jems 2020 mini banner 01

4170 - Neuromorphic computing using spin-wave scatterers

More
7 months 2 weeks ago #239 by jserra
Oral Presentation
Tuesday, December 8
Speaker: Adam Papp
Full Title: Neuromorphic computing using spin-wave scatterers

Please Log in or Create an account to join the conversation.

More
7 months 2 weeks ago #309 by jserra
Session questions:
  1. What is the loss in spin-wave power supplied by the CPW to the power arriving at the detection node? (slide 5)
  2. When I see the highly complex spin-wave beams, I wonder how stable the system is against defects? If the system is trained using the simulations before fabrication, might this be a challenge/problem?
  3. How do you induce the PMA scatterers experimentally?
  4. Does it matter in which direction/with which field the system is magnetized?

We kindly ask the speakers to reply to any questions left unanswered during the talk.

Please Log in or Create an account to join the conversation.

More
7 months 2 weeks ago #329 by AdamPapp
Q1. What is the loss in spin-wave power supplied by the CPW to the power arriving at the detection node? (slide 5)

We did not simulate actual antennas for spin-wave generation, we simply launched a wavefront in the YIG film using a time-varying in-plane field defined in a line. Thus, the efficiency of the waveguide is not included, but these merits are well known for different technologies (efficiencies are generally rather low, we are not aware of a game-changing breakthrough in this field yet). As a rule of thumb, generating a long wavefront is more efficient than many individual inputs, because the same signal launched in a long waveguide will generate spin waves along its whole length, instead only at a single point. Due to the low coupling efficiency, the RF signal level does not decrease significantly along the line.

The waves are propagating in a YIG film modeled with realistic damping, also, we used absorbing boundaries in this case. Reflective boundaries could also be used (finite size YIG film), in which case there would be no energy loss through the sidewalls. The training algorithm in that case would adapt to taking into account reflections from the sidewalls. In the simulations we run so far the training algorithm actually receives a normalized vector of time-integrated intensities from the detectors, so the algorithm is not trying to maximize the output power, only handles the relative amplitudes. For the current demonstrations this was fine, but in a real device the algorithm should also optimize absolute output-power levels. For this we only need to redefine the loss function.

Please Log in or Create an account to join the conversation.

More
7 months 2 weeks ago #365 by AdamPapp
Q2. When I see the highly complex spin-wave beams, I wonder how stable the system is against defects? If the system is trained using the simulations before fabrication, might this be a challenge/problem?

Good point, the algorithm optimizes for the parameters we provide, and thus in the actual fabricated device the parameters might be different, which could possibly destroy the functionality. In fact we have thought about this, let me address a few scenarios here.
  • Most obviously there could be material-parameter deviations. We run a few tests by varying the Ms and bias fields, and we found that there is a reasonable tolerance for a few percent deviation. This, of course could depend strongly on the actual device complexity, so more investigations are needed.
  • Errors in the magnet switching could also be a problem. We did some testing, the effect of a single-magnet error depends strongly on where the magnet is located. In our tests there were only a few magnets that significantly effected thou output loss, most of them had a low effect. We think that some redundancy in the system (using more magnets) could help to alleviate this problem. Also, as most gradient based algorithms, our code can find only local minima in the parameter space. Depending on the initial conditions, one might be able to find many different magnetic configurations with similar performance. Selecting the best according some preset design rules (one of which could be a prescribed error-tolerance level) could be feasible.
  • One might have concerns about the alignment of the output detectors to the magnet array, or one can ask what happens if the waves are focused to a slightly off location due to some fabrication deviations. The shape and size of the output areas can be included in the training process, so one might specify larger regions in the in the training compared to the detector size, so in case of misalignment the physical antenna would still be within the trained output region. Also, one could fabricate tapered fins in YIG at the output side, so detectors would be easier to align to physical structures.
  • Finally, if we would have a fabricated device, maybe we could apply random magnet configurations and record the outputs. With slight modification, our training algorithm would use the known magnet array, and would instead train the material parameters and alignment errors, to find an estimated parameter set for which the given inputs would produce the recorded outputs.

Please Log in or Create an account to join the conversation.

More
7 months 2 weeks ago #371 by AdamPapp
Q3. How do you induce the PMA scatterers experimentally?
The group of Markus Becherer in TU Munich has a long experience in PMA based NML (pNML) fabrication, we have also been part of some of these works. Proof of concept devices are usually programmed by using FIB to modify the switching fields of (individual or groups of) PMA magnets, after which switching can be achieved by global fields. This is of course not practical for integrated devices if reconfigurability is desired, but in that case a lot of technological questions are still open. One can think of using PMA magnets in normal NML operation, where information could be propagated sequentially from one side by a clocking field. In this case no individual electrical wiring would be needed, but some conditions have to be met to have sufficient coupling between magnets.

Maybe the most trivial answer to the question would be putting MRAM-like cells on top of YIG, with the free layer being close to YIG. But I don't know enough about the technological aspects of MRAM, any thoughts on that would be welcome.

Please Log in or Create an account to join the conversation.

More
7 months 2 weeks ago #372 by AdamPapp
Q4. Does it matter in which direction/with which field the system is magnetized?

As I briefly mentioned at the end of my talk as well, the training system is quite general, so anisotropic waves (in-plane, or tilted field) can also be trained. This makes it possible to design optical components for anisotropic waves as well, where analytic solutions are more difficult to acquire (we cannot just copy optics).
This is also important, because large bias fields might not be practical for portable applications, and might not be compatible with other magnetic devices in the proximity.

Please Log in or Create an account to join the conversation.

Time to create page: 0.166 seconds

Organization

INESC MN logo

Abreu Events - Lisbon Office

For general information about the congress, including registration, please contact us at:
 This email address is being protected from spambots. You need JavaScript enabled to view it.
 +351 21 415 6120