graham/prop
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Rewrite /about in a less-LLM voice

Browse source 
Drops the formal "What we're trying to do / The approach / The stack"
scaffolding for a shorter, first-person telling: line-of-sight is
wrong, here's what we're doing about it, here's what's in the box,
here's what's next. Adds a highlighted note that the scoring is
hand-calibrated and not updated in real time, with links to the
algorithm page and the submit page so readers can help close the
loop.
This commit is contained in:
Graham McIntire 2026-04-09 13:01:19 -05:00
parent 5724b33e8f
commit 087517252b
1 changed files with 87 additions and 73 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

160
lib/microwaveprop_web/live/about_live.ex
View file

@ -23,59 +23,76 @@ defmodule MicrowavepropWeb.AboutLive do
<section class="space-y-8 text-sm leading-relaxed">
<div class="space-y-3">
<h2 class="text-xl font-bold">What we're trying to do</h2>
<p>
Microwave propagation above 10 GHz is dominated by the lower atmosphere:
humidity gradients, temperature inversions, ducting layers, rain cells,
and hyper-local refractivity structure. At 10, 24, 47, 76, 122, and 241 GHz
these effects decide whether a contact happens at 50 km or 500 km. This project was
started by Jim KM5PO, then picked up by Graham W5ISP.
Line-of-sight is wrong for the microwave bands. Every operator on
902 MHz and up has made contacts over hills, through trees, and 200 km
past the horizon — contacts that shouldn't have worked if LOS were the
whole story. What's actually going on is the lower atmosphere:
temperature inversions, humidity gradients, ducting layers, and rain
cells that bend, trap, and absorb signals in ways that change by the
hour. The window for a good contact is often minutes, not days.
</p>
<p>
Traditionally microwave contacts are ruled by line of sight, but we've
seen that is incorrect. This project is an attempt to build a data-driven
propagation prediction model specifically for the amateur microwave bands,
using hourly numerical weather forecasts, historical contact records, and
eventually calibrated beacon measurements.
Jim KM5PO started this project to try to predict those windows from
weather data. Graham W5ISP picked it up and is iterating on it from
there. The goal: a map that tells you when to get on the air, per
band, based on real atmospheric conditions and what we've learned
from the contacts you and everyone else have already made.
</p>
<div class="rounded-box border border-warning/30 bg-warning/10 p-3">
<p>
<strong>Heads up:</strong>
the scoring you see is hand-calibrated from the data we have so far.
It isn't updated in real time — I (Graham) re-fit the weights
manually when there's enough new data to move the needle. The more
contacts and beacon reports we collect, the better it gets. See the
<.link navigate={~p"/algo"} class="link link-hover font-semibold">algorithm page</.link>
for what's actually under the hood, and please <.link
navigate={~p"/submit"}
class="link link-hover font-semibold"
>submit your contacts</.link>.
</p>
</div>
</div>
<div class="space-y-3">
<h2 class="text-xl font-bold">The approach</h2>
<p>We pair two things that most propagation tools keep separate:</p>
<h2 class="text-xl font-bold">How it works, roughly</h2>
<p>
Two things, paired up. Most propagation tools keep them separate.
</p>
<ol class="list-decimal list-outside pl-5 space-y-2">
<li>
<strong>Atmospheric state</strong> — hourly 3 km HRRR forecasts
(surface fields plus pressure-level profiles), hourly surface
observations from ASOS, 12-hourly upper-air soundings, and gridded
IEMRE reanalysis. From these we derive refractivity profiles,
ducting detection, minimum refractivity gradient, precipitable water,
boundary-layer depth, and a 9-factor composite score for every
0.125° cell on a CONUS grid, for each hour of an 18-hour forecast.
<strong>The weather.</strong>
Hourly 3 km HRRR forecasts give surface fields and pressure-level
profiles. We derive refractivity profiles, minimum refractivity
gradient, ducting detection, boundary-layer depth, and precipitable
water for every grid cell. ASOS surface obs, 12-hourly upper-air
soundings, and gridded IEMRE reanalysis fill in the gaps. A 9-factor
composite score is written out for every 0.125° cell across CONUS,
for each hour of an 18-hour forecast.
</li>
<li>
<strong>Historical contacts</strong> — 58k+ amateur microwave QSOs
tagged with the atmospheric conditions at both ends and along the
path at the exact time they happened. That gives us a ground-truth
dataset we can use to calibrate the scoring weights and, eventually,
train a machine learning model that predicts contact success given
conditions.
<strong>The contacts.</strong>
58k+ amateur microwave QSOs, each tagged with the atmosphere at
both ends of the path (and along it) at the time the contact
happened. That's a ground-truth dataset we can actually fit against
— "when you had this score, how far did the contact go, and did it
happen at all?"
</li>
</ol>
<p>
The current scoring algorithm is a weighted sum of nine factors
(humidity, time of day, T–Td depression, refractivity gradient, sky
cover, season, rain, wind, pressure) with band-dependent weights —
humidity helps at 10 GHz via enhanced refractivity but hurts at
24+ GHz via absorption, for example. Every coefficient in that
formula is a hypothesis waiting to be tested against the contact
and beacon data.
The scoring is a weighted sum of nine factors: humidity, time of
day, T–Td depression, refractivity gradient, sky cover, season, rain,
wind, pressure. The weights are band-dependent because the physics
changes as you go up: humidity helps at 10 GHz (more refractivity)
and hurts at 24+ GHz (absorption). Every coefficient is a hypothesis
I'm checking against the contact and beacon data.
</p>
</div>
<div class="space-y-3">
<h2 class="text-xl font-bold">What we've collected</h2>
<p>Live counts from the production database:</p>
<h2 class="text-xl font-bold">What we've collected so far</h2>
<p>Straight from the production database:</p>
<div class="grid grid-cols-2 md:grid-cols-3 gap-3">
<.stat_card label="Contacts" value={@stats.contacts} />
<.stat_card label="Weather stations" value={@stats.weather_stations} />
@ -91,67 +108,64 @@ defmodule MicrowavepropWeb.AboutLive do
</div>
<div class="space-y-3">
<h2 class="text-xl font-bold">The stack</h2>
<h2 class="text-xl font-bold">How it's built</h2>
<ul class="list-disc list-outside pl-5 space-y-2">
<li>
<strong>Backend:</strong> Elixir / Phoenix 1.8 with LiveView for the
real-time map, Ecto on PostgreSQL for storage, Oban for the
background data pipelines (HRRR fetch, terrain, ASOS, soundings,
IEMRE, solar indices, commercial links), Bandit as the HTTP server.
Elixir / Phoenix 1.8 + LiveView for everything web-facing. Ecto
on Postgres for storage. Oban runs the background pipelines
(HRRR pulls, terrain, ASOS, soundings, IEMRE, solar indices,
commercial links).
</li>
<li>
<strong>Frontend:</strong> LiveView with Leaflet for maps, Canvas
tile layers for the HRRR heatmap (we render 0.125° cells at
interactive frame rates), Tailwind v4 + daisyUI for layout, esbuild
for JS bundling.
Leaflet + Canvas tile layers for the propagation heatmap — we
render 0.125° cells at interactive frame rates. Tailwind v4 +
daisyUI for layout, esbuild for JS bundling.
</li>
<li>
<strong>Physics:</strong> ITU-R P.526-16 knife-edge + Deygout
3-edge terrain diffraction, ITU-R P.838-3 rain attenuation,
dynamic k-factor from live HRRR refractivity gradients, great-circle
geometry for link budgets, Free-Space Path Loss with frequency-
dependent O₂ and H₂O absorption per band.
Physics: ITU-R P.526-16 knife-edge + Deygout 3-edge terrain
diffraction, ITU-R P.838-3 rain attenuation, dynamic k-factor
from live HRRR refractivity gradients, great-circle geometry,
FSPL with frequency-dependent O₂ and H₂O absorption.
</li>
<li>
<strong>Machine learning:</strong> Nx / Axon / EXLA scaffolding for
a 13-feature (8 atmospheric + 4 cyclical temporal + 1 log-frequency)
feed-forward network, 64 → 32 → 1 sigmoid. Not trained yet —
waiting on a larger calibration dataset.
Data sources: NOAA HRRR (AWS S3, hourly analysis + 18 h
forecasts), Iowa Environmental Mesonet (ASOS and upper-air
soundings), IEMRE gridded reanalysis, SRTM 90 m terrain, SNMP
polling on seven commercial microwave links near DFW at 5-minute
intervals, and Copernicus ERA5 for pre-2014 contact enrichment.
</li>
<li>
<strong>Data sources:</strong> NOAA HRRR model (AWS S3, hourly
analysis + 18 h forecasts), Iowa Environmental Mesonet (ASOS &amp;
upper-air soundings), IEMRE gridded reanalysis, SRTM 90 m terrain
tiles, SNMP polling of seven commercial microwave links near DFW
at 5-minute intervals, Copernicus ERA5 for pre-2014 contact
enrichment.
There's Nx/Axon/EXLA scaffolding for a small feed-forward model
(13 features → 64 → 32 → 1 sigmoid). Not trained yet — waiting
on enough calibrated data to not overfit.
</li>
</ul>
</div>
<div class="space-y-3">
<h2 class="text-xl font-bold">Where this goes next</h2>
<h2 class="text-xl font-bold">What's next</h2>
<ul class="list-disc list-outside pl-5 space-y-2">
<li>
<strong>Beacon calibration.</strong> A distributed network of
amateur receivers continuously reporting CW beacon signal levels,
feeding ground-truth measurements back into the scoring algorithm.
Beacon submissions are now open to anyone via the Beacons page.
<strong>Beacon calibration.</strong>
A distributed network of amateur receivers reporting CW beacon
signal levels on a schedule. Beacon submissions are already open
to anyone via the Beacons page.
</li>
<li>
<strong>Scoring weight calibration.</strong> Fit the 9-factor
weights against recorded QSO distances / counts per band so the
composite score reflects real propagation rather than intuition.
<strong>Weight calibration.</strong>
Fit the 9 factor weights against recorded distances / QSO counts
per band so the score reflects real propagation, not my
intuition.
</li>
<li>
<strong>ML model training.</strong> Once we have enough labeled
contact+condition pairs, train the Axon model to replace or augment
the hand-tuned scoring function.
<strong>Training the model.</strong>
Once there are enough clean (contact, conditions) pairs, train
the Axon model to augment or replace the hand-tuned scoring.
</li>
<li>
<strong>Better inputs.</strong> MRMS for precipitation at
24+ GHz (where rain attenuation dominates), RTMA/URMA surface
analysis blending, GOES total precipitable water.
<strong>Better inputs.</strong> MRMS for precipitation at 24+ GHz where rain attenuation
dominates, RTMA/URMA surface analysis blending, GOES total
precipitable water.
</li>
</ul>
</div>