Periodic Reporting for period 1 - NeuroMag (Magnonic Matrix-Vector-Multiplier for Neural Network Applications)
Reporting period: 2018-05-22 to 2020-05-21
The most computationally intensive part in a neural network (both in training and operation) are the matrix-vector multiplications (large-scale linear transformations) needed to propagate signals between the different layers. To improve the efficiency of matrix vector multipliers (MVMs), it has been proposed to perform the linear transformations through physical processes instead of digital computation. Key properties of a matrix-vector multiplier (MVM) in artificial neural networks are its bidirectionality (ability to perform the multiplication with the transposed matrix in the opposite direction) and trainability (ability to adapt itself to the forward and the error signal).
Trainable bidirectional MVMs have been suggested based on the interference of waves. The implementation of NeuroMag explored and laid the groundwork towards MVMs using the propagation, interference and phase shift (training) of spin waves. High quality epitaxial Y3Fe5O12 (Yttrium Iron Garnet, YIG) films were developed as a medium for spin waves. The low magnetic damping in YIG limits detrimental effects due to spin wave attenuation. Electrical to magnetic transducers can act as generators, detectors and training mechanism, leading to full bidirectionality and trainability of the MVM.
To reach this ambitious goal, over the course of 2 years, starting from May 22, 2018 and ending on May 21, 2020, the researcher targeted 3 main research objectives:
1. To achieve a pulsed laser deposition process for epitaxial YIG thin films on Ga3Gd5O12 substrates as ferromagnetic medium for the propagation and interference of spin waves.
2. Develop magnetoelectric transducers to generate and detect spin waves.
3. Develop a magnonic MVMs and demonstrate its operation.
To meet these goals, each objective was addressed to a distinct work package (WP1-3) with tangible targets and specifications.
Trainable bidirectional MVMs have been suggested based on the interference of waves. The implementation of NeuroMag explored and laid the groundwork towards MVMs using the propagation, interference and phase shift (training) of spin waves. High quality epitaxial Y3Fe5O12 (Yttrium Iron Garnet, YIG) films were developed as a medium for spin waves. The low magnetic damping in YIG limits detrimental effects due to spin wave attenuation. Electrical to magnetic transducers can act as generators, detectors and training mechanism, leading to full bidirectionality and trainability of the MVM.
To reach this ambitious goal, over the course of 2 years, starting from May 22, 2018 and ending on May 21, 2020, the researcher targeted 3 main research objectives:
1. To achieve a pulsed laser deposition process for epitaxial YIG thin films on Ga3Gd5O12 substrates as ferromagnetic medium for the propagation and interference of spin waves.
2. Develop magnetoelectric transducers to generate and detect spin waves.
3. Develop a magnonic MVMs and demonstrate its operation.
To meet these goals, each objective was addressed to a distinct work package (WP1-3) with tangible targets and specifications.
The work performed within WP1 consisted on the development and optimisation of the epitaxy process and the deposition of ultralow damping YIG films by pulsed laser deposition (PLD). This was performed in tight collaboration with the PLD manufacturer, Solmates B.V. (Enschede, The Netherlands). This goal was successfully achieved within the targeted specifications (thickness range 50–100 nm; Gilbert damping α < 10-3). The characterization methods include X-ray diffraction, transmission electron microscopy and ferromagnetic resonance (FMR). Particularly, the FMR measurements enabled the knowledge transfer from the secondment group (U Paris Sud), with recognised longstanding expertise with FMR, to the host institution (IMEC).
The work developed in the scope of WP2 comprised the development of magnetoelectric (ME) transducers. These devices act as generators, detectors, and training mechanism of SWs, and promise the best scalability and highest energy efficiency of all approaches (e.g. with respect to microwave antennae). The work performed in NeuroMag was a significant step towards the understanding of ME transducers. Particularly, the significant material and fundamental physics challenges still hampering this approach. The material challenges concern the deposition of epitaxial YIG films with a conductive bottom electrode. Furthermore, the deposition of the piezoelectric material Pb(ZrxTi1-x)O3 (PZT) on YIG also depicted a strong impact on the magnetic properties of the YIG film. As for the more fundamental challenge, micromagnetic simulations have shown a very complex magnetization dynamics is generated by the magnetoelectric mechanism. This feature averages out the detected signal preventing the SW detection. These findings will help the ME community to find alternative approaches and geometries that will enable the fabrication of efficient and scalable SW transducers.
The learnings of WP1 and WP2 were applied to the device proof of concept of WP3. The YIG films (WP1) were used as SW medium, and a new strategy was defined using antennae with steps, to simulate the intended effect of ME cells (WP2) as a training mechanism (local magnetic fields). The device functionality could be demonstrated (particularly bidirectionality and training). Furthermore, the collaboration with U Paris Sud allowed the researcher to become an expert on SW spectroscopy and to perform the full knowledge transfer to the host laboratory. In fact, an RF electrical measurement setup incorporating an applied magnetic field was built at IMEC. Both the hardware and software development were performed in the scope of NeuroMag.
The work developed in the scope of WP2 comprised the development of magnetoelectric (ME) transducers. These devices act as generators, detectors, and training mechanism of SWs, and promise the best scalability and highest energy efficiency of all approaches (e.g. with respect to microwave antennae). The work performed in NeuroMag was a significant step towards the understanding of ME transducers. Particularly, the significant material and fundamental physics challenges still hampering this approach. The material challenges concern the deposition of epitaxial YIG films with a conductive bottom electrode. Furthermore, the deposition of the piezoelectric material Pb(ZrxTi1-x)O3 (PZT) on YIG also depicted a strong impact on the magnetic properties of the YIG film. As for the more fundamental challenge, micromagnetic simulations have shown a very complex magnetization dynamics is generated by the magnetoelectric mechanism. This feature averages out the detected signal preventing the SW detection. These findings will help the ME community to find alternative approaches and geometries that will enable the fabrication of efficient and scalable SW transducers.
The learnings of WP1 and WP2 were applied to the device proof of concept of WP3. The YIG films (WP1) were used as SW medium, and a new strategy was defined using antennae with steps, to simulate the intended effect of ME cells (WP2) as a training mechanism (local magnetic fields). The device functionality could be demonstrated (particularly bidirectionality and training). Furthermore, the collaboration with U Paris Sud allowed the researcher to become an expert on SW spectroscopy and to perform the full knowledge transfer to the host laboratory. In fact, an RF electrical measurement setup incorporating an applied magnetic field was built at IMEC. Both the hardware and software development were performed in the scope of NeuroMag.
NeuroMag fulfilled some of its key research goals, however, the development of ME cells (WP2) requires further research. Neuromag has pushed the boundaries of the state-of-the-art in several aspects: the correlation between PLD parameters to achieve high quality ultralow damping YIG films and the measurements at device level; further understanding on ME transducers; and the possibility of using local magnetic fields for SW phase shift to perform beyond CMOS computation. Additionally, Neuromag also leveraged the development of PLD as more widespread tool within IMEC, which is already being used to support several projects. This project also allowed for the knowledge transfer of RF electrical measurements from the secondment institution to IMEC. Furthermore, the main activities of NeuroMag and related scientific/technological developments were successfully communicated to diverse target audiences (general public, students, entrepreneurs and policy makers). Finally, the researcher had the opportunity to develop both his technical (PLD, RF electrical measurements) and soft skills (focus on supervision and leadership) through hands-on scientific experience, complemented with practical dedicated courses. This approach enhanced the future career prospects of the researcher.