Aerospace Learning Curve Application

Apply learning curves to aerospace programs: typical rates by product, lot midpoints, T1 estimates, rate and mix adjustments, and an 85% curve example.

20 min read 5 sections Guide 2 of 6, HPU Forecasting Updated

90% CURVEHPU
Learning curveEach time quantity doubles, hours per unit fall to the rate

New to this topic?

Read these first to get the most out of this guide:

85%
Typical Airframe Rate
Lot MP
Midpoint Method
3 to 5
Shop-Level Curves
±15%
T1 Analogy Accuracy

Typical Learning Rates by Aerospace Product

Learning rates vary significantly across aerospace product types, driven by labor content, process maturity, and the ratio of manual to automated work. Knowing the expected range before you start analysis prevents you from accepting a regression result that is physically unreasonable.

UNITCUM AVG
Unit and cumulative averageTwo curves: unit hours fall faster than the cumulative average that trails themThe Unit and cumulative average guide
Product CategoryTypical RateKey DriversExample Programs
Fighter airframe80 to 84%High manual assembly, complex geometry, extensive rework earlyA fighter center fuselage, a wing assembly
Commercial fuselage82 to 86%Repetitive structure, increasing automation over timeSingle-aisle fuselage sections and panels
Helicopter airframe83 to 87%Composite-intensive, smaller lot sizes limit learningCH-53K cabin, AH-64 fuselage
Aerostructures (nacelles, empennage)84 to 88%Mix of composites and metal, moderate complexityNacelle assemblies, horizontal stabilizer
Satellite bus88 to 93%Low volumes, high test content, design changes between unitsGPS III, communications satellites
Munitions78 to 85%High-volume, repetitive, assembly-line productionGuidance kits, munition airframes

Rate vs. Product vs. Shop

These are program-level rates. Individual shops within a program will vary. Assembly shops tend to learn faster (lower percentage) than test or systems integration shops. Always decompose to the shop level when accuracy matters, which is always on a proposal or EAC.

Lot Midpoint Calculations

Production is purchased in lots, not individual units. To estimate total lot hours, you need the lot midpoint: the unit number whose hours, multiplied by the lot quantity, equals the total lot hours. The lot midpoint is not the arithmetic mean of the first and last unit numbers.

LOT 3MIDPOINT
Lot midpointA lot's average hours sit at its midpoint unit on the learning curve
Lot Midpoint Formula Crawford Model

Lot midpoint unit number M is found by solving:

YM × Q = ∑ Yi for i = first to last unit in the lot

For large lots, an approximation:

M ~ [(Lb+1 − Fb+1) ÷ ((b+1)(L − F))]1/b

Where F = first unit in lot, L = last unit in lot, b = learning exponent

Example: Lot 3 covers units 51 to 100, learning rate 85% (b = −0.2345)

MethodMidpoint UnitNotes
Arithmetic mean75.5Too high: underestimates lot hours
Lot midpoint formula73.1Correct: accounts for learning curvature
Exact summation73.0Computed unit by unit: most precise

The difference between arithmetic mean and true midpoint grows larger for steeper learning rates and earlier lots. Using the wrong midpoint on a 50-unit lot can produce a 2 to 4% error in total lot hours.

T1 Estimation Methods

T1 drives the entire curve. A 10% error in T1 produces a 10% error in every unit estimate. Getting T1 right is the single most impactful activity in learning curve analysis.

PLANACTUAL2423242524
Analog boardA marker writes the hour's count on the whiteboard by hand, where everyone walking past can see it
MethodWhen to UseProcessTypical Accuracy
Bottom-up engineeringNew product, no analogiesSum touch labor by operation, add allowances for rework, support, move±20 to 30%
Analogous scalingSimilar product existsT1new = T1analog × (Weightnew/Weightanalog)CF±15 to 25%
Parametric CERHistorical database availableRegression of T1 against physical parameters (weight, area, part count)±15 to 20%
Regression from actualsProduction underwayPlot actuals on log-log, fit line, extrapolate to X=1±5 to 10%

On active programs, always re-derive T1 from actual data as soon as 5 to 10 units are complete. Early unit actuals are noisy, but by unit 10 the regression T1 is almost always more accurate than any pre-production estimate. Update the T1 at each major milestone or lot delivery to keep forecasts current.

T1 Is Not Unit 1 Actuals

Actual unit 1 hours include first-article inspection overhead, tooling debugging, initial process setup, and engineering support that will not recur. T1 is the theoretical starting point of the repeatable learning trend. Using actual unit 1 hours as T1 will overstate every subsequent unit and inflate the program estimate.

Adjusting for Rate and Mix

The basic learning curve assumes continuous production at a steady rate with a stable product definition. Real programs violate all three assumptions. Rate changes, product mix shifts, and engineering changes all affect the curve.

BREAK
Rate disruptionThe learning curve is running down when a break bumps hours back up, then it recoversThe Rate disruption guide
FactorEffect on CurveHow to Adjust
Rate increaseFlattens learning (overhead spread, workforce dilution)Add rate penalty factor of 2 to 5% per rate doubling
Rate decreaseMay improve unit hours (more time per unit) or worsen (workforce loss)Analyze historically; typically net neutral to slight increase
Product mix changeDifferent variants have different labor contentNormalize to a common configuration using complexity factors
Engineering changesResets learning on affected operationsEstimate learning loss hours as fraction of affected operations × disruption factor
Supplier changeNew parts may not fit as well, causing reworkAdd transition hours for 3 to 6 units after change point

Worked Example: 85% Curve on a 200-Unit Program

FORECAST →
HPU forecastActual hours run down the curve and the forecast extends it to the last unitThe HPU forecast guide
Full Program Example Step by Step

Given: 200-unit fighter wing production program. T1 = 12,000 hours. Learning rate = 85% (b = −0.2345). Four production lots.

LotUnitsLot MidpointHours at MPLot Hours
Lot 11 to 25Unit 10.56,780169,500
Lot 226 to 75Unit 46.84,620231,000
Lot 376 to 150Unit 109.23,780283,500
Lot 4151 to 200Unit 173.53,360168,000

Total program hours: 852,000

Average hours per unit: 4,260 (35.5% of T1)

Unit 200 hours: 3,120 (26.0% of T1)

Key observations: Lot 1 averages 6,780 hours per unit. By Lot 4, the average is 3,360, a 50% reduction. The steepest improvement is in the first 25 units. By Lot 3, the curve is flattening: the difference between Lot 3 and Lot 4 average hours is only 420 hours (11%), compared to a 2,160-hour drop (32%) between Lot 1 and Lot 2.

If this program has multi-shop decomposition (fabrication at 82%, assembly at 86%, test at 91%), the shop-level estimates would differ from this aggregate. The composite program curve may not be exactly 85%: it will appear to shift as the mix of shop hours changes across lots.

The Bottom Line

Applying learning curves to real programs requires more than the basic formula. You must use lot midpoints (not arithmetic means), select the right T1 method for your program phase, adjust for rate and mix changes, and decompose into shop-level curves for accuracy. The worked example shows that an 85% curve on 200 units drops from 12,000 hours at T1 to 3,120 at unit 200, a 74% reduction that must be captured in every proposal, budget, and EAC. Next: HPU Data Collection & Validation, how to gather and clean the production data that drives these curves.

Frequently asked questions

What is a lot midpoint?

A lot midpoint is the single unit number within a production lot whose individual unit hours, when multiplied by the lot quantity, approximate the total lot hours. It is NOT simply the average of the first and last unit numbers. For a lot from unit 51 to unit 100, the lot midpoint is typically around unit 72-74 depending on the learning rate. The lot midpoint shifts toward the higher end because learning is nonlinear: the early units in a lot consume disproportionately more hours than the later ones.

How do you estimate T1 when there is no production data?

Three primary methods: (1) Engineering bottom-up estimate: sum touch labor by operation for a first-unit build, typically accurate to ±20-30%. (2) Analogous program scaling: take a similar product T1 and adjust by weight ratio, complexity factor, or feature count, typically ±15-25%. (3) Parametric models: use cost estimating relationships (CERs) based on physical or performance parameters, typically ±15-20%. All three should be cross-checked when possible.

Why do multi-shop curves matter?

Different shops (e.g., fabrication, assembly, systems installation, test) learn at different rates because they have different labor content and process characteristics. A composite shop may run at 83% while final assembly runs at 87%. Rolling everything into one program-level curve masks these differences and produces inaccurate unit-level forecasts. Breaking the curve into shop-level segments improves accuracy by 5-10% on typical programs.

Share this guide: LinkedIn X Follow SymplProcess on LinkedIn

Take this to a room

The running order

For programme estimators. They should leave using lot midpoints and shop-level curves rather than one programme-level rate.

7 beats · 12 min
  1. 1

    The formula is not the hard part

    Applying learning curves to a real programme takes four things the textbook does not mention.

    • Lot midpoints rather than arithmetic means.
    • The right T1 method for the phase you are in.
    • Adjustments for rate and mix.
    • Shop-level decomposition.

    Ask the room Which of those four do we currently do?

  2. 2

    Lot midpoints, not averages

    A lot's representative unit is not the middle unit by count. It is the midpoint on the curve.

    • Using the arithmetic mean systematically misprices lots.
    • The error grows with lot size, so it is worst on the biggest buys.
    • This is a spreadsheet-level fix with a large dollar effect.
  3. 3

    T1 is not unit one

    Actual unit 1 hours carry costs that never recur, and using them inflates the whole programme.

    • First-article inspection, tooling debug, initial process setup, engineering support.
    • T1 is the theoretical starting point of the repeatable trend.
    • Use analogy or regression; analogy is typically plus or minus 15 per cent.
  4. 4

    Decompose to shop level

    A programme-level rate hides the fact that shops learn at different speeds.

    • Assembly shops learn faster (a lower percentage) than test or systems integration.
    • Three to five shop-level curves is typical.
    • When accuracy matters, decompose. On a proposal or EAC, accuracy always matters.

    Ask the room Do we have one curve or several?

  5. 5

    Adjust for rate and mix

    A curve fitted at one production rate does not transfer unchanged to another.

    • Rate changes shift hours for reasons unrelated to cumulative experience.
    • Mix changes alter the content of a unit, so the trend is measuring something different.
    • Both must be normalised before the curve means anything.
  6. 6

    The worked example

    An 85 per cent curve on a 200-unit programme, which makes the size of this concrete.

    • T1 at 12,000 hours.
    • Unit 200 at 3,120 hours.
    • A 74 per cent reduction, over the life of the programme.
  7. 7

    Why every number depends on this

    Proposal, budget and EAC all rest on the same curve.

    • Get the curve wrong and all three are wrong in the same direction.
    • Get it right and the three finally agree with each other.
    • The discipline is in the four practical steps, not the equation.

    Ask the room Which of our programmes should we decompose to shop level first?