour and lost sales. Annual AI operating cost must include licences, compute, support, monitoring, retraining and extra maintenance. Set cash realisation to zero when time is redeployed without reducing cash spend. The model excludes taxes, financing, residual value, inflation and construction cost.
Equations, sensitivity and what still needs evidence
Matched volume = minimum of annual demand and both cases’ mean daily good output × 250 days. Labour benefit = matched volume × minutes saved / 60 × hourly cost × cash-realisation fraction. Avoided repair cost = matched volume × difference in simulated first-test reject fractions × non-labour cost per event. Net recurring benefit = labour benefit + avoided repair cost − annual AI operating cost.
Simple payback = incremental investment / positive net recurring benefit. NPV = negative initial investment + discounted yearly net recurring benefit over the entered horizon. Fully allocated annual benefit per vehicle also subtracts straight-line capital allocation (investment / years). No launch saving is included in the recurring benefit. The synthetic ten-run results cannot establish real-world parameter uncertainty; validate both favourable and downside inputs with actual observations.
Before approval: collect supplier quotations, task time studies, maintenance/recovery workload, quality costs and proven AI effect sizes from held-out trials. Demonstrate quality equivalence or improvement. Report bottleneck movement, cumulative good output and labour hours per good vehicle alongside the euro result.
ONE PRODUCTION RECORD / ONE ACCOUNTABLE OWNER
Close the communication loop.
Product definitions, production orders, material identity, equipment events and quality decisions share explicit interfaces. AI reads authorised context and proposes actions; deterministic execution and named owners control release.
What is approved?
Product revision, manufacturing bill, process version and change effectivity.
What is required and available?
Demand, purchasing, verified stock, kits and delivery confirmations.
What may happen next?
Routing, serial genealogy, holds, tests, defects and authorised release.
What can execute locally?
Qualified sequences, motion, protective functions and controlled recovery.
What may need attention?
Versioned features, predictions, confidence, drift and grounded evidence.
Who resolves it?
One issue ID, affected parts, owner, containment, repair and verification.
ILLUSTRATIVE HANDOFF WALKTHROUGH
Minimum event contract and AI operating rules
event_id, station_id, event_time, part_id, parent_id, order_id, product_revision, recipe_revision, tool/calibration status, measurement, unit, approved limit, result, quality flag, disposition, owner and acknowledgement. An AI recommendation adds model_version, input timestamp, confidence/abstention, supporting evidence and reviewer.
Use consistent identifiers, clock synchronisation, schema validation, duplicate protection and retained local events. Define which operations can continue offline. Reconcile records before dependent release. An assistant cannot fabricate a missing measurement, clear a quality hold, change a protective function or silently revise a recipe. Start AI in shadow mode, compare with current decisions, pilot bounded assistance, then consider further authority only after validation.
TIME TO STABLE PRODUCTION
Ramp by readiness.
Measure the cost of learning.
Ramp-up is a programme of supplier readiness, commissioning, pilot learning, quality qualification and stable support. A faster nominal cycle does not prove an earlier successful launch.
Define & freeze
Scope, make/buy, material choices, interfaces, capacity, risks and controlled design release.
Accept & integrate
Supplier and site evidence, utilities, qualified protection, fault handling and interface acceptance.
Prove pilot builds
Traceable builds, measurement capability, known-defect trials, trained recovery and controlled changes.
Ramp & sustain
Increase output with adequate quality, supplied material and support; confirm stability over an agreed window.
Can this concept advance a ramp step?
Illustrative decision rules. These controls demonstrate governance, not actual plant readiness.
PROPOSED OUTPUT LADDER
40 · 100 · 200 · 300 · 400
saleable vehicles per day
HOLD — evidence incompleteLaunch schedule, experiments and ramp-cost measurement
Build a dependency schedule covering product freeze, tooling, utilities, installation, acceptance, pilot evidence and release. Use actual supplier lead times and accountable owners before assigning dates. Parallel work helps only where interfaces are sufficiently stable; otherwise it creates rework.
Compare baseline and proposed launch methods using cumulative good builds, launch scrap, engineering/recovery hours, late changes and time to sustained output. Ramp cost = launch labour + trial materials/scrap + supplier support + incremental rework/overtime + justified delay exposure. Keep it separate from steady-state savings.
Priority experiments: blocked logistics route; failed fixture or dispenser; changed material lot; increased defect rate; missing central data; insufficient repair staffing. Freeze the comparison conditions, repeat independent runs and review tails as well as averages. AI acceptance requires later-lot/shift validation and an effective abstention/fallback path.
WORKING MODEL / PACK LINE
What happens when
one station slows down?
A deterministic, finite-buffer simulation runs in your browser. Change the inputs and compare the same 14-hour production day. Every result below is calculated from assumptions, not measured factory data.
Interruption begins at hour 4. Both cases start empty. Proposed case: sealing cycle 145 → 100 s and recovery duration reduced to 25% of the entered interruption.
SIMULATED COMPLETIONS / 14 HOURSCalculated
Baseline—
Proposed balance + recovery—
—baseline ending WIP
—proposed ending WIP
—additional completions
Model rules, limitations & reproducibility
Five serial stations: placement 90 s, connection 100 s, seal 145 / 100 s, test 110 s, release 95 s. All timings are assumptions. Parts enter no earlier than the demand interval and only when station 1 is free. Each station holds a completed part until the next station or its finite queue has space (blocking after service). A stopped seal station retains its part and resumes processing after the outage. Zero buffer still permits direct handoff. No scrap, failures outside the injected outage, retests, shifts, changeovers or shared labour are modelled.
One-second time steps, deterministic inputs, no random seed. The 14-hour run includes start-up and leaves unfinished work in the line. Output counts completed packs, not a validated saleable-vehicle forecast. The model checks: entries = completions + ending work in progress. Scroll movement controls explanatory animation, not elapsed production time.
RESEARCH REGISTER / UPDATED 3 OCTOBER 2026
Every claim needs
the right evidence.
A targeted review of 18 selected primary-source records, including process guides, research papers, OEM reports and technical guidance. Access depth is shown for each. This is not an exhaustive literature review; one guide is retained as a bibliographic pointer because full text was unavailable.
Published findingProposed designAssumed model inputCalculated scenario result
A source supports the stated finding, not all proposed station choices or labour assumptions. OEM reports are examples, not independently verified savings for this factory. AI candidates here are design proposals; no factory-trained model or physical twin is connected.
Download research registerPlanning perspective: connect the equipment, logistics and vehicle interfaces before optimising isolated tasks. BMW virtual factory planning example. This project remains an independent concept, without a connected physical twin.
WHAT THIS PROJECT NOW CONTRIBUTES
A process architecture, 75 station-group decisions, AI trade-offs, workload estimation, a three-stream capacity experiment, an incremental cost model and a controlled release/ramp framework.
WHAT WOULD PROVE THE CASE
Measured process and recovery times, product-specific quality validation, supplier quotations, a staffed pilot, held-out AI evaluation and integrated production evidence. No plant-wide cost or labour reduction is claimed as demonstrated.