I was thinking of full grind. Obviously won't work if the top section is too thin as you said.
So... need to shave off round patch around 3mm at the bottom. AND find copper base with raised 3mm. Hmmm. 
LOL. I think we're going way above our heads here.
in the xml2 the bottleneck I think is not really the thermal transfer, by improving it you may gain a few percent output by running the die cooler, but you can not drive the led harder. A real bottleneck preventing more current/output is the bond wires that blow before the maximum output of the die is reached.
(but still I am totally for trying this out of course, great stuff here :-) )
You may have just opened up a whole new modding front line Comfychair.
Do you think it would be possible to unsolder the die, direct mount it to copper and then some how wire the positive to the bond wire points (that would be a trick and a half for sure)?
But but but... how do you know the bond wires can't be pushed farther if the die is kept cooler?
Even so, look at the increased output between AL & CU boards at higher currents. Shouldn't direct-die cooling increase the light output even below the level at which the bond wires are at risk?
yes you are right of course, it will improve output at the safe currents as well, and I am curious about how much :-) . But it would be so nice if you can push the amps higher as well, and because the bond wires blow right in the middle of the silicone, I don't think that better cooling of the die will help much.
maybe arms and even breasts too. 
Carry on.
There was a thread on CPF a couple years back where the poster shaved off about half of the substrate and glued the shaved emitter directly to a metal surface and demonstrated that the substrate WAS a major bottleneck.
Regarding the need for an electrically neutral thermal pad, did ya’ll forget about the original Luxeon already? And the SSC P4? The Osram Golden Dragon? They were built on metal slugs and the thermal pads were conductive.
In those days (way back, long time ago ;-) ) there was a dielectric layer under the center pad in the board. With the sinkpads now it is a problem.
Drivers that regulate on the positive side would work fine without anything being insulated, but Nanjgs and the like that do it on the negative side would not.
I suppose a layer of insulation (JB Weld!) covering an area the size of the base of the MCPCB would be preferable to a layer of insulation (the substrate) covering an area the size of the die. A relatively poor thermal material can be pretty good if the surface area is large enough to compensate.
Then the question becomes what's worse, 1mm of ceramic or 1 micron of dielectric.
Changing circuitry and such to match a shared thermal/electrical path seems like the easy problem. Actually separating the chip from the base and connecting to the bond wire terminals seems like the critical challenge.
It just seems if one is to go there, they should pull all the stops and completely eliminate the bottleneck.
Then the question becomes, how would you do that? No matter which design you use, there needs to be a layer of insulation somewhere.
Comfy already mentioned a couple ways. Also, one could just go DD. No insulation needed.
It sounds like from Comfy's reflow experience mentioned above, the chip can be removed from the ceramic base using heat. In one of Djozz's threads, someone mentioned a great idea of putting a small copper bridge under the bond wires. Maybe these ideas can be merged. Get rid of the thermal bottleneck Comfy has shown us and the weak bond wires Djozz has shown us.
You cannot direct mount the LED chip to a metal surface without some kind of electrically insulating layer in between.
Whether its a ceramic or silicon based substrate or a dielectric layer, you need something between the LED chip and the metal host. Otherwise, you’ll get a short.
Milling out a round pocket directly under the die, say 1.5mm/.060" diameter, would give you 98-99% of what you'd get by mounting the bare die directly to a new copper substrate. And you'd probably end up killing 10 of them for each one that worked afterwards.
"But wait a minute," I hear you all saying, "how the hell do you hold the tiny little LED while you mill out a 1.5mm pocket through the substrate?"
Good question!
Well...
Still an early work in progress so ignore the temporary zipties, aiming for the finished product to have a few M2 screws to clamp the two halves together so it can be held in place in the drill press vise.
To set the pocket depth, set the quill stop to zero out on the bottom (upper, since it's mounted upside down) of the center pad, then place shim stock under the vise to raise it the same amount as you want the depth of the pocket. Easy. Right? :bigsmile:
Inserting a metal slug beneath the die, similar to the design of the original Luxeon.
Question 1 is, if the slug contacts the cathode, how will you insulate the slug. And wouldn't the extra insulation offset some of the gains of the slug since the insulation would prevent a direct thermal path to the host.
Question 2 is , will the joint between the slug and the spot under the die be tight enough for the setup to actually work. Knowing how important a tight interface is when going from one layer to another, is this even worth trying if you don't have a way of ensuring a tight bond between the cathode layer and the slug?
Question 3 is, wouldn't it be simpler and perhaps more effective to just shave the substrate very thin, mount it to the host using a micro layer of highly conductive epoxy, and power the LED by soldering to the anode/cathode pads on top.
check out this thread: http://www.candlepowerforums.com/vb/showthread.php?185551
older, but still relevant.
1 - Slug will not be electrically isolated, it'll be soldered to the underside of the cathode plane under the die and to the center pad on the MCPCB, so the MCPCB will then be electrically negative. Not a huge deal to work around, the MCPCB just has to be electrically isolated from the pill. A layer of insulation here will be better than insulation under the tiny little die (surface area and thermal conductivity tradeoffs taken into account, having the worse material cover a larger area gives less negative impact).
2 - Solder. Cathode plane is copper, MCPCB is copper, pin/peg/pillar or whatever you want to call it inbetween the two is copper.
3 - The simpler method is to leave the majority of the substrate intact along with the +/- contact pads so it can be reflowed onto a standard MCPCB just like if it were an unmodified LED, aside from the extra steps for applying solder paste to top and bottom of the 1.5mm dia/.5mm thick copper slug.
I guess insulating the pcb from the host would take care of that problem. That's would be a pretty good work around.
But then you're introducing solder, which has relatively low thermal conductivity, and probably creating a thicker joint than the stock cathode-to-ceramic joint. One thing to keep in mind is that depending on the composition of the ceramic, the conductivity of the substrate is probably not that much worse than aluminum. You're not going to gain much by using a copper post if the joint isn't perfect.
I don't know that introducing additional layers, which is what you're doing by adding the post, is simpler. You want to minimize the number of layers as much as possible.
Interesting! That's similar to how some tissue samples are prepped for microscopy, embedded in resin then shaved with a very very expensive diamond knife. (no I've never even seen an EM in real life, I just read a lot of weird books) :p