Using AI for flight planning

[ilovecornfields]

So, my son is doing a generative AI camp this year and they have to come up with a project. I was wondering if anyone has used AI for flight planning. I tried asking Google Bard if the local airport would be below minimums tomorrow and got a garbage answer including how the minimums for a visual approach are one mile but three miles for an instrument approach. 
 

Seems like it might be pretty easy to check all of the METAR. terminal forecasts and MOS along the way, map out your possible diversion airports and give you some sort of probability that the weather would be VFR, MVFR, IFR or LIFR as well as computing the risk using one of those tools like the MMOPA-FRAT.
 

Has anyone tried this? Any thoughts? 

[PT20J]

I like the idea. I don't know if any of the current large language models are up to the task of predicting future events, however, as they seem best suited to answering questions by mining existing knowledge. But, they probably couldn't be any less accurate than the current methods.

[N201MKTurbo]

I just got an email from Visual Studio urging me to download their AI package. It says all I have to do is write the comment for a method and it will write the code.

I hate writing comments…

[gdwinc]

I am by no means an expert, and am not sure about Google Bard, but the ChatGPT AI was modeled on information through September 2021 and is thus unusable for real-time data.

[201er]

38 minutes ago, ilovecornfields said:

So, my son is doing a generative AI camp this year and they have to come up with a project. I was wondering if anyone has used AI for flight planning. I tried asking Google Bard if the local airport would be below minimums tomorrow and got a garbage answer including how the minimums for a visual approach are one mile but three miles for an instrument approach. 
 

Seems like it might be pretty easy to check all of the METAR. terminal forecasts and MOS along the way, map out your possible diversion airports and give you some sort of probability that the weather would be VFR, MVFR, IFR or LIFR as well as computing the risk using one of those tools like the MMOPA-FRAT.
 

Has anyone tried this? Any thoughts? 

[PeteMc]

So I wonder if you could ask AI how long it would take to learn to do flight planning and if you'd get a realistic answer? 

I like the idea and I'm wondering if it would ultimately be AI doing the flight planning or if AI would learn to create better algorithms and then just write a better App for us to use.  Is the current thought of those close to AI that everything will be AI in the future?  Or is AI going to be used to create and occasionally Update Apps that we'd still use day to day.  And if it's all AI, then when will the first Desktop AI computer be available??? 

And was the FAA calling the future ATC plan to eliminate the need for airways something like "Open Skies" or something similar?  When can we put AI on that project?!

[ilovecornfields]

1 hour ago, PT20J said:

I like the idea. I don't know if any of the current large language models are up to the task of predicting future events, however, as they seem best suited to answering questions by mining existing knowledge. But, they probably couldn't be any less accurate than the current methods.

Good points. The more I think about it the more clear it is that I don’t understand AI. My son said they just had a “machine learning” segment that involved linear algebra. I said “you mean like matrices, eigenfunctions and stuff like that?” He said yes so now we’re both confused. 
 

It’ll be interesting to see where this stuff goes 

[Fly Boomer]

With all the goofball answers I have seen generated so far, I don't think I'll be trusting anybody's program in my lifetime.  Plus, I worked in IT for 40 years, and I have seen the underbelly.  It ain't pretty.

[Vance Harral]

4 hours ago, gdwinc said:

the ChatGPT AI was modeled on information through September 2021 and is thus unusable for real-time data.

That's not correct.  Your run-of-the mill "deep learning" neural net that people are now ubiquitously calling "AI", does indeed need a set of old training data to produce a trained model.  But then new data is mapped against that trained model - that's the whole point of it.  If the nature of the new data is essentially the same as the data used to train the model, you most certainly can make real time predictions.  The nature of weather today isn't meaningfully different than it was a couple of years ago, ergo weather data from a couple of years ago can be used to produce a model of what the weather will be tomorrow.  This is, in fact, how pretty much all modern weather prediction models work.  The technology just wasn't ubiquitously known to the public as "AI" until recently.

It would indeed be cool to train a machine learning model specifically against aviation go/no-go criteria, rather than just general weather parameters.  But I'm not sure that would produce any kind of revolutionary change in flight planning or go/no-go decisions. 

[ilovecornfields]

5 minutes ago, Vance Harral said:

It would indeed be cool to train a machine learning model specifically against aviation go/no-go criteria, rather than just general weather parameters.  But I'm not sure that would produce any kind of revolutionary change in flight planning or go/no-go decisions. 

That’s what I was thinking. I have a friend who’s a relatively new VFR pilot and lives near the coast, but sometimes the TAFs are for places more inland so they don’t predict the marine layer very well. The MOS is better, but not great. Seems an AI could constantly compare the actual weather against the predictions and come up with a more accurate hyperlocal model.

[RoundTwo]

5 hours ago, N201MKTurbo said:

I just got an email from Visual Studio urging me to download their AI package. It says all I have to do is write the comment for a method and it will write the code.

I hate writing comments…

Yeah, that will be some solid code. That will require some extremely extensive unit testing. Glad I retired while coding skills were still valuable.

[1980Mooney]

3 hours ago, gdwinc said:

I am by no means an expert, and am not sure about Google Bard, but the ChatGPT AI was modeled on information through September 2021 and is thus unusable for real-time data.

It seems that you are mixing 

1. the breadth of available data sources that can be accessed in the design of the model
2. with the real time data that is processed in the model.

The model is based upon sites and sources of data that existed as of Sept 2021.  As more sources or improved sources become available then the model is modified/improved such as from https://home.pivotalweather.com/our-company

The AI model will then source real-time data - weather, airport conditions, etc.

 

 

[hais]

I think a reasoning system is more suitable for this task than a generative one. There's data available, and one can write rules to generate the answer. 

While that is outside the scope of your son's BootCamp, it is AI for your idea. 

One idea for a generative one close to flight planning would be itinerary planning: suppose you are looking for where to go...

[hais]

EZWxBrief is close to that - since risks are already generated, it would be easier to write for/no-go rules or get the system to explore alternate paths. I'll let @Scott Dennstaedt, PhD comment whether there's a feasible project there for a bootcamp...

[Scott Dennstaedt, PhD]

2 minutes ago, hais said:

EZWxBrief is close to that - since risks are already generated, it would be easier to write for/no-go rules or get the system to explore alternate paths. I'll let @Scott Dennstaedt, PhD comment whether there's a feasible project there for a bootcamp...

Yes, my PhD dissertation was on how to quantify risk using personal weather minimums that include ceiling, visibility, thunderstorms, icing, turbulence and wind.  And my progressive web app, EZWxBrief, was the result of that 3.5 years of research.  The key is to not use data that is based strictly at airports such as MOS or TAFs.  Those are not going to be relevant for a route.  High resolution numerical weather forecasts are the best to use in this case.  

[N201MKTurbo]

1 hour ago, RoundTwo said:

Yeah, that will be some solid code. That will require some extremely extensive unit testing. Glad I retired while coding skills were still valuable.

I keep trying to retire, but the world seems to have different plans for me. 

One of these days I need to tell the world to forget about it. I could think of some more colorful words….

[ilovecornfields]

2 hours ago, Scott Dennstaedt, PhD said:

Yes, my PhD dissertation was on how to quantify risk using personal weather minimums that include ceiling, visibility, thunderstorms, icing, turbulence and wind.  And my progressive web app, EZWxBrief, was the result of that 3.5 years of research.  The key is to not use data that is based strictly at airports such as MOS or TAFs.  Those are not going to be relevant for a route.  High resolution numerical weather forecasts are the best to use in this case.  

I think what you focused on is more relevant to the overall risk of the flight which is very useful, but what I was interested in was just the last 15 minutes. How likely is it that the field is VFR (or above your minimums.) I think this is very different than what you did because it would be hyperlocal and seems much less complex than the overall weather throughout the route of flight.

My programming experience is limited to using Basic on my Apple IIe and trying to program an Arduino, but it seems that with the technology available now it would be relatively simple (for those with the knowledge and skills) to make a program that predicts the terminal conditions, checks to see the predicted vs. actual and then modifies the model to try to improve accuracy. It seems that if you did this ever hour for a number of months to years the local model might get pretty good. If 

Again, not my field so maybe it’s more difficult than I’m imagining but at least from what my son is telling me the machine learning technology is already there. 

[Scott Dennstaedt, PhD]

12 hours ago, ilovecornfields said:

I think what you focused on is more relevant to the overall risk of the flight which is very useful, but what I was interested in was just the last 15 minutes. How likely is it that the field is VFR (or above your minimums.) I think this is very different than what you did because it would be hyperlocal and seems much less complex than the overall weather throughout the route of flight.

But it seems that with the technology available now it would be relatively simple (for those with the knowledge and skills) to make a program that predicts the terminal conditions, checks to see the predicted vs. actual and then modifies the model to try to improve accuracy. It seems that if you did this ever hour for a number of months to years the local model might get pretty good. 

I guess I don't understand the statement if the "field is VFR" and "last 15 minutes."  You can travel a good distance in 15 minutes.   Are you just looking for the ceiling/vis at the destination airport or along the last 30 miles or so of the route that includes the destination? 

MOS uses geoclimatic data for the airport which means it uses a historical record of observations along with a model forecast. Specific equations are written for each airport for ceiling, vis, wind, temperature, etc. When I worked at the NWS many decades ago, I helped to develop the NGM MOS...which has since been retired.  MOS has been around for over 50 years.  

[ilovecornfields]

8 minutes ago, Scott Dennstaedt, PhD said:

I guess I don't understand the statement if the "field is VFR" and "last 15 minutes."  You can travel a good distance in 15 minutes.   Are you just looking for the ceiling/vis at the destination airport or along the last 30 miles or so of the route that includes the destination? 

MOS uses geoclimatic data for the airport which means it uses a historical record of observations along with a model forecast. Specific equations are written for each airport for ceiling, vis, wind, temperature, etc. When I worked at the NWS many decades ago, I helped to develop the NGM MOS...which has since been retired.  MOS has been around for over 50 years.  

I apologize if I’m being unclear. I have very limited knowledge of both weather and AI so I’m sure both my thought process and communication are somewhat flawed.

What I was imagining was something like a RAIM prediction tool that would constantly update and predict the weather at the airport when you arrive. It sounds like MOS does this to some degree but I was imagining a more dynamic model that was constantly updating and would flag if there was a significant change (like no longe lr VFR at destination).

For example, once I flew to Arcata, CA (KACV) which is right on the coast. It was completely clear when I left and the TAF and MOS both predicted IFR conditions in the evening. Within a couple of hours of departure it was LIFR as the marine layer moved in much earlier than predicted and I barely made it in.

I was wondering if some of these machine learning models were able to determine in real-time when the actual weather deviates from the predicted and provide a more accurate prediction both in the short term and in the future. Maybe MOS does this already. I know most of us do this already when we’re flying and we notice the actual weather is different than predicted we start to doubt the forecast and behave more conservatively. 

Probably make less of a difference for those who are instrument rated but for VFR pilots it might be nice to have an another tool to let you know when you should be more cautious or expect to divert to your alternate. My hangar neighbor regularly has to retrieve her plane from an airport on the other side of the hills because she’s not instrument rated and can’t get in when there’s a marine layer. If she had a better way of knowing beforehand when the field was going to go IFR it might help her with her flight planning and airplane relocation.

[Scott Dennstaedt, PhD]

 

14 minutes ago, ilovecornfields said:

It sounds like MOS does this to some degree but I was imagining a more dynamic model that was constantly updating and would flag if there was a significant change (like no longer VFR at destination).

What is your source of MOS?  In other words, where do you get your MOS forecast?

Understand that TAFs are amended as conditions change since they are monitored in real-time by forecasters.    

[ilovecornfields]

2 minutes ago, Scott Dennstaedt, PhD said:

 

What is your source of MOS?  In other words, where do you get your MOS forecast?

Understand that TAFs are amended as conditions change since they are monitored in real-time by forecasters.    

I use ForeFlight. I thought the TAF was issued four times a day and only updated if there was a big change which required a human to notice the change, care about the change and then issue a new forecast.  I guess I haven’t noticed a lot of forecasts getting updated in time for it to be useful for me right before departure.

[Scott Dennstaedt, PhD]

4 hours ago, ilovecornfields said:

I use ForeFlight. I thought the TAF was issued four times a day and only updated if there was a big change which required a human to notice the change, care about the change and then issue a new forecast.  I guess I haven’t noticed a lot of forecasts getting updated in time for it to be useful for me right before departure.

TAFs are routinely issued four times a day, but for airports in high-impact airspace such as NY, Atl and Chicago, they will issue TAFs more frequently (e.g., every two or three hours). The TAFs are issued by meteorologists at your local weather forecast offices.  Those forecasters have a program called AvnFPS which is a "monitoring" system which I discuss in this post.  Below is one from the Greenville-Spartanburg WFO when I was there visiting a few years back. This allows them to constantly monitor the TAFs and amend accordingly when specific criteria are met.  

When I worked for FF, I helped them add in the LAMP/GFS MOS.  They let me go about 2.5 years later.  I don't use FF now, but I understand they changed from using the LAMP/GFS MOS I recommended to using NBM which in my professional opinion was a huge mistake.  LAMP has been around for a couple of decades and NBM is relatively new and still has some growing pains.

LAMP is a forecast that goes out to 38 hours for ceiling and vis and it is updated hourly.  Essentially it uses the GFS MOS forecast that is issued every 6 hours and is designed to ingest each hour new observational data. That means if the weather is deviating from the original GFS forecast, LAMP will adjust it accordingly and issue an updated forecast.  It also looks at the latest radar and lightning data and also melds in the HRRR model forecasts.  Of course, all of this is specific to the airport.  There are about 700 airports with a TAF and about 2200 airports with a LAMP forecast.  

 

Edited June 15, 2023 by Scott Dennstaedt, PhD

[jaylw314]

On 6/14/2023 at 12:39 PM, ilovecornfields said:

Good points. The more I think about it the more clear it is that I don’t understand AI. My son said they just had a “machine learning” segment that involved linear algebra. I said “you mean like matrices, eigenfunctions and stuff like that?” He said yes so now we’re both confused. 
 

It’ll be interesting to see where this stuff goes 

If you want to go down the math rabbit hole, this guys does an excellent series of videos on the math part of machine learning:

 

[Fly Boomer]

On 6/14/2023 at 7:02 PM, 1980Mooney said:

Really?  I don't know how old you are but any conclusion you draw like above is from "looking in the rearview mirror".  And since you have worked in IT you know, like Moore's Law, the capability and complexity will be growing exponentially.  Unless you are planning on passing in a few years I think you will be seeing it.

https://variety.com/2023/music/news/paul-mccartney-ai-recreate-john-lennon-voice-last-beatles-song-1235642277/

https://www.bbc.com/news/entertainment-arts-65881813

 

Fly Boomer’s First Law:  Nothing happens as fast as you think it’s going to.

[Fly Boomer]

1 hour ago, 1980Mooney said:

Apparently they have already been testing driverless semis on the interstates around Houston for some time. I read that Volvo Trucks will have trucks in testing here soon. No-one really even notices. They plan to be commercial without safety drivers by the end of 2024. 

Like some of the gadgets in airplanes (both GA and Airbus come to mind), everything is fine until it isn't.  Heard a bunch of people who had bad experiences with Tesla and other cars a few days ago.  Many said "never again".  One trucker said his truck had locked up all 18 wheels because it saw the shadow under an overpass.  No other cars on the road.  He too said "never again".  There seems to be a prevailing sentiment that a certain amount of death and destruction is just the price we must be prepared to pay for technological advancement.