eme, Open Research Institute

October 9, 2026October 9, 2026

Building a 2304 MHz Septum Feed for an EME Station

“Hello Giggy,” is a one-meter offset dish that needed a circularly polarized feed for 2304 MHz Earth Moon Earth digital communications experiments. This is the story of how that feed went from a parameterized OpenSCAD file to working aluminum feed horns. The journey began with paper dolls, bend allowance coupons, SendCutSend’s design rule checker, and three different electromagnetic simulation programs. We learned a lot from when things went wrong. The process from paper to working feed took four weeks.

Choosing the Design

Circularly polarized septum feeds are what are generally used on EME dishes. They are built from a square waveguide that is divided by a metal plate. The metal plate is cut on the diagonal in a stair-step style. These steps convert each of two rectangular ports into left- or right-hand circular polarization. One is used for receive and the other is for transmit. The first set of design work for this feed had derived all the dimensions from the classic Chen & Tsandoulas paper. It turns out that this describes a design that requires a dielectric to work properly. All-metal amateur septum feeds tend to use OK1DFC’s published 23 cm design, verified in HFSS by Paul Wade W1GHZ and built into a wide variety of EME stations.

So we started over. The final dimension set is the OK1DFC 1296 MHz spreadsheet values scaled by exactly 1296/2304 = 0.5625: an 81.51 mm square guide, a five-step septum, and a 1.4-wavelength flared aperture chosen from W1GHZ’s published tables for our dish’s f/D of 0.83. Every number in the design traces to one source through one scale factor and stays true to known-good EME feed designs.

One File to Rule Them All

The entire feed lives in a single parametric OpenSCAD file. It contains the verified 1296 reference values, the scale factor, a bend-allowance model, every flat pattern, a 3D verification visualization, and an echo block that prints a human-readable build sheet, bend line positions, check dimensions, and hole positions. If you change the frequency, the material thickness, or the bend parameters, then everything scales as it should.

There are two Python scripts. One generates a set of dimensioned drawings and the other generates vendor-ready flat-pattern cut files, based on SendCutSend’s file requirements. SendCutSend is the company we ended up going with to cut and bend the metal parts for the feed.

Paper Dolls

Before cutting any metal, we printed the flat patterns at full scale on cardstock and built the feed with tape and glue. This sounds quaint, but the process found real bugs, as expected. The septum was missing corner notches that the backshort cap needs to slip over the assembled tube. It found every place where holes seemed to be missing. This was forgotten and deliberate drill-after-fit decisions that hadn’t been correctly documented. A fellow SBMS builder’s paper doll was assembled with the connector at the wrong end, and this sparked a discussion about why the probe seems too close to the backshort. The probe sits only 24 mm from the backshort instead of the quarter-guide-wavelength rule of thumb. This is because the fat probe and the close backshort and the target frequency are a matched set.

The paper doll also revealed the right order to assemble the parts and what tools to make sure we had on hand in order to reach the fasteners. Both machine screws and rivets were used to assemble the feed, and the paper doll process let us sort it out before any metal was cut.

Bend Allowance, Coupons, and a Big Surprise

Sheet metal shrinks into its corners. A hat channel, which is the type of bend we were using for the waveguide body clamshells, is shorter than the sum of its formed dimensions by one “bend deduction” per bend. That deduction depends on the material, the tool used, and the operator. We designed a calibration coupon. This is an 80 by 30 mm strip, scribed at exactly 20 and 60 mm, and bent into a U-channel. Measuring before and after gives us the bend diameter of the shop brake we were using. This number goes into the OpenSCAD, so that the flat pieces will end up the right dimension when bent.

Paul bent five coupons. Some were with the grain and some were across the grain. Both sides were measured so that any skew could be averaged out. Results were that the grain effect was about 0.1 mm (3003 aluminum), and repeatability was about +/- 0.25 mm per bend. Stacked across the clamshell’s four bends, hand-brake repeatability alone could make about a millimeter of difference. And then we got a big surprise. Neither of the two shop brakes at Remote Labs could physically form the deep hat channel anyway! This part belonged on a press brake.

SendCutSend and the Design Checker

We sent out dimensioned drawings and got a quick and polite reply. Their instant-quote pipeline wants cut geometry, not a bunch of documentation. Their published bending calculator gives the K-factor and bend radius their press brakes use for each material. For 0.040″ 5052-H32, K = 0.45, R = 0.6096 mm, bend deduction was 1.5755 mm, which our own formulas reproduced to a thoroughly ridiculous four decimal places. Those numbers went into the OpenSCAD file as parameters, and the flat patterns exported. That is the entire vendor work flow. We take their constants, use our geometry, and then export the right type of document to upload to their website.

Their checker then gave us a compressed education in bend-line file conventions, which had to do with how a box or pan is specified for their bending and cutting machines. In order, here’s what had to be corrected:

A stray contour becomes a separate part. A leftover 0.3 mm sliver of scribe geometry in one file was interpreted as a second, below-minimum-size part and derailed the whole preview.
Bend lines that meet at corners form a closed loop, which looks like an interior cutout. Our pan cap’s four bend lines joined into a square and the checker insisted that the face was a hole.
Bend lines that cross are rejected as intersecting bends. Extending the lines past the part edges (a convention some CAM systems require) made things worse, and inflated the reported part size by a really large amount.
The actual fix for perpendicular flanges is corner relief: a little circle at each bend-line intersection, radius at least bend radius plus material thickness plus 0.5 mm. With four R2.5 reliefs cut at the corners, the one-piece pan finally sailed through. The breakthrough came from refusing to believe a sheet-metal shop simply could not make a box, and going back to read their guidelines until we found why our pan/box/backshort refused to import into their workflow.

Parts arrived and were perfect. Our acceptance gauge was the septum itself. The full-height section was cut to exactly the 81.5 mm guide width at fit perfectly into the bolted clamshell pair.

Probes

The probes are 1/8″ brass rod. This design is W1GHZ’s “fat probe” that substitutes bandwidth for tuning screws. Each rod was machined and center-drilled 1.3 mm by Paul Williamson. The rods were soldered over to the precisely cut pin of a flange-mount SMA. The SMA had extended PTFE dielectric which was trimmed flush with the inside wall. The rod seats against the Teflon face. This makes the wall end of the probe coplanar with the side of the waveguide, and exactly as long as it needs to be for 2304 MHz.

Soldering brass rod to an SMA pin can be done with a capillary joint. This can be intimidating, but with flux in the socket and a hot enough soldering iron, solder can be fed into the seam as soon as it is hot enough. While some people have had success with this method, the four probes (for two feeds) did not cooperate fully with this plan. The probe had to be lifted up slightly and the SMA pin and probe socket heated, and solder flowed into this joint, then the probe dropped fully onto the pin. This worked for all four pins. All four probes were soldered this way, and continuity checked before anything left the bench.

Simulation

We modeled the feed in openEMS, the free open-source FDTD solver, rebuilding the geometry from the same parameter block as the CAD. Brian Yee W6BY independently reconstructed the design in SolidWorks and ran it through CST, a commercial finite-element solver, with real lossy metals. Matt Smith AF0FO is making progress with HFSS.

The two completed solvers agreed on everything they could both compute. Axial ratio converging to about 1.2 dB (W1GHZ’s own tolerance analysis predicts about 1.5 dB for this dimension set), port isolation 23–25 dB, a 13 dBi pattern with 40 degree beamwidth and −18 dB sidelobes. This is good for the aperture. This is what our 0.83 f/D dish wants to see. Radiation efficiency with real aluminum and copper was 100% to within thousandths of a dB. These were all really good results.

The one number both solvers produced that caused a bit of concern was match depth, reading about −11 to −12 dB. Another dive into the W1GHZ literature explained why. The published dimension set seems to describe a thin probe compensated by a tuning screw. The fat-probe, screw-free variant that everyone actually builds was W1GHZ’s recommendation, but not as many updates about this design.

The Rigol answers the open question

With both feeds assembled, we calibrated the Rigol spectrum analyzer’s VNA mode with a short, open, and load at the far end of the cable chain, and measured all four ports.

Return loss at 2304 MHz: 17 to 19 dB on every port of both feeds, SWR between 1.24 and 1.34, with the two independently built feeds matching each other to within 0.1 SWR. Port-to-port isolation, measured as vector S21: 22.5 dB on one feed, 23.7 dB on the other, against a simulated prediction of 22.8.

The trace itself answered a question Paul asked that deserves a permanent answer: shouldn’t we see a nice dip at the design frequency? No, and the flat trace is the design working. Narrowband antennas match through a single sharp resonance and show a V at their frequency. The fat probe is deliberately low-Q. It makes a broad shallow bowl, under 1.4 SWR across a 500 MHz window. The hardware came in 5 to 8 dB better than the simulations.

Then we carried the feeds around the neighborhood with a HackRF connected up to each one. We confirmed they worked with real radios attached and then we demonstrated the mirror rule of circular polarization, that two same-handed feeds facing each other are cross-polarized, on the air. The feeds work!

What’s Next

Next is connecting the feeds to a boom and sliding carriage to replace the dish’s Ku-band LNB arm. The feed’s phase center gets peaked on sun noise. Deep Space Exploration Society was interested a high-power variant of this same parametric design for their 18-meter dish. This would be the same OpenSCAD design with thicker walls, 7/16 DIN launch, and no flare for their deep prime-focus optics. Same file but with different parameters.

All the design files for the feed are in the ORI EVE repository at https://github.com/OpenResearchInstitute/EVE

Thanks to Paul KB5MU, Brian Yee W6BY for the CST validation, Matt Smith AF0FO and Chuck Hill for HFSS support, and the San Bernardino Microwave Society for additional paper dolls, technical support, and enthusiasm. Open Research Institute develops open source digital radio technology for experimentation and education.

October 8, 2026October 9, 2026

EME Station for EVE Testing: Signal Construction

The Earth Moon Earth (EME) station for Earth Venus Earth (EVE) testing effort continues! Let’s walk through the signal construction and explain how it’s received.

The challenge of EVE is to bounce a signal off Venus and receive it on Earth. Venus rotates (slowly) and it reflects back a small percentage of radio waves aimed in its direction. The amount reflected back to us is related to the reflectivity, or albedo, of the surface of the planet. This percentage ranges from 10 to 15 percent, depending on what part of the planet happens to be facing our radio transmitter. Since the surface of Venus has been mapped, and since we can calculate this in advance, we have a good idea of how much of our signal will be reflected back.

That reflected signal isn’t just weakened by a lossy reflective surface. Venus is very far away, so the path loss is immense. And, since Venus is rotating, the reflected signal is smeared from the rotation. This is called Doppler Spread, and it has a direct effect on how we design, contruct, transmit, and receive our signal. The enormous path loss and the Doppler Spread (low coherence time) of the signal are the two biggest challenges for a signal designer.

Pete Wyckoff tackled this problem by directly confronting the Doppler Spread. The signal design is based around this physical limitation of the channel. The reciprocal of Doppler Spread is coherence time, which is the amount of time we can coherently integrate the signal before we lose phase information. Integration is essentially gathering up energy over time and figuring out where the power appears. Higher power values at particular frequencies indicate a high probability of the presence of a transmitted signal. Most digital signal designs assume an infinite or very long coherence time. Dealing with Doppler Spread is not something that a typical radio signal designer has to face. When we deliberately bounce a signal off a reflector like Venus, then we have to account for this effect.

After this coherence time expires, we can still make use of the magnitude information for our signal integration, but we lose out on phase. This means we have a penalty to our receiver signal to noise ratio compared to a signal that doesn’t have any Doppler Spread.

Spiral, our recommended modulation and coding scheme, is a frequency shift keying design, with forward error correction. It’s a 4096-ary design, where each tone represents 12 bits of information. A symbol is a single tone. A frame is 11 tones. 11 symbols times 12 bits gives us 132 bit-slots. We set up 90 message bits and use 16 bits to give a cyclic redundancy check (CRC). This gives us 106 payload bits. We use BCH(127,106) forward error correction encoder. This adds 21 parity bits. We have 127 frame bits fill 132 slots, with 5 leftover slots zero-padded. A frame is 11 or 12 ASCII characters, depending on the type of ASCII encoding.

Doppler Spread is 1 Hz at 1299.5 MHz and 2.8 Hz at 2304 MHz. Coherence time is the reciprocal of Doppler Spread. We have a 0.5 second coherence time at 1299.5 MHz and 0.35 second coherence time at 2304 MHz. These two frequencies are the ones most commonly considered for EVE work.

Using the 0.35 second long receive window, we set our receiver tone bandwidth to 2.87 Hz. We set a tone spacing of 5.74 Hz, so that we have an entire “guard band” for each of our tones.
Using the 0.5 second long receive windows, we set our receiver tone bandwidth to 2 Hz. We set a tone spacing of about 4 Hz, and we run the symbol length as long as needed to get CNR in 1 Hz target.

We run the symbol length as long as needed to meet or exceed the CNR in 1 Hz target. If we’re at a station where we are receiving our own signal, then we send as long as we can without exceeding the round trip time. This is about four minutes for Venus.

The most straightforward way to do this (if you have to receive your own transmission) is to send one symbol per transmission. Send one symbol, which is one tone, for the entire round-trip time. Then turn the transmitter off, and receive the entire echo. If you have another station listening while you transmit, then you can send one tone after another. Each tone is sent as long as it takes to build up enough integrated signal power in order to recover the transmitted tone.

How do we receive? For each symbol, the receiver calculates a series of coherent windows using the coherence time. For example, the coherence windows would be either 0.35 or 0.5 seconds long, depending on the transmission frequency used. Within each of these short windows, we coherently combine. We collect these windows, one after another, as long as the signal is received. Once all these individual windows are collected up, we non-coherently combine all of these results. So, coherent within the coherence time window. Non-coherently add up all these individual coherent results.

From this collection of windowed results, we determine the tone what tone was sent. We collect a full frame of 11 tones. We determine the 12 bits indicated by each specific tone. We decode and then return the message.

Since Venus isn’t usually close enough to try things out, we decided to use the Moon as a test case. Althought the Moon appears to not be rotating, it is wobbling back and forth. This is called Libration, and it causes a Doppler Spread too. We thought this would be an excellent test bed for EVE signal design. We weren’t expecting it to be as interesting as it was!

The libration, or cyclic wobbling, is very much like a rotation. It causes Doppler Spread. Unlike a steady state rotation of a planet like Venus, the Moon’s Doppler Spread varies. The Libration cycle goes from fast to slow and back again, back and forth, over the course of a month. The amount of Libration can be tracked in the Python library Skyfield.

When we calculated the amount of Doppler Spread for EME tests of our EVE signal, we found that it varied above and below what we would be getting with Venus. For one particular date, sending EVE signals to the Moon to test would incur a penalty of over 5 dB, simply because the coherence time was much shorter than Venus! The effect varies depending on where you are on Earth. Higher latitudes don’t see as much Doppler Spread as lower latitudes. The Earth’s equator is the most challenging place to be.

In order to get the most performance out of EVE signals tested with the Moon, we adjust the tone spacing of our signal, and we adjust the duration of the coherently collected windows, based on the worst-case Doppler Spread for when the transmission was made. We round to the nearest hour, calculate the worst Doppler Spread a receiver would possible experience, and set the tones and coherence windows based on that Doppler Spread. That is how we adjust the Spiral communications protocol for EME.

Is this strictly necessary for EME communications? No. There are many ways to bounce a signal off the Moon. This scheme, where we adapt to varying Doppler Spread, provides a real-world case study of low coherence time signal theory. For very low SNR situations, where every dB counts, techniques like this can make the difference between copying a signal and receiving nothing.

October 8, 2026October 8, 2026

Science and Technology for EVE 2028

Do you like amateur radio, amateur radio astronomy, and digital communications? Are you excited about the Earth-Venus-Earth communications attempts in 2025 and 2026?

The Earth and Venus are next aligned on the same side of the sun on 1 June 2028. Looking at the link analysis, the geometry of 2028 is a bit different than 2026. Venus passes much closer to the Sun’s disk in June 2028, at a separation of only about 0.8 degrees, compared to 6.5 degrees away in October 2026. In June 2028, Venus will be located high in the northern sky with a positive declination of approximately 22 degrees 54 minutes.

If you would like to help organize an EVE communications attempt (bouncing a signal off of Venus to be received back here on Earth), then please get in touch. We’ll do all we can to support, educate, and publicize. We have a modulation scheme designed specifically for the low coherence time EVE channel. We have an EME test station at ORI so that research and development can happen year-round for EVE work. Welcome aboard!

https://openresearch.institute/getting-started