SpaceX progress towards a 100 launch year

[eeergo]

An exponential increase is the most natural and obvious one. SpaceX can’t obviously flip a switch and change every single bottleneck of flightrate. It’s ludicrous to think so. Exponential increase is from making a whole bunch of incremental improvements that occur roughly stochastically over the year.

And of COURSE a linear function fits a curve if you look at a very short portion of the curve, that’s literally true for basically every smooth (and non-pathological) function!

If linear fits so well, then (if you mean linear increase) tell me what year SpaceX had a negative launch rate? Or (as I assume you mean linear cumulative launches) did they have a 100 year launch rate last year, too?

To make linear work requires just a bunch of overfitting and piecewise assumptions. It’s a crap model and actually over-complicated, requiring a bunch more assumptions.

(A logistical or sigmoidal function I could take more seriously, ie that it would plateau soon… but exponential is just the most simple model.)

Thanks for agreeing with the bolded.

Now, defining what "short" is in this case can obviously be up for debate, but you will also agree it should be short enough that trying to linearly fit segments with such lengths yields a much better R (or χ² if you will) than any other function. That's the case for yearly segments, as demonstrated with any moving window average - which is the data xyv is plotting. If you are just interested in overall multi-year cadence (with some saturation cutoff as you imply by the sigmoid), I agree some kind of exponential may be more representative in spite of periods like 2021. But not for yearly periods. No need to get all nasty with "crap" linear regressions.

Regarding the italicized statement: a negative slope doesn't imply a negative launch rate as long as it isn't too steep and the initial rate isn't close to zero to begin with. It really isn't hard unless you try to make it so.


But you keep ignoring xyv's and Comga's plots. What do you get fitting year-on-year exponentials to that data? You have the numbers... EDIT: Actually never mind, I have done it myself with the 2022 data and standard Excel fitter. Guess I don't really need to report the R value. I don't think we have Comga's data, but I would bet the goodness of fit would be a toss between linear and exponential for that 3-year period too.

[FutureSpaceTourist]

Some background to F9 ops and flight rate:

twitter.com/superclusterhq/status/1630962514430246912

In our latest, author and contributor @dwbwriter sits down for an exclusive interview with @SpaceX VP of Falcon Launch Vehicles, Jon Edwards.

The past, present, and future of Falcon 9:

https://www.supercluster.com/editorial/the-man-who-wrangles-the-workhorse-falcon-9

https://twitter.com/superclusterhq/status/1630964921486671877

Edwards is certain SpaceX can beat their record turnaround time for re-launching a Falcon 9.

@elonmusk is pushing for a bigger goal:

“Why do we have any refurbishment at all? We shouldn't need any refurbishment. You should just land it and fly it again—fix it so it does that.”

[alugobi]

Whatever the shape of the slope, as of today:

[Robotbeat]

An exponential increase is the most natural and obvious one. SpaceX can’t obviously flip a switch and change every single bottleneck of flightrate. It’s ludicrous to think so. Exponential increase is from making a whole bunch of incremental improvements that occur roughly stochastically over the year.

And of COURSE a linear function fits a curve if you look at a very short portion of the curve, that’s literally true for basically every smooth (and non-pathological) function!

If linear fits so well, then (if you mean linear increase) tell me what year SpaceX had a negative launch rate? Or (as I assume you mean linear cumulative launches) did they have a 100 year launch rate last year, too?

To make linear work requires just a bunch of overfitting and piecewise assumptions. It’s a crap model and actually over-complicated, requiring a bunch more assumptions.

(A logistical or sigmoidal function I could take more seriously, ie that it would plateau soon… but exponential is just the most simple model.)

Thanks for agreeing with the bolded.

Now, defining what "short" is in this case can obviously be up for debate, but you will also agree it should be short enough that trying to linearly fit segments with such lengths yields a much better R (or χ² if you will) than any other function. That's the case for yearly segments, as demonstrated with any moving window average - which is the data xyv is plotting. If you are just interested in overall multi-year cadence (with some saturation cutoff as you imply by the sigmoid), I agree some kind of exponential may be more representative in spite of periods like 2021. But not for yearly periods. No need to get all nasty with "crap" linear regressions.

Regarding the italicized statement: a negative slope doesn't imply a negative launch rate as long as it isn't too steep and the initial rate isn't close to zero to begin with. It really isn't hard unless you try to make it so.


But you keep ignoring xyv's and Comga's plots. What do you get fitting year-on-year exponentials to that data? You have the numbers... EDIT: Actually never mind, I have done it myself with the 2022 data and standard Excel fitter. Guess I don't really need to report the R value. I don't think we have Comga's data, but I would bet the goodness of fit would be a toss between linear and exponential for that 3-year period too.

You start the exponential curve earlier, not arbitrarily at Jan 1st, FFS. I’ll do it when I get home.

It’s like y’all have never done proper curve fitting before.

[xyv]

You start the exponential curve earlier, not arbitrarily at Jan 1st, FFS. I’ll do it when I get home.

It’s like y’all have never done proper curve fitting before.

Well at time = 0 (1 Jan) any exponential will be 1 x whatever the multiplier is so some extra fiddling will be required.  As I said, I am not doing any curve fitting just keeping score - are they ahead or behind goal.  And goals (aka management pressure) do have an effect - think fundraising thermometer.  The trend this year does appear to be steeper - maybe after a couple more months if they are still tracking to it I will plot the two years sequentialy.  Right now it appears they indeed have made the rate "tip up".

Back for a quick rethink.  I forgot that to make the x axis read in 2023 dates I have to use Microsoft date format.  So 1 Jan 2023 = 44,927.  I will be interested to see the compensations Robotbeat has to make to fit an exponential.

[eeergo]

You start the exponential curve earlier, not arbitrarily at Jan 1st, FFS. I’ll do it when I get home.

It’s like y’all have never done proper curve fitting before.

Listen man, enouh with the "FFS", "crap", "y'all have never" and other qualifiers ok? Unless you would like to receive the same treatment.

Any date you choose as starting point will be as arbitrary as January 1st - you're just applying priors to the data, which is not blind fitting and reflects some a-priori knowledge of why the data should be like that (i.e. in this case, your personal conviction). With the data xyv has provided, you either use that or arbitrarily truncate the time series.

Either way, we were discussing the best fit to a *yearly* dataset, and you yourself were referring to it - not some other period. The exercise was rethorical: it's visible already by eye an exponential curve does not fit the bill for yearly periods. Indeed, a *logarithmic* curve ironically fits it better! Exponential may be ok for the whole historical launch data, but I doubt it would be significantly better than a linear at this point in time, courtesy of periods like 2021, as shown above.

[Robotbeat]

Every smooth(-ish) function looks linear if you zoom in far enough. That doesn’t make it a good model of how you’d expect growth to work over time. Continuous compounding improvements given you an exponential curve, and that also fits better with SPaceX’s long term launch rate trend.

Getting to 100 launches per year while the previous year had 61 launches implies a big improvement in capability for SpaceX. Is it realistic to expect that to happen overnight on January 1st, 2023? Obviously not. It’ll take a series of improvements over time. Launch rate will be lower in the first half of the year than the second probably. That was true for 2022 as well.

[eeergo]

Are you going to provide those exponential fits or not?

[wannamoonbase]

Every smooth(-ish) function looks linear if you zoom in far enough. That doesn’t make it a good model of how you’d expect growth to work over time. Continuous compounding improvements given you an exponential curve, and that also fits better with SPaceX’s long term launch rate trend.

Getting to 100 launches per year while the previous year had 61 launches implies a big improvement in capability for SpaceX. Is it realistic to expect that to happen overnight on January 1st, 2023? Obviously not. It’ll take a series of improvements over time. Launch rate will be lower in the first half of the year than the second probably. That was true for 2022 as well.

You make some good points.  One of the first times I remember seeing Elon speak about anything was talking about EV efficiency where if you get a 1-2% improvement on different components each year that over time it really adds up.

I think the F9 launch cadence is what to watch, maybe the last 10 or 20 flights is a good average.

At the end of this year they could be hitting a 100 flight per year Candence and still only flight 80-90 flights for the year.

I still think they need another ASDS, given weather and payload delays, as well as maintenance required, a 100 flight calendar year requires a near perfect year with their current equipment.  Possible, but will really push the limits.

[Robotbeat]

The end of 2022 had SpaceX doing higher than 61 flights per year cadence. 61 was the average for the whole year. If spaceX made zero improvements to flight rate over what they did at the tail end of 2022, they’d get about 80-85 launches per year. But why would they stop improving? That would make no sense. Consistent compounding improvement is more likely.

[Robotbeat]

Are you going to provide those exponential fits or not?

Sorry, I was busy last night. I’m on my phone now, which is annoying to do that sort of thing with

[Robotbeat]

But an illustration of flight rate improvement over time while you wait: https://twitter.com/EricWollberg/status/1630992258861334546?s=20

[eeergo]

But an illustration of flight rate improvement over time while you wait: https://twitter.com/EricWollberg/status/1630992258861334546?s=20

Nice animation, but you have Comga's graph to more precisely show this.

Yet you keep moving the goalposts: your critique, and my posterior answers, referred to the annual scorekeeping xyv provided and showed against a linear trend line, to then generously proceed to provide the data he compiled for the graph - not the overall rate you keep referring to.

[Robotbeat]

Sorry, it takes time LOL.

[alugobi]

At the end of this year they could be hitting a 100 flight per year Candence and still only flight 80-90 flights for the year.

I still think they need another ASDS, given weather and payload delays, as well as maintenance required, a 100 flight calendar year requires a near perfect year with their current equipment.  Possible, but will really push the limits.

This is my thinking, too, and the root of my skepticism about making 100.

[Robotbeat]

As they’re already achieving a rate of about 80-90 per year (currently 84/year and will get to 90/year if the Starlink launch in about 5 hours happens… even without perfect weather so far this year), I don’t see any reason why they couldn’t improve the rate by another 10-20% by the end of the year even without another droneship. If droneships are the bottleneck, they can do more RTLS flights and use Vandenberg more, plus a few from Boca Chica.

I just don’t get the argument of “80-90 per year, which is exactly their current rate in the early part of the year, is all they will be able to do for the whole year even as they make improvements!”

I was down at Port Canaveral in January, and when I first arrived there, the droneship had arrived with a booster. I was worried they would have it processed immediately and I’d miss a close-up. But I was wrong! The booster and droneship sat there for a few days. So I know for a fact there’s still some slack in their current droneship operations. 3 days of slack with 14 days average in between catches per droneship means you can increase rate by 20% even without faster tow speeds, etc. And that’s ignoring additional use of RTLS, or very short droneship trips, or Starship launches.

Heck, even with OCISLY totally out of the picture (due to being upgraded), SpaceX achieved 7 launches in the 31 days of December 2022, an annualized launch rate of over 82/year. So with OCISLY now in action, 100 is easily feasible. 

[Robotbeat]

(A droneship going 630km downrange for a typical Starlink launch, if towed at a maximum demonstrated 9 knots—which is high but I think 10-12 knots could be done as the hull speed of the barge might be as much as 17 knots and a European study suggested 12 knots top speed was feasible for autonomous barges like SpaceX was using—takes just 1.6 days to reach its station, so transit time may only be about 3.2 days per launch. 0.8 day at station implies each droneship could in theory turnaround a launch every 4 days. With 5 droneships, that’s over 273 droneship recovered launches per year maximum rate. The European study assumed a 6 day mission time for the droneship approach, giving a still-impressive 180 droneship recoveries per year. A significant amount of ground pad recoveries could help extend this, plus the Starship launches… so there’s margin for slower towing speeds, weather interruptions, etc, and STILL reaching 100 per year. They’re not near the upper limit even for droneship recoveries, so there’s plenty of reason to think they’ll succeed.)

Droneship study by Europe: https://www.researchgate.net/publication/335611322_A_Systematic_Comparison_of_Reusable_First_Stage_Return_Options

[Redclaws]

I do think the argument over the best fit function for the launch rate is pretty silly.  It’s obviously controlled by many, many different overlapping factors, many with large discontinuities, like adding a drone ship or changes in payload availability.  *no* equation should be expected to give much meaningful forward insight unless you can point to a primary limiting process that might reasonably follow that type of equation.  We can fit something retrospectively, sure, but it shouldn’t be expected to have much predictive power going forward.  The increase isn’t actually linear or exponential - it’s governed by a herky-jerky set of complex discontinuous overlapping processes/limitations.  It’s not even just production rate increase, which could maybe be linear or exponential, except we know they’re mostly not rocket production limited.  Or they are *sometimes* and not others.

[Robotbeat]

I agree that expecting ANY simple function to fit all the data perfectly is silly…

BUt.. The overall principle that you should expect launchrate to improve during the year if they’re expecting a launch rate 64% higher than the previous year is completely sensible whereas the assumption of NO additional rate improvement during the year is ridiculous.

[Nomadd]

 Those ASDS's don't cost that much and betting the farm on one not going offline unexpectedly seems kind of short sighted. Having a spare doesn't mean they'd have to have all the tugs and support gear for it.