Wednesday, October 13, 2021

Driver for an electromagnetic pendulum

In a previous post, I described some experiments with circuits for driving an electromagnetic pendulum. The results were not very satisfactory, and for the Thriecan clock, I ended up using a commercial module. I have since revisited this, and now have something I am happier with. To recap: the pendulum contains a magnet. As it approaches the center of its swing, a sensor detects its presence, and then energizes a driver coil to impart some momentum to the pendulum. Ideally, you use one coil as both the sensor and driver. The circuit I present here uses an Arduino and a small number of discrete components.

Some initial experiments

In the first attempt, I was unable to get a strong signal from a coil: only a few millivolts. This is probably because of the geometry of the coil, which had a narrow diameter but was quite long (about 25mm), based on the one in the instructions for the Toucan clock. It is better to have a shorter, fatter coil, so that more of the winding is close to the magnet as it passes by. I rewound the coil on a bobbin with a 6mm inner diameter and a length of 10mm. I did not count the number of turns or measure the length of wire. Based on its resistance (35 ohms) and the resistance per unit length for 32 AWG magnet wire, it is around 60m. The winding is about 30mm in diameter.

I wanted to know how much difference the specific magnet made, so I constructed a pendulum with the magnet on the end of a 300mm brass rod and measured the voltage when the pendulum was dropped from a known position. The magnets I had lying around were:

  • 20mm x 4mm N38
  • 12mm x 3mm N35 (I think)
  • 12mm x 3mm N42
  • 10mm x 3mm N52
All but the first gave about 0.3V. If you add a second or third one, this goes up by 0.1V per additional magnet. The 20mm one gave 0.5V or around 0.65V with two. I was a bit surprised that the grade of magnet didn't make a difference, as each unit of N rating is supposed to correspond to an extra 1% in magnetic field strength (or something like this).

The distance between the magnet and the coil also makes a difference. My test setup didn't allow me to change it much, but an increase of distance between the coil and the magnet from about 3mm to about 6mm dropped the voltage by half.

The circuit

Here is the circuit. I'll explain the ideas behind it in a moment.

(Diode: 1N4001. Transistor: PN2907.)

The voltages I measured aren't enough to register as a digital signal when connected to an Arduino and are maybe too low to switch a transistor which could be attached to a digital input. An alternative is to use an analog input of the Arduino and then set a threshold in code based on the analog reading. With a resolution of about 5mv per step, we should be able to do this. A reading of even 0.1V would be detected as a value change of 20 on the analog input.

If you search the web for advice on connecting a coil as a sensor to an Arduino, there is a lot of unclear (and possibly misleading) information. The main concern is that transients from the coil could exceed the input voltage range of the Arduino and so damage it. Various solutions involving voltage dividers or Zener diodes can be found. I don't think this is necessary. The Arduino inputs have protection diodes which limit the voltage to just above 5V and just below 0V. The issue with the protection diodes is that they can only handle a limited amount of current. 1mA might be OK, 100uA definitely is. More than that would likely burn them out. In the circuit above, the flyback diode on the coil should eliminate most of the risks, but to be sure, I designed the circuit to allow for a 10V swing. In this case, with a 100k resistor on the analog input, the current would be 100uA and we are OK. Relying on the protection diodes is something that people argue about on various forums, leading to some of the other solutions. An application note from Atmel has an example in which an analog pin is connected safely to mains at 110-240V AC. Also take a look at this.

There is one further issue with using the analog input. The Arduino datasheet recommends that the input impedance to analog pins should be 10k or less, and we want to use 100k. Looking into this in more detail, the reason appears to be that the analog to digital converter works by charging a capacitor internally and them sampling its voltage. The internal capacitor is 14pF and it should be charged in 6us or less at standard sampling rates. With a 100k external resistor, we have a time constant of 1.4us, so we should be OK.

The remaining input element is the 10uF capacitor. It largely gets rid of noise and ringing on the input without slowing down the signal too much. Looking at the signal on an oscilloscope, there is still a small and short duration spike at an acceptable level. With no capacitor at all, there are spikes for about ten milliseconds before the circuit settles down.

On the output side, we simply drive a transistor from a digital pin. When the pin is low, the transistor turns on and the coil is energized. The sensor can't be read at the same time, and that's OK as we don't need to. The flyback diode helps to protect the transistor when it turns off and the field in the coil collapses.

In previous experiments, I found that I needed external power to get enough drive in the coil, mainly due to use a coil with less good geometry. For this version, the 5V output of the Arduino worked fine. Its Vin can also be used if a higher voltage or current is needed. I experimented with an Arduino Uno, but will move to using a Nano soon.

The code is similar to the version I gave before, namely:
  • read the sensor
  • when it exceeds a threshold (20), wait (10ms)
  • turn on the coil (200ms)
  • turn off the coil, and wait a little longer (10ms)
  • keep going
I still have to tune the exact timings and the threshold. These values give a very strong swing to the pendulum. Currently I am not using a weight bob or tuning to 1s per cycle, and this may affect the timing and possibly the voltage needed to drive the coil.

I've not fitted this to the Thriecan clock. My intention is to use it for a possible future design.

Sunday, October 10, 2021

The Thriecan Clock

 The Thriecan Clock is a modified version of Clayton Boyer's Toucan clock, adapted for 3D printing. The key feature of this clock is a electromagnetic pendulum. There is a magnet attached to the end of the pendulum and a coil hidden in the base. As well as providing the timing, the pendulum also provides the energy to drive the clock. As the pendulum approaches the vertical position, the electronics attached to the coil senses the magnet and applies a pulse to the coil. This pushes the magnet and hence the pendulum away and the process continues. The top of the pendulum is attached to an arbor and this turns a cam with two pawls. One pawl, called the pick-up pawl, pushes on the escape wheel. The other pawl then holds the escape in place while the pick-up pawl moves back for the next cycle. The remainder of the clock is a standard gear train to divide the rotation of the escape wheel down to minutes and hours.

The version you see here is not quite final. I am going to replace the battery pack with power connector so I can connect it or a different power source. There is enough room to locate the battery pack in the base, but it disrupts the operation by attracting the pendulum. There's also a few screws sticking out while I make final adjustments and I will then replace them with shorter ones.

Adapting the design

The original design was intended to be made out of wood. You can see many examples on the Toucan clocks YouTube channel. I bought a copy of the plans in DXF format and loaded them up into Fusion 360 to provide a starting point for the sketches. For parts where the dimensions were critical, I used these sketches directly, and for others I either adapted them (for example, the cam) or replaced them entirely (for example the weight bob). The original design used 3/8 inch arbors, and I replaced these with 3mm ones. Instead of printing separate gears, connectors and pinions and then gluing them together, I merged them into a single part. This is something which works nicely for 3D printing, but it difficult or impossible in woodwork. I undersized many of the holes for the arbors and for the pendulum shaft and drilled them out to either a tight fit or a loose one. I've not always been successful in getting good tight fits and so where possible I also provided holes for set screws.

One of the hardest parts to adapt was the frame. It is much too big to fit on the printer bed. It's possible to split up large pieces like this and then glue or otherwise connect them. However, I didn't like where the splits ended up so I changed the shape of the frame. It still needs to be split into three pieces: the main part of it, the left foot, and the curve reaching down to the right foot. You lose the nice curve on the left hand side of the original design by doing this. The joins are hidden behind the dial on the left and where the arc on the right joins the vertical part. In each case, as well as gluing the parts, there are some metal pins joining them. These help keep the parts aligned while the glue sets and provide a little extra strength. The dial is also slightly smaller, and has a central bar to help support one arbor and the hour wheel. It's held on to the frame with two M3 screws. I didn't really think ahead here, so I ended up having to use 45mm screws with a bit of padding behind the frame to get the length just right.

One reason for wanting to try this specific design was to see how easy it is to start with a plan designed for woodworking and adapt it for 3D printing. It worked out somewhat well. The process was a bit hindered by the ways the plans are supplied: the DXF is a single gigantic file some parts of which are actual plans, and some parts of which are descriptive text, assembly diagrams and other auxiliary information. It seems an odd way of doing things. Fusion 360 gets a bit heated when you load all this in, and so a fair amount of initial pruning was needed. I'm also a novice with Fusion 360, so the way I did some things might not be best.

The pendulum

The pendulum is a 2mm brass rod with the magnet holder as the bottom and the weight bob around the 25cm mark. Adjusting the position of the weight is a bit fiddly as it involves loosening the screw underneath it, and sliding it up or down. The bob holds quite tightly to the shaft so adjustment can be down without tightening the screw until it is in its final position. The weight itself is a few M3 screws in the shiny little bucket.

The pendulum is connected to the cam with a 3mm brass shaft. This is a problematic part of the design, as the weight of the pendulum is greater than that of the cam and pawls, causing the shaft to tilt to the back. There are a few ways this could be fixed; perhaps a counterweight on the front, or a longer support glued to the back of the frame. Another option would be to move everything in front of the from forward and lengthen the shaft. For now, it works OK as it is.

One tooth or two?

Looking at the examples of the Toucan clock on YouTube, there are two different philosophies about the position of the pick-up pawl and the stop pawl at the point where the pick-up pawl starts to push on the escape wheel. They can either be on adjacent teeth like this, or separated by one tooth like this. The clock runs either way, and you can set the way it works by adjusting the angle between the pendulum and the cam. It may make some difference to the required swing of the pendulum. Looking at the first 10 videos on the YouTube list, I saw 8 were adjacent (the Toucan closes its beak) and 2 were separated (the Toucan eats with its mouth open). I went for adjacent, which sets the cam a bit anticlockwise from the pendulum.

Driving circuit

I experimented with various driving circuits. As mentioned in a previous post, I had no success with getting a 1 or 2 transistor circuit to work, and after fiddling around for a long time, I wasn't happy with any of the other options I tried. In the end, I decided to buy a ready made module from carveshop. It works well, but is a little pricy. I plan to spend some more time looking into this, in part because I have a second electromagnetic pendulum clock I would like to try.

Sunday, October 03, 2021

An electromagnetic pendulum

 For a future clock project, I would like to construct an electromagnetic pendulum. The pendulum has a magnet at its tip, and base contains a coil of thin wire. We aim to detect when the pendulum is approaching, and then turn on the coil for a short time to add extra energy. This can be done by turning on the coil to attract the magnet before the pendulum reaches the center (vertical), or waiting until it has passed the center and turning on the coil to repel it.

There are many different ways of doing this, and it was used in commercial clocks before electronic movements took over. The German-made Kundo clocks are one example. If you look on the web, you will find many articles about electronic pendulums, whether for clocks or just as toys, with a variety of different circuits. One of the best descriptions is Kundo battery clocks by Rod Elliot, with several possible circuits and a lucid explanation of how they each work. There is a 2 coil, 1 transistor design, used in Clayton Boyer's Toucan clock, and two variants of a 1 coil, 2 transistor design from the Kundo clocks themselves. In another article on the same site, Rod Elliot notes that there is some trial and error in getting these circuits going. After playing around for a few days without getting to anything reliable, I agree. I also found a site with a very similar circuit and a comment thread full on people saying how they never managed to get it to work. I'm not sure why I had so much difficulty. I know that for a few components (capacitors mainly), I didn't have the exact values and had to come up with a substitute. I tried out several different coils that I had around. Some didn't produce enough voltage to turn on the transistors. They would need more turns or more cross-sectional area, both of which affect the induce voltage. Others could clearly turn on the transistors but didn't impart enough energy the the pendulum.

There are other solutions, both analog or digital. One analog option is to replace the transistors with two or more stages of op-amps. You can then tune the sensitivity of the triggering better, and also add some delay before the pulse to the coil happens. Other solutions use a simple microprocessor like a PIC, or a TI chip.

I decided to go with a simple and highly controllable solution. I had some KY003 Hall effect sensor boards in my stock of parts, and so I just connected one to an input pin on an Arduino. The Arduino can't drive the coil directly, so I hooked up an output pin to a transistor, like this:

The transistor is a PN2222, and the Arduino output pin connects to the 22k resistor.

The code is quite simple:
  • wait until the Hall effect sensor turns on.
  • wait a while to allow the pendulum to reach the center.
  • energize the coil for a while.
  • check the Hall effect sensor has turned off (in case we are still within range).
The pendulum itself if a 30cm brass rod. I want a period of about 1 second, so there is a weight around 25cm from the pivot. The magnet is at the end. The delay after detecting the sensor should be long enough for the pendulum to reach or just pass the center. If the angle of the pendulum at its extreme is P, and the angle at which we detect the sensor is S, then the delay should be arccos(S/P)/6.28, to a first approximation. The 6.28 comes from the angular frequency being 6.28 radians/second for a 1 second pendulum. See here. None of this is exactly right, as it assumes a small angle for P, whereas I actually see around 20 degrees in each direction. It also does not allow for extra energy being added to the system. But it will do to get roughly the right values.

This is the prototype working. It needs a small nudge to get going:

Here's the code. Note this is a prototype and might change before the final version.

// Electromagnetic pendulum controller.
// Detect the magnet with hall effect sensor on this pin...
constexpr int hall_sensor = 3;

// ... then wait this many milliseconds ...
constexpr int detect_to_activate_delay = 23;

// ... then energise this output pin ...
constexpr int output = 7;

// ... for this many milliseconds ...
constexpr int activation_duration = 200;

// ... with this much slop in milliseconds when checking the sensor is out of range.
constexpr int sensor_cooldown = 10;

// Use this pin for LED.
constexpr int led = 13;

void setup()
  pinMode(led, OUTPUT);
  pinMode(output, OUTPUT);
  pinMode(hall_sensor, INPUT);
  digitalWrite(led, LOW);
  digitalWrite(output, LOW);

void loop()
  // Sensor reports LOW on detecting a magnetic field.
  if (digitalRead(hall_sensor) == LOW) {
    digitalWrite(output, HIGH);
    digitalWrite(led, HIGH);
    digitalWrite(output, LOW);
    digitalWrite(led, LOW);

    // Wait until the magnet is out of range of the hall effect sensor, and then allow a little longer.
    while (digitalRead(hall_sensor) == LOW) {}

Additional notes

After some experimenting with this form of drive, I think there are some good and bad points.

Good points: it is very easy to build as the circuit is so simple. It's convenient for tuning the timing. You can add a few extra lines to report the time between ticks. Note that the activation delay and duration make very little difference to the timing, so this is about adjusting the position of the bob on the pendulum.

Not so good: getting both the Hall effect sensor and the coil close enough to the magnet is a challenge. I mounted the sensor on top of the coil, but this then requires a large coil to supply enough impulse. The force on the magnet goes down (if I remember correctly) as the square of the distance from the coil. To make life easy, my coil was whatever was left on a 4oz could of 30AWG magnet wire, probably about 300 metres. Anything much less than this didn't work. This could be solved by more careful design, such as mounting the probe embedded in the top of the coil or in the hole in the middle.

Wednesday, September 22, 2021

Mending broken DXFs in Fusion 360

I recently downloaded some DXFs originally designed for use with a CNC machine, with a view to converting them into something suitable for 3D printing. Blender is my go-to for quick manipulations, but in this case I wanted some more advanced tooling and so I decided to use Fusion 360. I had multiple problems, and so I decided to document them here in case this helps others in the same situation. The root cause of most of the problems is disconnected segments. It seems that some programs which output DXFs will leave a small gap between segments. As a result Fusion 360 can't find closed curves and this prevents it doing many important operations such as extruding. The DXF file I am working with is about 4MB and contains around 30 separate items, some for parts such as gears, and a few to illustrate how the parts fit together. I'm only interested in the former. I also tried cutting one part from the sketch into a new sketch for some tests.

You can recognize when there are broken segments if extrude and similar operations are not available, or if you hover over what looks like a closed curve and it does not all highlight. Generally you have to zoom in a long way before you can find the segments that are not joined. If there were only a few of these, fixing them by hand would not be too much of a burden, but when you have each gear has two or four of these per tooth, it quickly becomes time consuming.

A first thing to note is that Fusion 360 has two ways of importing DXFs. In both cases, the result is one or more sketches. The options are:

  • From the Design > Insert menu, there is an Insert DXF function. It seems to be essential to set this to "One sketch per layer" to avoid it taking a very long time. In the file I was working with there were 9 layers.
  • Use the DXF Import Add-in, from the Fusion 360 App Store. I was hopeful about this, as it includes an option to fix the sketch by joining close elements, either on import or as a separate operation. The fixing stage appears to be extremely expensive. I tried importing the large file with the fix option turned on, and after several hours it was still going. I doubt that it would complete in a reasonable time, as by that point Fusion 360 was using around 7GB of RAM, forcing my PC to continually swap and so grind to a halt. With the smaller sketch, I didn't run up against the memory limit, but the fixing stage simply didn't do anything. Note that the Add-in does not allow you to set the units (DXF files don't specify what units they are in), a curious omission.
I then looked at other scripts and add-in designed to fix cases like this. I found three:
  • ConnectTheDots is a Python script. The example video shows it working on exactly the case I wanted. It operates on the selected part of sketch, which I though might allow me to concentrate on just the parts I was interested in. Unfortunately, the script does not work with current versions of Fusion 360. It raises a Python exception. It appears to be unmaintained.
  • ConnectTheDots (another one) is a fork of the original version. It does not raise an exception, but never seems to complete.
  • FillGaps is a paid add-in. It costs $10. There is a free preview version to try it out, though the preview is not very useful as you can't examine the resulting segments. It can join nearby disconnected points by either adding a new segment or merging them. It doesn't take a huge amount of memory, though it can be quite heavy on CPU. It works on the whole of a sketch. Unfortunately, on my large test file it got slower and slower. At the start, the progress bar (which shows percentages) advanced by 1% every 10-15 seconds. From 99% to 100% took multiple minutes, and it then stayed at 100% for an hour and a half until I cancelled it. On the small sketch (cut and paste of a single gear), it finished quickly and accurately. So this looks like a good option, with the inconvenience of having to do a lot of cutting and pasting.
One other option is to load the DXF into blender, select a part from it, convert it to a mesh and then use Blender's Edit Mode operation to combine vertices by distance. This works, but leaves the result as a mesh and in some cases seemed to add some distortion. You can also only export meshes as DXF from Blender and not curves: if you try to export as a curve, then it silently does nothing. However, if you want to continue doing all your modeling in Blender it might work.

A final possibility is to treat the DXF as not being part of the model at all, and just use it to guide creating new sketches with them a guidelines. This isn't great for complex geometry such as gears, but could be combined with the Fusion 360 gear generator add-in to generate gears matching the ones in the DXF.

Thursday, September 16, 2021

Steve Peterson's Stepper Clock

Steve Peterson has a design for a desk clock driven by a stepper motor. I am mostly interested in purely mechanical designs, but as I was impressed an earlier design of his, I decided to give it a go. As usual, here's some pictures and video:

The clock consists of a stepper motor driving a gear train to turn the seconds, minutes and hour hands. One of the arbors, called the "gear 5 arbor" and located at the top right, has two gears held together by friction. There is a spring clutch just behind the larger gear. This allows you to turn the knob at the back to adjust the time. The clock is driven by an Arduino Nano and a NEMA 17 motor. Normally a Nano would not have enough power to drive a NEMA 17, but in this case an uncommon type with a higher coil resistance and lower current requirements is used. There is an intermediate board designed and sold by Steve Peterson which simply connects five ports of the Nano to each coil of the motor. Each port is connected through a resistor, with values which are approximately in powers of two. The control program can then advance motor with 1/32 microstepping. The circuit has a passing resemblance to a resistor ladder DAC, though it is not quite the same. It's quiet and works well, with low power requirements and minimal circuitry.

A few places in the mechanical part of the clock rely on parts being a tight fit on their arbors. I'm not keen on this approach. With 3D printing, you have to undersize the hole and then hope to drill it out to a suitable diameter, then force the part into the right position. I found all of the parts were already too large for the arbors. The holes have a diameter of about 1.65mm and fit onto 1.5mm shafts. With some printers and filament, the hole might close up enough to make this work. For me it did not. I decided in the end to print the parts with a small transverse hole and hold them in place with M1.5x5 screws. One piece where I did this is the gear 5 insert that holds the spring for the clutch in place. It needed some care to avoid the screw interfering with the large gear at the back (gear 2, in the design). I also found that you have to be careful in attaching the knob on the back of arbor 5, used to adjust the time. If it drags on the back of the frame at all, this is enough to stop the minute and hours hands.

The stepper runs quite smoothly, but when it is attached to the gear train, there is noticeable judder. In testing, I started with just gear 2 and saw a lot of rebound on each step. It gets less as you put the rest of the gear train together. The following video illustrates:

I used Mika 3D silk PLA throughout. The gears are so-called rose gold, which is in fact a pale pink. The frame is silk black, which is more like a dark grey. I'm not all that pleased with the appearance. Either a darker black or a bright color would have made for better contrast on the numerals. Two of the parts (the stepper gear and the smaller gear on arbor 5) warped during printing, like this:

I was surprised by this. It used to happen a lot when I had a poor quality printer and used blue tape as the print surface. I have not had it happen at all since I got the Prusa MK3. The parts were OK on reprinting them.

With the code as downloaded the clock was accurate to within 2 seconds over an hour. I am tuning it further by tweaking the parameters in the code.

And a small update

The first clock in this series, also by Steve Peterson, had been working quite well and accurately for several weeks until recently. It now occasionally stops and the pendulum swing is less than it used to be. I am intending to add a bit more weight to the weight shell, but also think I probably need to take it apart and check for anything that might have slipped out of alignment, particularly on the pendulum arbor.

Wednesday, September 15, 2021

M5Stack Clock

With all that I've written recently about 3D printed clocks, I thought I should not neglect a purely electronic one. Some while back I bought a M5Stack. It's a packaged ESP32 with a screen and (in this model) some neopixel LEDs. I wrote code to use this as a clock. It periodically resyncs to a reference time if it has a wifi connection and otherwise tracks the time internally. This variant of the M5Stack doesn't have a real time clock, but there is a library for tracking the time without one. The colored LEDs on the side very gradually change color: a color encoded as HSV is picked at random and the color change until it reaches the target value. It can run off the internal battery or from USB power. Here's a few minutes of it, at 30x speed:

Not super interesting, but practical and useful.

Tuesday, September 07, 2021

Brian Law's Clock 27

Brian Law is a British designer of clocks. Most of his designs are intended to be made out of wood, and he sells plans for them on his site, He has also adapted a few of the designs for 3D printing. There is a simple beginner's design, and three designs which are more interesting mechanically. I decided to make his Clock 27 design. This is a weight driven clock with a pendulum, and the interesting feature of it is that is uses a novel escapement, designed to minimize friction. A blog post explains how the escapement works. I like the design and found it fairly easy to get it running, with only a few issues. I do think there are a few ways in which it can be improved to make it easier to print and assemble. I'll come back to these in a moment, but first some pictures and video:

The clock running:

And in slow-mo. Make sure you have the sound on for this.

When I first assembled the clock, it would run for a few seconds (20 was the longest) before stopping. The pendulum was not swinging enough to keep the clock going. I tracked this down to the gravity arm, the piece on the left shaped like a crescent moon. The blog post says that "The finger on the end of the Gravity arm needs to be adjusted so that it causes the trigger to release the Escapement wheel just before the fork in the fork in the Lifting lever [The L shaped piece] reaches the curved tooth", and this was not happening. As a test, I stuck a small piece of plastic on the tip of the gravity arm. The clock then ran, though a bit unevenly. I reprinted the gravity arm with the tip extended by 1mm, and took the chance to check over and clean up all of the remaining parts, and after that it ran smoothly.

Brian Law has several parts that are to be printed separately and glued together. For his version, he used ABS, and there are plenty of effective solvent glues for it. I used PLA, where the only glue that works well is gel cyanoacrylate. It requires the surfaces to be smooth and quite close fitting for the glue to hold. One case where it is suggested is to assemble parts like this:

There are two gears here, the big 60 tooth one and the smaller 15 tooth one. They are printed separately because of the overhang and then joined into a single piece. A single part this shape would otherwise need support to print. I found the parts did not glue well together as there is a fairly large gap around the mating piece (which is part of the upper gear). One way of improving this would be to make it a tighter fit. A better way would be to make the mating piece and the hole it goes into hexagonal. And best of all is to extend the teeth of the upper gear down and then merge the pieces into a one that can be printing as a single entity, like this:

This is the solution I adopted. I did glue some of the pieces as recommended, for example the wall spacers on the back of the frame, but also found a few more places where it was easy to modify and then merge the parts.

The clock relies on some of the gears being tight on their shafts, and the design undersizes some of the holes so they can be carefully drilled out to fit the arbors. In most cases, you could get away with a looser fit and extra spacers as an alternative. The one place where there has to be a tight fit is the 60T gear which mates which the ratchet. It directly drives the shaft it is on and thereby the minute hand. I found it was too loose at first, and so tightened the interior hole as well as adding a couple of hole for M2 set screws. This seemed to work; even without the set screws the fit was tight enough that I had to tap the rod with a small hammer until it was in place and the screws are only there in case it loosens up over time.

Note that there is an error in the bill of materials for the clock. Instead of needing three 100mm shafts and one 69mm one, it actually needs one 100mm and three 70.5mm; similarly, you need one 31mm shaft, not three. I used several different materials: stainless for the 100mm, carbon steel (McMaster Carr High Tolerance Rod) for two of the 70.5mm rods, and brass for the remaining one. This was based purely on what I had available. Brass is my preference as I can cut it with hand tools, unlike stainless, and it doesn't corrode. The carbon steel corrodes a little, and as I live near the ocean this can become a problem over time.

A few of the parts are on flying shafts (I mean ones which are only held at one end). Law suggests adding a magnet to keep them in place. I prefer a simpler solution: make the shafts slightly longer and print a tight fitting cap for them. Take a look at the second picture above for some examples.

One other issue I encountered is that once you have attached the dial to the frame, two of the screws which hold the front of the frame to the back are inaccessible.

Before I modified the gravity arm the beat was very patchy and I worried that this might be due to uneven teeth on the gears. However, once I made the slightly longer gravity arm, it ran evenly. It has met my basic test (running for 1 hour). I have yet to tune the timing and as you can see in the pictures it is still mounted on a test frame rather than on the wall.

Where next?

This is the fourth clock I've made recently, each based on a slightly different principle: hanging pendulum, seesaw pendulum, balance wheel, gravity escapement. It's not that I especially love clocks, more that I just find it interesting to make them and then get them working, and I enjoy finding complex printing projects. There's a few more designs I will probably try out. I don't leave most of them running: if I want to tell the time, there are plenty of much more accurate electronic devices within easy reach. The only one that I do leave going all the time is the Peterson clock, in part because it runs for 7 or so days without rewinding. It has stalled a couple of time in the last few days after a few weeks of working well, so some maintenance might be needed.

A footnote

A day after posting the above, I moved the clock from the testing frame to a permanent position on the side of a bookcase, and found it didn't run for more than about 45 seconds. I think this is just that it was not balanced. It took a while, but after slowly adjusting the angle it was hanging at and also adding a washer to the gravity arm to give it a bit more momentum, it seems to be running better - at least two hours without stopping.