paging dr. frink

Part II (of three) from Kevin Hall on modern sailing electronics and software…
It was impossible to predict where the leaps would come from:  In 2023, a fifteen year-old girl in Portugal singlehandedly disrupted the way we do downwind laylines in big waves. Gybing on a wave is so important where she grew up, that predicting and communicating which wave to gybe on while everyone kept hiking until that wave was under the boat, was key. She worked out a way to…well, just look at the code yourself.
We now run a live in-memory database looper with late materialization of data and cache-aware structures that allow inertial models to self-tune based on the conditions and on inputs that we can either pull from the experience of the community, or play with ourselves. In post processing, we focus a lot on tuning the live onboard tuner, and on ways to give this more feel and simpler inputs. Also: you’d be amazed how much you can actually use video to tune the system too, once you teach the footage the hang of it.
This is what our software does upwind for us now: it waits until we settle out of a tack, and then creates a super-smart, right-here-right-now performance “baseline”. It’s not that sensitive to calibrations, because we’ve taught it a bunch of ways to look at change rates in data and learn which ones represent the boat’s performance changing for better or worse, and which ones represent the conditions changing. A few slider inputs to match some graphic representations and voila! Instant perfect “daily calibrations”. It’s amazing what you can do with sail code, runner tension, mainsheet tension, traveler position, heel, trim, and rudder angle but no windgear! Ongoing old-way cal improvements happen almost organically, while on race day it doesn’t matter much how bad the cals are because the wind triangle solution isn’t God anymore.
The old ways, combined with onboard live-VPP outputs, give us all these expectation curves, and the computing horsepower onboard now allows us to teach the system to feel what’s going on and decide why? better, and more importantly, what to output as a result. We’ve taught it that shear looks like this, bad waves look like that. Similarly, we’ve taught it that telling us our mode is better right after we start pinching, but while the big, dead, lead bulb is still doing it’s inertial thing, is really bad for winning races. Similarly, we’ve taught it that we have to twist the sails first, then put the bow down just a little and hike extra hard, then wait for the speed to come so the leeway decreases, and then – and only then! – should it tell us whether the new mode is better or worse.
All of the sailors on the rail have an easy way to input their take on our mode. We figured, since they see so much and are such great sailors, we should add their intuition that we’re going well despite a bad lane, or that we’re going poorly despite a good lane, to the dataset. This gives us so much interesting stuff to look at in post analysis. In the past we had noticed that the tactician thought we “must have been slow” on days with bad results, and the speed team knew we were “always in bad lanes”. But the guys outside the cockpit usually knew the truth and were forever deciding whether they’d still have a job if they spoke up. (Some teams enter rail performance data anonymously, and other teams have contests to see which person’s inputs best match the targets at the middle and the end of the season).
Our displays are very graphics-oriented. We hung around in an airplane cockpit for a while and noticed how much intuitive information is communicated by the little picture of the airplane on the instrument panel – pitch, roll, and target pitch and roll in a glance. So we tinkered for awhile and realized that we could put a bunch of stuff in a picture without things getting confusing, and iterated until it was awesome.
Laylines are really cool now. We have graphics which represent the risk-reward surface of a layline call. As the layline approaches, the whole boat can see the odds click upwards of tacking and laying the mark. Based on historical COGs on the beat, we predict our final track to the mark using the relationship between the wind data before a tack and COG data after a tack, then fold in things like time on the new layline, wind trend before tack, and a couple other custom things depending on the boat. A good leeway model in big waves, which takes into account the speed going into the tack to be able to fold in the leeway out of the tack, for example. This allows the entire team to be in the loop using their own assessments, and to prioritize hiking. If there’s only a 5% chance that opposite COG will lay the mark tacking under the starboard tack layline train, we don’t expect the tactician to hallucinate a future where we make it – we all know with a glance that we’re ducking, winding up then tacking wide. And vice versa. If the call was beyond the group and a shift starts coming through quickly so the picture says we have a 60% chance of making it, and that chance is climbing –- and THAT is the beauty here, is the rate of change of the chance of making it tells us so much – we might even start a very gentle main windup just in case the software and the tactician and the tiller all get onto the same page.
Downwind, those same graphics include colored zones for whether to drop the staysail before the last gybe or not – it’s amazing how much quieter the boat is around this now! – and since we’ve input our best and average drop times by condition into the database, the whole boat can cue off the exact same info. This allows people to keep moving their weight for speed and still manage getting to their next position for the maneuver themselves. Occasionally, we listen to an old-school tactician who misses all the power he used to feel in getting ten people to run around on his countdown command, but he usually simmers down by the second beer that night. We also noticed that boxing him in the corner of the bar between two grinders seems to help soften his clinginess to the old ways.
Our downwind performance display leverages processing, nimble live database power, modeling, and new graphics to output an “Apparent Wind Snowball”. The goal is to get all the info that is communicated by tiller load, and by that load’s relationship to the deck pushing or dropping away from the helmsperson’s feet, as well as a model of the vital info communicated by the feel in the trimmer’s sheet, and also the historical BSP and TWA targets, onto one display that everyone can see. Instead of focusing on BSP and target TWA, we focus on Apparent Wind Angle, Heel, and rate of change of Apparent Wind Speed. The favorable BSP and TWA pairings  are an output of a sensitivity to today’s live-database modeled ideal AWA, AWS, and heel, not inputs that we chase. And high rudder angles with high heel? Easy to depict in the same graphics. Some boats even have their displays flash sometimes. One boat uses sound cues, too.
We display a circle that shrinks and moves toward the soak side of the boat as the actual sheet goes soft, and grows and moves toward the hot side of the boat as the AWA goes narrow and the AWS increases. The same circle has a line through it which, if lined up with the actual horizon, tells us we’re right on target heel. A few teams use colors to represent 1 minute averages compared to targets (blue = too flat, red too heeled). The living database knows the apparent wind/true wind optimums that pair with the best actual recently produced inertial-massaged VMG, (that bulb means we have to be so careful with what we decide is working!) and the graphics offer a little target ball to sorta chase to help the sailors achieve that optimum. (Note that the human “neural net/mindbody/athlete” with a bit of tiller time is amazingly good at this).