wafer.space Community - 🏗️-die-sorter

I have been bad about posting things, sorry about that. First off, the tape release was one of the biggest hurdles from the first run, so this needed to be perfect. It's taken way more revisions than I would like to admit to get it perfect. There's a ton going on in a very small space. I showed a bit about the clamp/stretcher before, but here's another shot. I also have a joystick wired in that helps with interfacing and debugging. Things should start moving a lot faster now that's behind me.

2

1

Andrew Wingate

I have been bad about posting things, sorry about that. First off, the tape release was one of the biggest hurdles from the first run, so this needed to be perfect. It's taken way more revisions than I would like to admit to get it perfect. There's a ton going on in a very small space. I showed a bit about the clamp/stretcher before, but here's another shot. I also have a joystick wired in that helps with interfacing and debugging. Things should start moving a lot faster now that's behind me.

So, the device sucks the backing tape down onto it and then the pushers push up? Do you have control over the different push pins, or do they all go up at the same time? I assume the other things next to it are the UV leds?

Tim 'mithro' Ansell

So, the device sucks the backing tape down onto it and then the pushers push up? Do you have control over the different push pins, or do they all go up at the same time? I assume the other things next to it are the UV leds?

Yep, the tape is sucked down with a vacuum just like the previous version and modeled after the commercial machine i found a while back. No, all the pins go up at once, that's why we were asking about packing in the lower left corner. Now that i have played ( a lot ) could probably feel comfortable designing something with more fidelity, but would probably no longer use steppers. Or at least steppers as the driver. Yep, those are uv leds. 365nm. There isn't any extra space so 2 were needed and one will have to be used depending on the quadrant.

Andrew Wingate

Yep, the tape is sucked down with a vacuum just like the previous version and modeled after the commercial machine i found a while back. No, all the pins go up at once, that's why we were asking about packing in the lower left corner. Now that i have played ( a lot ) could probably feel comfortable designing something with more fidelity, but would probably no longer use steppers. Or at least steppers as the driver. Yep, those are uv leds. 365nm. There isn't any extra space so 2 were needed and one will have to be used depending on the quadrant.

Steppers driving what? FYI, something like https://odriverobotics.com/ can upgrade from the 2-phase NEMA steppers to 3-phase BLDCs w/ encoders doing closed loop field oriented control. Though it probably shouldn't be an issue anyways.

ODrive

ODrive - High performance motor control for robotics.

namibj

Steppers driving what? FYI, something like https://odriverobotics.com/ can upgrade from the 2-phase NEMA steppers to 3-phase BLDCs w/ encoders doing closed loop field oriented control. Though it probably shouldn't be an issue anyways.

The stepper drives a screw that pushes up the pins. The pusher part is just an insert that slides down inside. Given the timeframe, I have to choose what is most likely to work, even though you can see it still took a number of revisions (I don't even think that was all) In the end, the number of pins had to be reduced (kept getting 'bed of nails' effect) and the vacuum needed to be embedded in the tip with some internal chambers as I was having a hard time getting seals any other way. As I've understood the design requirements more, I think I would try for some voice coil/solenoid options for the pins and that would allow for smaller space and fidelity on the for pin selection? That's the idea, but need to move to other areas now. *re: *odrive, I don't think it's not really necessary for what we have at the moment as simplicity is desired.

Andrew Wingate

The stepper drives a screw that pushes up the pins. The pusher part is just an insert that slides down inside. Given the timeframe, I have to choose what is most likely to work, even though you can see it still took a number of revisions (I don't even think that was all) In the end, the number of pins had to be reduced (kept getting 'bed of nails' effect) and the vacuum needed to be embedded in the tip with some internal chambers as I was having a hard time getting seals any other way. As I've understood the design requirements more, I think I would try for some voice coil/solenoid options for the pins and that would allow for smaller space and fidelity on the for pin selection? That's the idea, but need to move to other areas now. *re: *odrive, I don't think it's not really necessary for what we have at the moment as simplicity is desired.

If you think there's practical potential to modify this very design a little and add load cells into the structural backing of those pins (ideally individually) wired out to something that could eventually record and collate to what the machine was tasked with to mostly-passively gather information about the forces actually involved in the pushing work..... While ofc not matching individual linear motor actuation of any particular variety (there's a reason "limited slip" differentials are a thing for 4WD cars....), it'd still give understanding of the forces involved and thus what nuanced actuation in the future would need force-wise/power-wise.

2:35 a.m.

Re: odrive, I thought you may have referred to the motion platform of the die sorter when you eschewed steppers

That's where I'd see such a change to be something to consider if the motorization performance isn't making you happy enough.

namibj

Re: odrive, I thought you may have referred to the motion platform of the die sorter when you eschewed steppers

That's where I'd see such a change to be something to consider if the motorization performance isn't making you happy enough.

speed, accuracy, or power Pick 2 or pull out your wallet

For the vacuum seal I'd suggest to consider a simple cut to outline soft silicone sheet/gasket stuck to this, just saying

namibj

If you think there's practical potential to modify this very design a little and add load cells into the structural backing of those pins (ideally individually) wired out to something that could eventually record and collate to what the machine was tasked with to mostly-passively gather information about the forces actually involved in the pushing work..... While ofc not matching individual linear motor actuation of any particular variety (there's a reason "limited slip" differentials are a thing for 4WD cars....), it'd still give understanding of the forces involved and thus what nuanced actuation in the future would need force-wise/power-wise.

I think you may be able to do something similar with voice coils yes? In the end, I'm not sure how valuable this data would be tho

namibj

For the vacuum seal I'd suggest to consider a simple cut to outline soft silicone sheet/gasket stuck to this, just saying

Is has to be very flat though. You can't suck a die down into it's neighbors

2:38 a.m.

and the seal issue was all 'behind' the head, not on the tape side

well, the operating principle of odrive can give you a lot of speed for the seek while letting you control the deceleration good enough to keep the mechanics from ringing (much) and then still using the encoder resolution to do the fine positioning. If you buy from them it's somewhat more expensive than steppers but you do get more speed offered and substantially higher efficciency. Though ofc regenerative braking doesn't matter as much for this as it does for a battery powered robot.

Andrew Wingate

Is has to be very flat though. You can't suck a die down into it's neighbors

yeah.

Andrew Wingate

and the seal issue was all 'behind' the head, not on the tape side

oh yeah then don't keep it nice hard

Andrew Wingate

I think you may be able to do something similar with voice coils yes? In the end, I'm not sure how valuable this data would be tho

voice coils themselves dont' really give you position feedback

2:44 a.m.

for these pins I'd probably have aimed at solenoid style with a non-ferromagnetic ideally not aluminium levels of electrically conductive front (maybe non-magnetic stainless steel should work great?), and a solid soft iron or at most ferrosilicon (transformer sheet alloy) active chunk; then a coil in similar encasing that sucks the pin forwards and if needed a separate one to pull them back town again.

2:46 a.m.

a non-magnetic tube should give easy winding mandrel and later inner solid holder for the windings that would be done. But without knowing forces it's not at all easy to check if it'll reasonably behave if made a particular way.

2:49 a.m.

(I think there'd ideally be a magnetic tube of similar cross section or slightly above compared to the inner rod's magnetic iron part, around the outside of the windings, acting as shielding and to channel the magnetic field where it's wanted. But switched reluctance motors like those technically would be are vulnerable to kinda-non-intuitive design-effects/behaviors, hence kinda preferring to sim. Thankfully the almost trivial geometry involved would make this part of the design almost trivial to pull off.)

I will have a bunch to share in the coming days, for now I'll just share a snippet of code that basically runs the whole show. I thought this bit was kinda beautiful.

def sort(self, todo):
    """Do the sort
    Machine consists of 3 major components. pusher, tobor, feeders
    Certain components can be working on the next die while waiting on others 
    run_together() allows for this by blocking until both are completed
    """
    for placed, die, upcoming in pipeline(todo, size=3):
        yield from self.run_together(
            self.pick(die),                   # tobor: down onto die N, pins up, lift
            self.feeder_ready(placed, die))   # feeder: past N-1's pocket, right tape, pocket found
        yield from self.run_together(
            self.carry(die),                  # tobor: up camera, align, into the pocket
            self.pusher_under(upcoming))      # pusher: already off under die N+1
        yield from self.checkpoint(die)

-# yes, it's generators all the way down

1

repeatability seems pretty good waferspace

spoke too soon haha. n7 was off for some reason. Looking into it. @Tim 'mithro' Ansell I'm taking pics with wafer name and slot location. The order also says how it's in the tape. Any data that's useful to try and capture?

"WSLG C5R1 S1_3": {
   "state": "placed",
   "at": "2026-09-27 21:46:25",
   "n": 9,
   "project": "WSLG",
   "cr": "C5R1",
   "shot": "1_3",
   "design": [
    51.179,
    69.167
   ],
   "run": 11,
   "picked": "2026-09-27 21:46:09",
   "placed": "2026-09-27 21:46:25",
   "order": 9,
   "photo": "partial1/WSLG_009_n9_C5R1_S1_3.jpg"
  },

Andrew Wingate

spoke too soon haha. n7 was off for some reason. Looking into it. @Tim 'mithro' Ansell I'm taking pics with wafer name and slot location. The order also says how it's in the tape. Any data that's useful to try and capture?

"WSLG C5R1 S1_3": {
   "state": "placed",
   "at": "2026-09-27 21:46:25",
   "n": 9,
   "project": "WSLG",
   "cr": "C5R1",
   "shot": "1_3",
   "design": [
    51.179,
    69.167
   ],
   "run": 11,
   "picked": "2026-09-27 21:46:09",
   "placed": "2026-09-27 21:46:25",
   "order": 9,
   "photo": "partial1/WSLG_009_n9_C5R1_S1_3.jpg"
  },

All of it, all the time!?

4:14 a.m.

@Andrew Wingate - Did one of the die flip over?

Tim 'mithro' Ansell

All of it, all the time!?

haha, yeah, after I posted that I changed it so it basically saves every bit of data that's created. Also, no not flipped, it's just resting on the edge and so is reflecting the light weird. This spotlight light angle gives super good resolving of the pocket and its location, not good for much else though haha. I'm also playing with some verification code that attempts to resolve if a die is in the pocket and in location.

Andrew Wingate

haha, yeah, after I posted that I changed it so it basically saves every bit of data that's created. Also, no not flipped, it's just resting on the edge and so is reflecting the light weird. This spotlight light angle gives super good resolving of the pocket and its location, not good for much else though haha. I'm also playing with some verification code that attempts to resolve if a die is in the pocket and in location.

That lighting kinda looks pretty good for showing if the die landed correctly?

Tim 'mithro' Ansell

That lighting kinda looks pretty good for showing if the die landed correctly?

For humans maybe. Haven't gotten an opencv test to be super good. I'll get there I'm sure. I can see these pretty good, I'm fuzzing off an empty pic. Once there's much more data I'm sure we'll settle on something decent.

Well - I mean the failure case is pretty clear -- not so much the success case?

I'm also saving the pic from the up camera to see if we get anything useful from that. Not sure we ever would, but..

Tim 'mithro' Ansell

Well - I mean the failure case is pretty clear -- not so much the success case?

Yeah, that may be all that matters

4:36 a.m.

up camera for those interested

What's an up dog?

Tim 'mithro' Ansell

What's an up dog?

Not much

Computer vision is hard. I'm getting about 90%+ correctly verified. Hoping to improve with more data.

Andrew Wingate

spoke too soon haha. n7 was off for some reason. Looking into it. @Tim 'mithro' Ansell I'm taking pics with wafer name and slot location. The order also says how it's in the tape. Any data that's useful to try and capture?

"WSLG C5R1 S1_3": {
   "state": "placed",
   "at": "2026-09-27 21:46:25",
   "n": 9,
   "project": "WSLG",
   "cr": "C5R1",
   "shot": "1_3",
   "design": [
    51.179,
    69.167
   ],
   "run": 11,
   "picked": "2026-09-27 21:46:09",
   "placed": "2026-09-27 21:46:25",
   "order": 9,
   "photo": "partial1/WSLG_009_n9_C5R1_S1_3.jpg"
  },

Wdym, "data" that's "useful to capture"?

Andrew Wingate

Computer vision is hard. I'm getting about 90%+ correctly verified. Hoping to improve with more data.

How about some light setup that enhances the effective contrast? A simple polarization filter on the camera and one on the lamp with their angles and the angle of the light (between normal and parallel to the plane of the die) tweaked while looking at the camera feed probably does wonders. Also, consider throwing a tiny variant from a modern AlexNet/MobileNet evolution at the task using some classic binary/ternary classifier layer over it's embedding just trained with manual labels ("good placement"/"too close to edge"/"literally overlapping the edge") and fed with a suitably carefully handled cropping to keep it from looking at irrelevant environment.

Andrew Wingate

up camera for those interested

What do you mean, "up camera"? What's it looking at?