Linear Vibration and Mixing in Continuous Spatial Particle ALD

Julia Hartig – University of Colorado Boulder



Continuous vibrating spatial particle ALD reactors were developed to achieve high powder throughput while minimizing reactor footprint. Unlike fluidized bed reactors, continuous vibrating spatial particle ALD reactors operate below fluidization, using linear vibration to convey particles through alternating regions of precursor gas. Fine powder convection and mixing in these vibrating bed reactors is still not well understood, so a cohesive discrete-element-method (DEM) approach was developed to investigate the solids flow behavior. DEM modeling results and experimental images of dyed glass microsphere mixing revealed hopping convection behavior and low top-bottom diffusion during continuous bed vibration. Frit modifications were proposed to induce convection currents and improve surface titration uniformity in deep beds with gas diffusion limitations.


0:00:02   thank you for coming to my talk. My name is Julia Hartig, and today I will be presenting on linear
0:00:07   vibration and mixing in continuous spatial particle ale. De particle ailed is typically performed
0:00:14   and fluid ized bed reactors as shown here, we load our powder into the fluid eyes bedchamber. We
0:00:20   then dose, are precursors in an alternating fashion, separated by purge doses to remove any
0:00:25   unredacted precursor or gaseous byproducts. And this process is repeated as many times as necessary
0:00:32   to produce a film of the desired thickness. I'm showing here the binary sequence for alumina L D,
0:00:38   which is a very common coding, chemistry and L. D. And also the chemistry I worked with for my PhD.
0:00:44   This is a temporal process, meaning that we space out the precursor doses in time. And this is also
0:00:50   a batch process, meaning that we have no solids, inflow or outflow during the A L. D. The challenge
0:00:58   is how do we deal with large volume production. This is typically done by increasing your system
0:01:03   size, which results in higher costs and higher system footprint. You can also use deeper beds, but
0:01:10   this comes with its own set of fluid ization challenges, especially when working with very difficult
0:01:15   to fluid eyes powders. So for high powder throughput, meaning 3000 to 12,000 kg per day, we really
0:01:22   need a continuous process. And this is where continuous spatial particle ailed comes into play. In
0:01:27   contrast to batch temporal ailed where we load our powder into the chamber, perform a Siris of
0:01:34   reactions and then remove that powder. We can remove some of this dwell time by continuously flowing
0:01:40   powder through alternating regions of precursor gas. This way we can accomplish both half reaction
0:01:46   simultaneously just in different regions within the reactor itself, The continuous spatial particle
0:01:53   ale de reactor we have at the University of Colorado looks as follows are particles enter on the
0:01:58   left hand side. He's then pass over a porous space plate through which the gas flows. A first
0:02:05   encounter, a purge region, followed by our first precursor region, which is T. Emma here. Another
0:02:12   purge region, our second precursor region. And this unit, which corresponds to a single cycle of L.
0:02:19   D, is repeated four times along the length of the reactor to complete four cycles of L D. Before the
0:02:26   particles drop out on the right hand side, and these manifolds here dose gasses into the reactor at
0:02:32   velocities below fluid ization. The particles themselves air conveyed through the bed using a
0:02:38   process called Libra Torrey Conviction. Essentially, we have this pneumatic vibrator here, which is
0:02:43   just a self reversing piston that oscillates the reactor bed supported on a pair of leaf springs.
0:02:49   And these oscillations induce net forward motion of the particles through the reactor. There are
0:02:54   some aspects of the particle motion that are difficult to explore or investigate experimentally, So
0:03:01   we developed a cohesive, discrete element method model to provide some insight into the particle
0:03:06   dynamics, in particular in two key areas. The first is to investigate the role of particle particle
0:03:13   and particle wall cohesion. Typical studies on vibrate Torrey conviction have focused on the
0:03:18   conviction of objects and the conviction of course, powders where cohesion is often neglected. But
0:03:23   in the case of particle ailed, where we deal with Fine guild are a and even ultra fine builder Tse
0:03:29   powders. Cohesive forces play a significant role and cannot be neglected, and we're interested in
0:03:35   the effects of these cohesive forces on the particle ailed process. We're also interested in what
0:03:42   form of I batory conviction is occurring here. There are two dominant mechanisms discussed in the
0:03:47   literature. The first is ah hopping conduction mechanism where particles lift off of the fret during
0:03:53   portions of the vibration. The second is a continuous contact conviction mechanism. We're particles
0:03:59   slide forward more than they slide backward, leading to net forward motion. Continuous contact
0:04:05   conviction is more common under non signing soil excitation, whereas hopping conviction is more
0:04:11   common under signing soil excitation. So we need some more information about the vibration wave form
0:04:17   before we can determine which of these is occurring in continuous spatial particle ale de, we also
0:04:23   need to make some simplifications. This reactor contains billions of particles, so we shrink our
0:04:28   domain using a periodic box model. Essentially, instead of modeling all of the particles throughout
0:04:33   the entire length of this reactor, we just looked at a single periodic slice. This d A model still
0:04:40   requires information about reactor vibration. So an accelerometer was fitted to the top of the
0:04:45   reactor to monitor reactor movement. And this data is then used as an input to the D M model. And
0:04:52   what we can see from the acceleration data is that the UAE acceleration, which is along the primary
0:04:57   flow direction and the Z acceleration, which is in the direction of gravity, have some significant
0:05:02   differences. And if we perform a fast Fourier transform on this acceleration data, we can reveal the
0:05:09   key amplitude and frequency components as shown on the right hand side here. So another challenge is
0:05:15   when we're talking about GM modeling and inputs to D M models. We really want to continuous function
0:05:21   for this reactor motion. But this accelerometer data is discreet. So what we can dio is we can
0:05:27   choose the top three frequencies from this FFT decomposition and some them together to rebuild the
0:05:34   original accelerate barometer data as a continuous function for acceleration. So we will call this
0:05:40   the FFT model and you can see this fitting procedure in the figure I've shown is an example on the
0:05:47   why acceleration data. We see pretty good agreement between the model and the original signal. And
0:05:52   if we integrate this once to get velocity and then twice to get position, we see that although the Y
0:05:58   and Z accelerations were quite different, the wind's velocity and positions air fairly similar. And
0:06:04   this is because the high frequency components that dominated the Z acceleration data actually drop
0:06:09   out upon integration. And we also see that the reactor trajectory follows a profile that somewhere
0:06:16   between an ellipse and a line. So now we have a continuous function for reactor acceleration,
0:06:21   velocity and position that we can use as an input to our D M model. And the way we decided to do
0:06:28   this is to treat the Y data as a fluctuating wall. Why Velocity and the Z data as a fluctuating
0:06:34   gravitational acceleration, which avoids the computational complexities of actually physically
0:06:39   having to move the mash up and down in the Z direction. But before we can look at these results, we
0:06:45   need to know what actually happens in the experimental set up. So we replaced the stainless steel
0:06:51   upper chamber with an acrylic flow channel and died 50 micron glass particles to act as tracers in
0:06:57   the flow. And by looking at these experimental results, we can see there's fairly good plug flow
0:07:06   behavior under vibrate Torrey conviction.
0:07:16   We can also see that if we replace these acrylic walls with some stainless steel shim tape, which
0:07:21   more accurately mimics the stainless steel upper chamber that we just replaced. We get even better
0:07:26   plug flow behavior.
0:07:37   So these results give us some context by which to evaluate our simulation results. And what we see
0:07:43   is that the simulations also show plug flow behavior. And by slowing this down to about 10 times
0:07:49   slower than real time, we can actually resolve the conduction itself.
0:07:56   So I'll play that one more time,
0:08:00   and you can see that there are clear regions of lift off where the particles separate from the frit
0:08:05   and contact where they reconnect.
0:08:10   We're also interested in quantifying mixing, and I'll explain a little bit more. Why later in the
0:08:15   presentation. But there's a few different ways to measure mixing. We chose to use diffusion
0:08:20   coefficient as our measure of mixing here. Essentially, what we dio is we track the deviation of
0:08:26   particle position around a trajectory defined by the bulk conduction velocity. And what we find,
0:08:32   which is perhaps unsurprising based on the videos you just saw, is that mixing is fairly slow and
0:08:37   it's nearly Aisa tropic. So these diffusion coefficients for self diffusion are fairly low, and we
0:08:43   see that there nearly equal in both the Y and Z directions, so you may wonder why worry about mixing
0:08:51   in a mono disperse mixture? If we don't have size differences and we don't have density differences,
0:08:57   why would we be concerned about something like self diffusion? The reason for this is because when
0:09:02   we're talking about a reacting system with gas solid reactions, poor top bottom mixing can have
0:09:09   implications on coding uniformity. I'm going to illustrate this using an optimization thought
0:09:14   experiment. So again, the main advantage I would like to remind you of these continuous spatial
0:09:21   particle ailed assistance is the ability to achieve high throughput on when we talk about high
0:09:25   throughput. We're really talking about our mass flow rate or in this situation where I'm showing
0:09:30   volumetric flow rate were upper limited on how high we can drive some of these values. We need to
0:09:36   stay below the fluid ization velocity of our powder, and we also need have some limitations on
0:09:43   packing fraction that limits are void fraction and our width of the plug is limited by the reactor
0:09:49   geometry. So on the fly, the main parameter here that we can adjust is going to be our bed height.
0:09:56   So when we talk about maximizing production, you may imagine that we really want to drive this bed
0:10:00   height as deep as possible. But this isn't the only thing we're concerned about. We're also
0:10:06   interested in ensuring good surface coding quality and what can happen if you have a bed that's very
0:10:13   deep and very gas. Diffusion limited is you could get concentration great aunts in your system. And
0:10:18   if you have poor top bottom mixing, these top layer particles will not see us high of a
0:10:24   concentration of precursor as the bottom layer particles. And if you're precursors own isn't long
0:10:30   enough, these top layer particles may not get to full coating surface titrate Asian. This contrasts
0:10:37   to the idealize mono layer situation. I'm showing on the bottom here, where all particles are seeing
0:10:42   sufficient concentration, or at least the same concentration of precursor. So you can imagine we're
0:10:48   constantly playing this trade off where we want to run our beds as deep as we can to maximize
0:10:53   production. But we don't want to sacrifice our coding quality. So we asked ourselves, How do we
0:10:58   guarantee equal exposure time without sacrificing throughput? In other words, can we induce
0:11:04   convection currents without impeding that forward motion? Is there a way to improve this top bottom
0:11:10   mixing that would allow us to use these deeper beds without having to worry about coding variability.
0:11:17   And so we asked again if improvements can be made here and when thinking about how to achieve this
0:11:24   in the continuous spatial particle A. L D system, we can take some lessons from powder mixers. We've
0:11:29   got ribbon mixers, which I've shown on the left hand side here. They're also rotating drums with
0:11:34   baffles in them to induce mixing. We have V blenders. We also have KNX mixers, which have baffles in
0:11:41   them to separate and recombine the flow. But in general, all of these different mixing techniques
0:11:47   are unified by the same underlying principle, which is called the pastry maker technique. So
0:11:52   essentially, in order to achieve a homogeneous mixture, you want to split re layer and repeat this
0:11:59   process as many times as possible. So we want to keep this underlying principle in mind when
0:12:04   thinking about modifications we can make to improve mixing and the continuous spatial particle ailed
0:12:10   system. And what we decided to try was whether front baffles can improve top bottom mixing. In other
0:12:17   words, can we modify this frit geometry in order to induce conviction currents that would enable us
0:12:23   to use these deeper beds without having to worry about surface coating uniformity. So future work
0:12:28   will investigate the role of this frit buffel geometry. In mixing in summary, we've developed a
0:12:35   cohesive D M model for continuous spatial particle A L D. Experiments and simulations revealed plug
0:12:42   flow with hopping. Conviction is the dominant mechanism, and fit battles will be incorporated to
0:12:47   improve top bottom mixing and future work. With that, I would like to thank the Weimer Research
0:12:52   Group and the onus of goalie proposal funding my work, and I hope you enjoyed the presentation.