Where does the target heart rate come from for suggested workouts?

The full quote from the support article you linked is:

How you configure your personal heart rate zones does not impact advanced performance metrics such as VO2 max, training status, training load, load focus, or aerobic and anaerobic training effect feedback available on compatible Garmin devices. Overestimating or underestimating your maximum heart rate can, however, affect the reliability of these metrics.

Just a guess, but maybe Garmin determines its own standard zones from your max HR and uses that for both performance metrics and suggested workouts.

I also wouldn't be surprised if it used the measured lactate threshold HR (if such exists) to recommend threshold runs, etc.

Jastitizer said:

Therefore, I would assume the target heart rate for suggested workouts doesn't rely on the zones either. So what does it rely on?

As the article says, the watch metrics, predictions and suggestions don’t use your training zones. 
They only use your “V02 max model” as I call it. Based on your training and performance data, the watch establishes a model of your ventilation needs and the corresponding output metric (pace or power for biking).
In the typical coaching tool box, a pace duration curve and a power duration curve is used to establish targets for intervals of any duration. 
With the watch, all the targets are expressed as a target as a %of VO2Max.

The watch uses HR and HRV to estimate VO2.

Finally there is a model linking V02 and pace.

So you can see there is way for the watch to calculate, for a target interval duration and a target % of VO2 Max what the target HR, pace or power should be.

We don’t know the exact way, and because the watch is using a machine learning chip, it is possible that it is much more complicated than linear or non linear models.

We don’t know either what is data scope that it used for the various models, in terms of time.

From my experience, I’d say the watch lets data drop off after a few weeks, but data as recent as several days is enough or has more weight.

i don’t think the watch is using the threshold values that a user would input manually but tracks it as part of the model in addition to publishing it to the user.

To get good targets, I noticed it is better to perform regularly maximal effort intervals, if you can. Otherwise the watch will have to predict the pace/duration/VO relationship.

@Etupes25 , you wrote "because the watch is using a machine learning chip", are you sure that an actual AI chip / NPU is included in Garmin watches, not just running simpler machine learning algorithms on the CPU? I'm asking because in the watch teardowns I've seen (http://www.f-blog.info/garmin-fenix-7x-solar-teardown-non-destructive/ for example), I've never seen any mention of any machine learning chip (the SoC chip is NXP MRT595SFFOC, and I don't see any machine learning acceleration in that either). And considering power consumption, I'm not sure it would make any sense either.

I got the info from here 7:10

m.youtube.com/watch

FWIW, my cycling power targets for suggested workouts seemed to be derived (spot on 95% for threshold intervals, if I remember correctly) from my 20 min power from the power curve from the last 4 weeks or so. Which was off since I was doing ramp test lately, rather than regular 20 min max effort FTP tests :-)

I assume indeed that the watch is using a power duration curve to match a target to a duration, and looks up the expected HR for that power and time from history to precalculate EPOC, hence TE, training load etc. This is speculation of course. Since there is a machine learning chip in there, the statistical analysis of the interaction of multiple variables, pattern matching and predictive metric models can be more elaborate.

Also there are well established ways to derive the FTP from the power duration curve, as well as VO2 max calculations. Maybe the watch is using some of these, or maybe not :-)

Etupes25About the machine learning chip: In another teardown the chip meantioned in the Youtube video you linked was mentioned to be the NFC chip (and so does one comment on the video you linked). I don't know for sure in either direction, but maybe it's best to take that machine learning chip information with a grain of salt.

Right, the gentleman in the video I linked says "appears to be a machine learning co-processor", while the author in the linke below just doesn't know what the chip is. There is a comment from a user asserting that 100T chips from NXP are NFC. Using Google search, I was not able to find any link from NXP about silkscreen 100TB2n. So it is a mystery.

But first, we know that Garmin implements Firstbeat algorithms using "neural network" for VO2 Max (emphasis added).

"METHOD FOR VO2 ESTIMATION

To construct a more accurate estimate of VO2 when compared with previous HR-based estimates, additional information describing bodily functions was added to the model. Neural networks were used to construct a model that derives VO2 from R-R intervals (time between successive heart beats), using respiration rate and on/off-response information. Schematic illustration of the model is presented in Figure 1. This VO2 estimation method is implemented in the Firstbeat Technologies Ltd. software."

Microsoft Word - VO2_white_paper_Revised_B.doc (firstbeat.com)

Second, we know that HR, HRV processing and EPOC modeling use neural networks. Looking the reference of neural networks at Firstbeat, there is an interesting dissertation on how to use neural networks in a variety of applications for (HR) data series filtering, EPOC modeling (from analysis of exercise intensity and ventilation), respiration frequency and peak detection (see paragraph 6 for these applications of neural networks, but the entire dissertation gets you a good learning/review of neural networks and data time series mathematical processing :-)

https://jyx.jyu.fi/bitstream/handle/123456789/13267/951391707X.pdf

Third, there are suspicions that neural networks are accelerated by a co-processor chip on the watch. Since hardware acceleration is nice but not a must, we can live with the uncertainty about the role of the 100TB2n NXP chip.

In summary, on the one hand, the watch has several models that prepares, analyzes and link the data necessary to compute a target to achieve a certain level of training intensity with respiration and HR, and on the other hand, it uses machine learning.

So it is possible that the calculation of HR, pace or power targets uses machine learning rather that more simple regression/cluster analysis linear or non-linear formulas.

Thanks everyone for the spirited conversation! If nothing else, it instills me with confidence for the insane paces my watch has me going now that I'm using target heart rate.