Large 140W PCB Hot Plate

October 17, 2026Leave a comment

I’ve used my original PCB hot plate a lot. It is compact, heats up quickly, and most importantly is right sized for my application – big enough for almost any PCB I would work on, and low profile enough to fit under my microscope stand or to do assembly directly on. Its bigger than the MHP-50, and way smaller than something like the typical generic blue lab hotplate that people use for reflow.

It’s also much more sophisticated and powerful than the original PCB hotplate, in pretty much every way. I think the only vestiges of the old design are that it uses the same microcontroller and the same I2C pendant that I designed for the last version.

Handling More Power – 140W EPR

This plate uses a CH244A to ask for 140w EPR at 28V. The extra power is welcome, but its not straightforward to deploy. The original plate “just worked” by having a heater that would draw just over 5A at room temp, which was PWM’ed, and slightly less than 5A current at the hot temperature. The resistance started out low, and it grew proportionally less at high temperature and left more margin at high temperature. My first 140W heater design failed because I started out with a higher resistance value, to limit the low-temperature DC current. Once it grew by ~70% due to increased thermal resistance, it was impossible to deliver enough power to it to at high temperatures to keep it getting hotter. This design was self limiting.

In order to deliver maximum power at higher voltages, I needed a resistance that would be close to 5.6 ohms at ~200C. That’s pretty close to the original 3 ohm heater I designed, but with one catch – at 28V, the 3 ohm heater would draw almost 10 A instead of around 6A. This seemed like a less manageable inrush current to deal with than before.

Current Control

To prevent overcurrent from the power supply on the low resistance end of the heater design, I used a current-feedback buck converter. This is a typical buck converter, but the feedback pin is hooked up to a current shunt and non-inverting amplifier so that the feedback pin achieves its target voltage when the current output is high enough, not when the voltage is high enough. This type of converter will try to maintain a constant output current by servoing the output voltage until it hits the supply voltage. So for a 28v supply, it will try to deliver 5A until close to 28V.

This design is fairly optimal because at low heater resistances, 5A still delivers a lot of power. the heater power is P =I^2 R, so as R increases, and I^2 is held constant, the power actually increases with increasing heat – this is good because a hotter plate takes more heat to maintain, since heat creates more convection and higher rates of cooling. This also delivers enough power at low resistances to “kick start” the heater and increase the resistance.

The design is not truly optimal in terms of always delivering maximum power because the buck is not delivering 140W at the lower resistances- remember, the delivered power scales with R. A buck converter can trade voltage for current, so at the low resistance levels it could deliver more current, but in this case that is not necessary, and it would mean bigger magnetics and more expensive parts, as well as more control system complication.

One Temp Sensor to Rule Them All!!

One thing I didn’t like about the last design was that there were two independent temperature sensors on the heater – one for a dedicated overtemperature circuit, and one for regular temperature feedback. Both were NTCs, and its irritating to design a circuit for NTCs that detects both overtemperature and sensor loss (disconnection) conditions – it ends up needing multiple window detectors, and then those need to be properly combined and latched. And while having two sensors is perfectly safe, it seems like losing the measurement sensor and relying on the plate getting too hot before shutting down is silly – if any sensor is lost, it should turn off.

The NTCs also had a linearization problem, requiring a bank of resistors to make temperature measurements over the range I needed to measure. This also creates challenges around having a reliable overtemp limit, because the resistance changes very rapidly. Because of this it is hard to set a tight limit, because being off by “a little” either in the resistor bridge or trigger voltage creates a big swing on the overtemperature cutoff.

I finally realized my vision for an elegant temperature measurement and cutoff circuit with a RTD. There are three main elements to the circuit – a current source feeding the RTD (R18), a level shifting/range mapping circuit, and a latch.

Current Source

This is just a run of the mill application of a current source IC to drive the RTD. The source provides 1mA, and the voltage at TP4 ranges across 1 V to 1.9 V across 0-250C. This minute voltage swing is tackled in the next block, but the important thing is that if the sensor were to become disconnected, the voltage would land on the same side as an overvoltage- meaning that one side of the voltage limit (instead of two) can be set as the error state.

Range Mapping Circuit

The problem with the current source is that given my 10 bit ADC, the roughly .9v of voltage change across my 250C temperature swing only has about 300 ticks- that’s a little less than one C per tick. I wanted more resolution, so I used what I am calling a range mapping circuit to project the range [1.0, 1.9] onto [0.3, 3.0]. This gives me about 2.7x the resolution. This circuit is neat because it incorporates both a voltage subtraction and a subsequent gain. I chose the output range to avoid being near 0 to avoid any weird behavior near the op amp rail, and I left .3v on either side to avoid any poor ADC performance near the extremes of the range, although I wouldn’t expect much trouble with this particular ADC, unlike the original ESP32 ADC that it replaces (which is bad at the extremes).

The circuit is explained in great detail here, but its worth having a quick explanation here. Basically:

V_{out}=V_{in}(1+\frac{R_{then}}{R_{13}})-V_{Vthenivin}\frac{R_{then}}{R_{13}}

V_{thenevin} = V_{pos}\frac{R_{22}}{R_{22}+R_{22}}

R_{then}=R_{22}||R_{23}

This is solved by solving a system of equations at the operating points, it is convenient to use the high/low input/output range for this.

Latching Circuit

Once the overtemperature voltage at the inverting input reaches the limit set by R7 and R8 (about 3v), the output snaps low. This lets D2 conduct, and the non inverting input reaches about .3-.4v or whatever the diode Vf is. This keeps the circuit locked out, until the output voltage of the sensor gets very low (near 0 C), or until SW1 is pressed. This signal, nOverTemp is anded later with the power control signal to block it in case of an overtemp event.

Case Details: Snap fits + Expansion Joints + Mechanical Details

In addition to all the other improvements, I also focused a lot more on the construction of a case for this project. The last hotplate had a lot of issues with warping, softening, and generally tearing itself apart with differential thermal expansion. I am hoping that printing this in ABS will prevent the baseplate from softening too much, and causing warping. So far so good!

Overall case Joining Strategy

3d printing a box can be more challenging than you would think, once all he requirements are added up. In this case, I needed a very low profile case that when “open” kept the heater element low on the desk and that was “tight” to the heater profile. This helps to prevent parts from being dropped and lost under it, and it lets you get your hands right up to the plate when doing assembly on the plate itself. The case also needs exits on two sides for protruding connectors, and a sliding lid for the heater area.

One case style I’ve used before, and that seemed promising in terms of a sliding lid, was a case that slides together, meeting in the middle, perpendicular to the PCB. However, last time I did that, only the front and rear had connectors/stuff that stuck out, and I had none of the thickness constraints.

What really helped to keep this feasible was a dovetail along the length of the case. In the back, permeant half of the case the sliding lid dovetail ends in a square slot with a square snap. Its only ~1mm deep, so it can be snapped back off with some bending of the lid, but that wont happen by accident since it requires a lot of asymmetrical force – something that wont happen if you pick the case up. In fact, pinching the rear of the case holds these snaps in even more.

The front lid is also held in by a similar dovetail and snap, but the front snap has a ramp on both sides of the snap and in the groove. This makes it easier to print, since there is no sharp overhang, and it makes it easy to slide the case lid on and off, while retaining the lid.

You can also see some other details in this cross section – namely the reveal around where the case parts come together, and some relief/clearance in the back snap joint. The front snap joint has less clearance because the tapers will fit even if they don’t totally bottom out, and I wanted the joints to remove all play- but the rear joint needs the square snap fully seated in the hole, or it will bow the whole case out.

Flush Mount Heater and Expansion Joints

One detail to notice here is that the mounting to the baseplate is done flush to the surface of the plate- there are no screw heads to bump into. This was achieved by tapping the aluminum PCB and then screwing it to the screw through the bottom, through the standoff, and into the PCB itself, and then cutting off the nylon screw head flush with the plate. Putting the screw all the way though the threaded standoff creates a lot of friction. This lets me leave the screw head fairly loose in the counterbore, while still keeping the whole assembly captured to the plate.

This loose screw serves as an expansion joint, allowing some “float” of the heater element when it (or the baseplate) expands. Hopefully under constraining the system will prevent stress on the plastic when hot, avoiding creep/plastic deformation. I am not sure if just the heat or heat and expansion killed my last plate, but it will be interesting to see how this one holds up and if this part actually needs to be metal.

Magnetic Feet

Finally, the case has a cover and a flat bottom, to keep the electronics safe and to allow for vertical storage on the bench. As a “nice touch” The base also has magnets in it, so it could be hung up on a metal tool board or for extra grip on a steel plate like the one under my microscope- this should help prevent it from being knocked around, and should keep it handy for assembly tasks.

Smooth is Fast: G3 continuous curves for better 3D prints

October 1, 2026Leave a comment

The future is weird. This morning, I used my timeshare supercomputer subscription to (somehow) guess how write a plugin for a piece of modeling software I don’t own (you are welcome, autodesk). This little piece of code solves a problem that people have been aware of since the beginning of design, but that is hard to put your finger on. To put it less literally, I vibe coded a g3 curvature-continuous sketch fillet tool for fusion. You can get it here. If you care to know more about the problem, or if this table edge bothers you and you don’t know why, read on:

Curve Continuity

The edge of the table above is (supposed to be) made of two lines coming into an ellipse. Even though the lines are tangent to the ellipse (at east to the best of my ability), they look kind of strange. That is because an ellipse (or more commonly a circular arc) has curvature, and a line does not. So even for a tangent line going into a circular radius, there is a discontinuity.

Here is an image of what’s going on – the transition is from the line (no curvature) to the ellipse, which does have curvature. That jump is something that you can see and feel. The little lines are called a curvature comb, and the length is related to the curvature where it touches down, and they comb lives on the “outside” of the curve.

The fix is to use a G2 or G3 continuous curve. Note that both combs now start at zero curvature and ramp up. The difference is that a G2 will have a constant initial “velocity” (first derivative) of the curvature, while a G3 curve accelerates that change over time. So a G3 is smoother.

How this helps my 3D prints, and other things

The curve on the left is a radius, and on the right is a G3 curve. You can see the very tiny shadow on the radiused side, while the G3 curve is smooth. I think this is both a print artifact from deceleration and a result of my careful lighting, but based on the geometry I also expect a little discontinuity there. I can also feel a slight “hitch” in the smoothness as I slide my finger around the curve.

This is a different corner, where the radius is on top and the G3 curve is on the bottom. You can see where the deceleration happens in green, as the print head cannot move smoothly around the corner at high speed.

Unsurprisingly, these smooth curves have tons of more important uses in the real world, particularly in transportation, where smooth curve transitions are used to prevent trains from derailing, cars from flying off of roads, and (least importantly) they are famous for being the non-radius shape for corners that apple uses. They are also used by industrial designers to terrorize mechanical engineers everywhere. There is nothing quite like getting a model with no flat surfaces and no radii that explodes when you sneeze at it, and these curves seem to be amazing for building that kind of geometry.

In my own life, I might start using them to replace radii in my 3d prints, at least when they are in the xy plane of the printer, so they come out extra nice, and ever so slightly faster. I also intend to use them on future woodworking projects to avoid the odd looking corner that I shared above, which is the real reason I made this plugin.

StenciLaser – How to cut solder paste stencils on the Xtool

July 21, 2026Leave a comment

After my initial success with cutting metal stencils on the xtool F1 ultra, I realized that the re-ordering and careful setup of the stencil by hand would really put a damper on my stencil making enthusiasm. Having to re-order hundreds of vector paths by hand is time consuming, and its not easy to visualize and certainly difficult to optimize. Fortunately, this is a fairly straightforward problem of ingesting a file, doing some algorithmic stuff to it, and looking at the output. The perfect kind of problem to take on with my new pal, Claude.

What StenciLaser does:

StenciLaser is a program that takes a DXF and creates an .xs or .xcs file for the F1 ultra where the cuts are ordered to minimize heat affected zones and warping. These thermal effects happen when the laser focuses a lot in one area. For thick substrates, its not as much of an issue since stress will just build up in the part, and be relieved later when it cools. However, with thin parts, the heat is enough to cause permeant deformation of the the part.

To avoid cutting a lot in one place, we need each cut individual cut not to warp the material, and then cuts also need to be spread out as much as possible. I can’t claim that the approach I took will never warp a stencil, but its pretty good. StenciLaser takes all the apertures and breaks them into groups of equal quantity, then goes round robin around the groups and chooses one aperture to cut. The chosen aperture is the furthest from the last cut, and is weighted to also avoid the last cut inside the group. The round robin chooses the furthest group from the last cut.

Finally, the program packages it all up into an .xc or .xcs file. One thing that seems to work, but that is not officially supported is infinite layers. This is important because the only way to get the xtool studio to respect the cut order is to go to processing->cut order->by layer. With only 16 layers (the default) xtool studio will not fully respect the optimized cut ordering, which is a big problem.

Issues+Future work

One issue with this program, possibly due to the infinite number of layers, is that processing in xtool studio takes a long time, especially with many repeats. I found it easier to set repeats to 10-50 and just press the button multiple times rather than wait forever for the processing to be done. This step can take several minutes, but since it happens in the xtool code I can only guess why it takes so long. Possibly it struggles with so many layers, or maybe the number or unique cuts and “travel moves”.

The algorithm is also not perfect. There are some trivial degenerate cases where it might produce bad output, like a very dense grid of rectangles, or having very few cuts to do that are close together. One issue is that the algo is “greedy” and might run out of good candidates while maximizing distance in round robin, instead of maximizing average distance. So far the greedy distance algo seems to be ok in most cases, but it could be an improvement to change to maximizing average distance.

The other issue is that there is no way for the program to “pause” right now. So if there are two rectangles right next to each other, they will just get cut one after another, even though it will cause a lot of local heat. There could be a solution to this too- even though there is no “pause” to put in, a large path at a low speed/intensity could be drawn to waste time. However, this hasn’t been an issue since usually a stencil has a lot of cuts in it.

There are cool improvements to be made too! For example, laser cutting stencils can leave burrs for a couple reasons- either from melt-freeze or from “tabs” that get broken either from heating/cooling and expansion/contraction, or from later manual removal. It would be interesting to add a deburring or even an aperture (vertical) keystoning pass similar to the “aperture enhancement” that is sometimes offered on commercial stencils.

How to get and use StencilLaser

Go here: https://github.com/Aylo6061/StenciLaser. There are instructions for use there. Be sure to set cut order->by layer in the processing menu of xtool studio! or it won’t work.

What I think about Claude Code

This is the kind of problem that in the past, would have required a lot of work on my part that I have historically not been great at, or very interested in – setting up a GUI, making sure the data stays synced with the gui, making sure buttons are the styled right, making sure dividers and cells resize nicely etc. In this case I was able to get a tested prototype done in an afternoon, which is incredible. It even has an ok installer, which should make it easy for me to share this, and which is something I would have avoided making in the past.

There are many complaints about various AI programming tools, but I’ve been overwhelmingly impressed. I think this technology will really let people write super niche software, to solve really specific problems for specific users way faster than anything else ever has. More importantly, I’ve taught a few non-technical people to vibe code, and they have been able to develop software for their own problems. It will be exciting (and scary for some people, like IT admins) to see what people do with this new tool.

PCB Stencils (or tiny metal parts) on the Xtool

July 14, 2025July 22, 2025Leave a comment

Lately I have gotten access to an xtool F1 Ultra. One of the advantages of this laser is its ability to ablate (I wont say cut) metal. I’ve seen videos of people making pcb solder paster stencils, or cutting really thin metal on similar fiber lasers, and I wanted to try too. All of this is applicable to making small “positive” parts with the xtool as well. Here are my notes.

Settings and Cut Ordering

When working with thin material, its important not to heat the metal up too much in one spot. This means making many light passes is better than a single, really aggressive pass. This is especially true for cutting out tiny windows, or cutting out a thin “web” between two windows. Too much heat too fast will distort the final geometry.

The settings I used were 100% power, 8000 mm/s speed, and a variable number of passes. For a single large cut/window (like a ~3mm square or above) you can just go ahead and set it to 500 passes and let it rip. The cut is long enough that by the time the laser comes around again, the heat has had time to conduct away.

Below about that size of window, I had better results from repeated cuts – e.g. cut 50x times, wait 10 seconds, then repeat 10x for a total of 500 cuts.

In addition to taking many passes, I needed to semi-randomize the ordering of the cuts. The Xtool sofware optimizes for time, by cutting from the top left to bottom right. This means that cut profiles that are next to each other together are done at the same time. This leads to overheating local overheating, and warping. The left stencil was randomized, the right one was not.

Sadly, there is not feature to do this in the xtool studio. The best you can do is select->ungroup your stencil, then go through and assign each cut in your stencil to a different layer. Then, using the “…” menu near the process button, choose processing path-> user defining and choose “by layer”. There are about 10 default layers, so the laser will go “somwhere else” for about 9 cuts before returning to where it was cutting before. This greatly reduces local heating.

Workholding to Prevent Warping

Warping happens when some part of the aluminum gets so hot that it expands, and causes plastic (permeant) deformation in the rest of the sheet of aluminum. To avoid this, I constrained the sheet with tape. This prevents deformation in exchange for stress in the aluminum. By minimizing heating, we minimize this stress.

This fixturing also holds the aluminum material so it doesn’t move due to the exhaust fan.

The best way to achieve this fixturing seems to be an aluminum block (safe to laser), then a layer of removable double sided tape, then another layer of painters tape, then the aluminum stock. Without the layer of painters tape, the aluminum card is hard to remove without bending. The painters tape releases from the doublestick pretty cleanly, so a single doublestick tape bed can last many laserings.

Focus Focus Focus

One issue I have run into repeatedly with this laser is that the autofocus does not work, at all. I guess its better than nothing, but I find that its off by about -3mm-4mm. So every time I create a new fixture, turn the machine on, or jog anything, it needs refocusing. I just start by taking the autofocus distance, subtracting 3, then I focus by .5mm increments until I go through a “sweet spot”. It seems like for metal ablation, the beam is tight enough to be focused over about 1.5mm

Surprisingly, the laser cutting makes a distinct snapping/humming sound when its cutting well. I assume this has to do with the metal rapidly expanding/contracting as the laser heats it up and deforms it.

Stencil Performance

My hope is that this can make stencils for everyday use, and for emergency use, like pcb rework. PCB stencils usually cost a couple bucks and cause a few cost related problems for ordering PCBS – specifically in the US, the stencils are relatively expensive (if coming from somewhere like sunstone), or you have to get them at a separate website with separate shipping (osh park/osh stencils), or if you are ordering small boards from overseas, stencils make shipping really expensive as dimensional weight goes up (the box is bigger).

This stencil was quite usable, but it seemed thick (around .18). In the future, I’ll try a slightly thinner material to reduce paste application, especially for “middle”/large pads. The good thing is that its easy to delete or reshape the windows for components if they cause trouble!

Lingonberry – First Leaf!

November 4, 2023November 8, 20231 Comment

UnbeLEAFable! This is a tiny leaf growing from a callus culture, produced from a tiny cutting harvested from the wild. This is exactly what I expected (and hoped) would happen with this project. In a few more months, I hope to have rooted plantlets that I can transfer to normal soil media for hardening! Once the plants are rooted in soil, I will write up a detailed procedure on the whole process.

Wide Temperature Range Thermistor Use

October 27, 20231 Comment

Reading a thermistor seems like a pretty straightforward task, and there are a lot of guides on how to do it, either using just the beta value, or using the full set of coefficients for the Steinhart-Hart equation. I wanted and needed to do something a little different, because I wanted to measure temperatures somewhat accurately over a wide range (25-200C) with a single thermistor, for an upgraded version of the PCB hotplate.

Measuring this range is difficult because the resistance of the thermistor changes a lot over the range I care about. at 25C, this particular model is 100k ohms, and at 200C it is about 1k ohms. It is very not linear, changing very rapidly at first. Thermistors are often used in a voltage divider, where the output voltage is related to the temperature of the thermistor:

This causes a problems for my implementation because choosing other resistor in the divider causes the most sensitivity near the value where Rt=R2. That is, the temperature sensing will be easiest near where the thermistor and the other divider resistor have the same value.

This is a chart showing the expected output of a voltage divider with three different resistor values, computed from the temperature-resistance chart plotted above. The steeper the curve, the better, since a small change in temperature will create an easily measurable output. But none of these curves are linear across the whole range that I care about, from 25-200C. Instead, the 100k looks good from 0-50C the 10k looks good from 50-120 and the 1k looks good from 120-200. At the center of each of these ranges is where Rt=R2, and the output of the voltage divider is about half the input voltage.

My solution was to use several different resistor dividers instead of a single resistor divider. The final implementation uses an analog mux instead of discrete FETS, and included a 100k resistor divider as well.

Calibration

While it would be possible, through first principles and careful measurement, to figure out some mega-equation for all these components, I decided it would be a lot easier to write a script to calibrate each “channel” of the measurement. These thermistors are calibrated against a thermocouple, which was the original sensing element on the hotplate. To collect calibration data, I just measured the hotplate with the thermistor and thermocouple taped together. This produced the curves above, as expected from the simple simulation. Shown in black dots are fitted curves- the bottom two fit a sigmoid/logistic curve, and the top one (which didn’t have enough data to fit to a logistic curve easily) is fit to a 4th order polynomial. The 100k resistor values (green, top line) were pretty useless above 40C, quickly flattening out.

These curves give us a value that predicts ADC ticks for an input of temperature. The inverse is really what we are after, but that is easy to do mathematically.

Looking at the two remaining 10k and 1k curves, the question is where to swap from one to the other. The goal is to keep the amount of ADC counts per degree C as high as possible over the whole range. This can be found by inspecting the derivatives of the fitted curves and finding where they meet (in this case, just around 120C).

This chart also shows why exactly a single value for R2 in the resistor divider would be bad. For example, below 40C, the 1k resistor shows less than 1 tick per degree C, so a 1 degree change would be hard to measure there.

Limitations

This was done with a single set of data, so it may not be accurate for all time. Taking more data in the future would be neat, especially because the python script for analyzing the data will just spit out numbers. It would be interesting to compare the calibration coefficients for various data sets to figure out how much they change, run to run. Hopefully it would be a small amount.

Solution Cost

The reason I didn’t go with a thermocouple is that the reader and the thermocouple alone would cost about $10 in parts. I usually buy multiple ICs per prototype (in case of an accident) which would cost something like $15-20 in parts. The parts for the equivalent thermistor solution cost about $2.50 per unit, which is cheaper than just the thermocouple reader IC.

Lingonberry Cell Culture – Rooting

October 21, 2023Leave a comment

I have reached the point where I have enough lingonberry cell mass to try rooting a few of them. This is done on different media, and since I want to have these develop into plants with leaves and roots, I can’t do it in a petri dish!

Instead of using a purpose-made plant growing container, or even a polypropylene takeout container, I am using some condiment containers I have left over from the original cell culture experiments I did a few years ago. These are half polypropylene, which is autoclavable, and half polystyrene, which is not. To use these, I autoclave the bottom half and soak the other half in bleach while the autoclave is running- when the parts come out, I pour media and fish a lid out of the bleach. it seems to work well enough against contamination.

The media is woody plant media with MS vitamins and zeatin, per the recipe in Plants from Test Tubes. I poured a few CM into each cup (these are large condiment cups, not the tiny ones).

During rooting, it seems important to have a source of light so that the plants can start making their own sugars. These are growing in a basement, which is roughly “room temperature”, but there is not much light down there. I built a rooting tray with overhead lighting for the plants. It is mostly built off of the hardware for the “SunDown” lamp that I built. The hardware is a little overkill, but modular enough to be recycled into future lighting projects.

Lingonberry Callus Cell Culture

September 22, 2023September 26, 2023Leave a comment

I’m back at trying to propagate a lingonberry plant from a cell culture gathered from the wild! This time, I seem to be having a lot more success in terms of growing something that at least resembles a callus culture. Encouragingly, I have had similar growths from various parts of plants that I plated. I tried to excise buds from the meristem, without magnification, so its hard to say exactly what tissue the cells came from, because they were very small (generally, they are from in the buds).

In the future it might be worth looking at callus induction from leaves, a technique I did not know about.

These cultures are growing relatively slowly (months between re-plating), possibly due to the totally uncontrolled and unoptimized conditions they are in, aka an funky basement. Surprisingly, contamination has not been a huge issue (transfers have been done in a pcr hood).

The next step will be to take some larger cell masses and try to get them to grow a bud, shoot, root or leaf- to start turning back into a plant. My plan is to grow them on a deeper media in a different container, undera lighting, without 2ip on something like WPM with zeatin (although zeatin is expensive, so maybe just WPM!).

DIY Canister Dive Light – Bringup

July 7, 2023Leave a comment

With the dive light parts in hand and assembled, it was time to test it out and fix all the inevitable mistakes. I expected this to go pretty well, and bringup for most low power functions was a breeze- however, at higher currents, this thing is a pretty good heater. This led to buying a fun tool – a thermal camera!

ESP32 bringup:

This was my first bare-chip-down ESP32 design, and it was actually pretty easy to do. I used the variant with internal flash to save room, so the only external components were a crystal + caps, some passives (R’s and C’s), and a switch for the RESET line, and a switch on the BOOT (gpio 9). However, there are two additional strapping pins (GPIO 0 and 2) that needed pulldowns. This was annoying since these pins were not pulled down on the pcb – some of them were pulled up, so the parts that did that needed to be removed, and the functionality of those circuits needs to be checked some other way.

I think in the future, it would make sense to avoid using the actual boot and reset buttons and use something like the ESP-PROG to toggle all the lines. Needing to push buttons to upload code is a drag!

I also consistently had issues soldering the QFN package the ESP comes in- I think this was due to too much paste on the center thermal pad. This causes the chip to float, and sometimes this causes one side of the chip (or a few pins) to lift. Next time I will reduce the amount of paste on the center pad to fix this, but I resolved it here by removing the chip, solder-wicking the thermal pad, and then re-soldering the part. Using low temp solder paste made this really easy.

LM3409 Driver bringup:

The LM3409 bringup went relatively well, but thermal issues started to crop up pretty much immediately. There are some surprising deficiencies in the layout and parts that I will be modifying for future revisions. The thermal camera really helped to visualize the issues quickly, and given the number of thermal/power projects I have worked on recently I am sure I will get to use it quite a bit.

PFET Gate Charge

One hair raising issue was that during bringup, the board started to smoke. This is almost always a bad sign. In my original estimation, this was caused by the buck PFET getting hot due to having a high gate charge. The odd thing is that the freewheel diode also got really hot, and stayed hot even if the led was very dim.

A very basic thermal investigation with two thermocouples showed that the FET was getting hotter than the diode, at least initially. However, I did smoke several diodes in the process.

original fet gate voltage. roughly 200 ns of switching time.

Reducing the gate charge did appear to solve the problem, and there was even a note in the datasheet (that I glossed over) about keeping the gate charge below ~30nC. Having a high gate charge is problematic for two reasons- it keeps the FET in a high RDS region for longer during switching, and it requires more power from the gate driver on the LED controller, which also heats up. However, this was not the real problem.

Overheating and Limp Mode:

Note that initially P1, the driver heats up to ~70C. Once the driver gets hot, P2, the diode, gets VERY hot (150C!)

Gate charge was not the entire story. With the FET replaced with a lower gate charge/higher RDSon model, there was still an issue! With the LED aggressively cooled (to allow it to live), the device would still go into some kind of limp mode after a few minutes of operation at high currents (3.6A avg, 12V, about 42W).

This seems to be due to the driver itself overheating, which is due to my poor thermal design. The driver is marked U in the layout above- this is supposed to have a thermal pad connected with vias to ground. I omitted the vias, essentially insulating the chip- no good.

Second, at high currents, the regulator needs to switch frequently to keep in regulation. Estimating from the gate voltage above (in terms of time), and assuming maximum 1A gate drive currents, I estimate that the total switching time is at least 100nS with a period of about 1375nS. The average voltage is about 16.5V, so 16.5V*1A*D = ~1.2W. This would probably be easy do dissipate if I hadn’t insulated the chip. Reducing gate charge helped reduce the power this chip was dissipating, but it was not the root cause.

To make things worse, I didn’t add much copper to dissipate heat on the FET, and at least half the copper is on the side of the driver, which actually makes things worse (its easier to for the FET to heat the driver). Adding some vias to a large plane on the back, and increasing the area on the front will increase the area the heat can easily dissipate from.

The reason the diode gets hot is because usually, the diode only free wheels for a short amount of time- when the FET is off. When the driver goes into limp mode, it has to dissipate all the energy stored in the inductor, many times a second.

So the full story is- this driver will self destruct if the driver overheats and goes into hiccup mode. Better thermal design (and eventually cooling it in the ocean) will probably improve the situation, by preventing driver overheating.

A wiser choice of package for the FET might also help dump heat to the PCB, which will be important when this is running in an enclosure. Convection won’t be an option then, and all the heat will need to go into the case.

Efficiency + Thermals:

With the driver sorted, the next thing to do was to smoke test the lamps at high currents. This went reasonably well, with only one LED burning out. As you can see above, the XHP70.3 required a heat sink and passive water cooling in order to prevent burnout. Given that its running at about 42W and only ~25% of the energy is turning into light, this is a 32W heater in a 70x70mm footprint.

This is going to be very hard to keep cool, even with the die almost being in contact with the seawater. The good news is at lower settings (around 2A) I should still get a very respectable 3000 lm out of the LED- more than the sola light!

Overall efficiency from power supply to LED was acceptable- in the mid 90’s for lower currents up to about 1400mA, and in the high 80’s above that. while that might seem like poor efficiency, this takes into account relatively long lead wires to and from the pcb, the multimeter, etc, so I actually expect it to improve in the final product, especially once the thermal design is improved.

Conclusion:

This PCB needs to be re-routed with thermal concerns and strapping pins in mind, but overall, it works! This iteration has given me a lot of confidence that the final design will do what I want it to do. I’m also really excited to keep playing with the thermal camera. It gives me a lot more information than the sizzling finger test, and it lets me dig into the context and sequence of the heating.

Sola 2000 Teardown

July 5, 20231 Comment

My beloved sola 2000 has finally, after many years of service (and after buying it secondhand) flooded. High power LED drivers and seawater do not mix well, and this light is at its EOL.

I have always wondered what was in this wonderful light, and now that its toast I don’t have any issues tearing it down to find out! There are a lot of very clever design decisions in this light, and a lot to learn.

Light head

The light head has two rings of LEDS- one for spot lighting and one for flood lighting. the small inner ring of three LEDs is the spot light, and the outer LEDS are the flood light. Given the package size and general shape the LEDS, I assume the inner ones are CREE XP-E2 LEDS, and the outer ones look like CREE XM-L2 LEDs.

from XHP50.3 datasheet- note that 2x current does not reach 2x luminous flux. additionally, 2x current increases Vf, which makes it even less efficient than expected from this plot

Having multiple LEDs is very smart, because LEDs run more efficiently (in terms of light per heat) at lower currents. It also spreads out the heat loading of the PFB- CREE says to estimate 75% of LED power to turn into heat instead of light- that means that a 4W LED needs to dissipate 3W of heat.

Given the stated lumen output of the light, I would guess the flood lighting runs around 9-10W, or 1A at about 9V. The ring of 6 flood LEDs probably run at about the same current (1A) but at about 18V, for an output of 18W. This means that about 7-14W needs to be dissipated. This seems to be done through good contact of the aluminum PCB with the metal ring that goes on the front of the light. If you look at the PCB photo above, you will see a smear of thermal grease along the edge.

Running multiple LEDS efficiently means spreading the heat around, especially for the radial LEDs, which are closer to the heat-sinking bezel. Its also important because the light is powered by a relatively small battery pack.

To prevent overheating, it seems like there is a single thermistor on the PCB as well. This will let the controller throttle the output when the emitters get hot.

Optics

The optics look a lot like they are made by carclo (wild conjecture). There are two styles- reflectors for the spot lights, and a total-internal reflection style optic for the spot light. I do wonder if there is any attempt to collimate the spot beam to make it extra tight, by biasing the three spot beams inwards.

LED driver

The LEDs are driven by an LT3755 wide-input rage LED driver. This driver is probably operating in boost mode all the time, since the battery voltage is too low to drive any of the LEDs. Interestingly, there is only one driver (just like in my design) but there appear to be two current sense resistors. I suspect these resistors are switched in for the spot and flood modes (on the high side, with a PFET), and then the overall brightness is controlled by PWM. This makes sense because these LEDS are probably operating at reasonable efficiency, so there is no advantage to turning down the average operating current.

As you can see, most of the corrosion/damage happened near the boost converter. It is close to where water can come in, and it is also where the highest voltages exist on the board. Seawater can cause a short between the current driver outputs, which would then tend to increase the output voltage until the current set point was reached, or until the driver maxes out or reaches some thermal limit- in other words, its a vulnerable circuit.

Flood sensor?

This board has something pretty unusual on it- a big floppy ribbon with conductors on one side. At first, I thought it might be some kind of temperature sensor, but I think it is a flood detection circuit- if it gets wet, it will alert the micro to shut down the led driver. This makes a lot of sense, both to protect the battery and the PCB. If the lamp is dried out after being protected from a flood, I imagine it would be just fine.

Magnetics + Micro

The micro is a pretty basic PIC16F884. The interface is much cooler. There are three evenly spaced IC’s marked “14E” on the PCB. I am guessing these are made (or were made) by NVE, since a lot of their ultra-low-power magnetic switches have that as a portion of the part number. This allows for control of the light without having an extra hole (leak point) in the case.

Power connections

Surprisingly, the power connections are made by wedging the pcb onto the gold contacts in the back of the case. Never in a million years would I have though that this would work so well, but some clever ribs in the back of the housing push the skinny pcb cutouts/contacts onto the gold plated pins.

Autopsy

I’ve always liked the sola lights because they are “factory sealed” and there are no waterproofing components that need changing regularly. Inspection of the front oring didn’t yield any interesting results, but looking at the light pipe seal, I have some suspicions that this might be where the light failed.

The corrosion is right under this seal, although that is also the most likely place for corrosion to happen, so its not a slam dunk. However, this is a circumferential static seal, and there are things I dont like about it.

Specifically, the surface finish of the light pipe is not very smooth on the contact area of the oring (difficult to photograph), and the whole light pipe can rock gently (although this is somewhat prevented by the bezel). The bend radius of the oring is also pretty tight compared to the diameter. The radius is about 2x the diameter, where the best practice would be about 6x the diameter.

Obviously this is a fine design, given that this light has lasted many many years. That said, I am suspicious of this seal.

Closing thoughts:

puffer-palooza at folly cove, illuminated by this light…

This light is pretty tidy from an engineering prospective, and it was a great dive light. I am still curious about what is shared between different models- how is it different than the sola 1500? from the spot lights (the driver here could easily drive a COB)? How did they reuse parts between the designs? And what on earth is that funny three-tier connector for (different models?).

I won’t be getting the answers to these questions but its fun to see what made this thing tick for so long.