Periodic Reporting for period 2 - k-NET (k-space Neural computation with magnEtic exciTations)
Período documentado: 2022-01-01 hasta 2022-12-31
With k-NET, we envisage a neural-network implementation that is qualitatively different from all existing propositions. Instead of designing and structuring individual nonlinear elements (neurons) and their interconnections (synapses) in real space, we propose a different paradigm in which these elements are constructed in reciprocal or k-space of large dimensionality.
- From a theoretical side, a new computational tool to study nonlinear magnetization dynamics in reciprocal space has been developed. This code is capable of computing the nonlinear time evolution of the spin-wave-like modes associated to a given ground state of the magnetic system. When the amplitude of modes is sufficiently small, the dynamics of the system is linear and the modes evolve independently from each other. On the other hand, when amplitudes are excited to reach nonlinear regimes, the modes have mutual interactions that can be computed through evaluation of suitable coupling coefficients. This computational tool is crucial for the prediction and the modeling of reciprocal space based neural computations. Simulations have also been conducted in order to confront the theoretical findings with the experimental measurements in order to progress toward magnetic-modes labelling.
- From an experimental side a series of garnet film processing ultra-low damping have been grown and integrated in radiofrequency devices. Different designs have been used in order to address the modes spectrum using the different spectroscopic techniques available within the consortium. Nonlinear regimes have been reached using either harmonic excitations or parametric excitations. Mode mapping have been studied. Preliminary attempts to selectively populate the selected modes using time-gating radiofrequency signal have been conducted. Micromagnetic simulations were able to reliably predict the complex behavior of the modes in the highly non-linear-regime.
The output of k-NET will be a TRL3 demonstrator based on a micron-scale magnetic device made with an insulating ferromagnetic yttrium iron garnet (YIG). The quantized eigen-excitations of this microstructure will form the neural network (NN) where the spin waves modes will act as neurons and their intrinsic mutual coupling, that can be dynamically tuned, will act as synapses. Radiofrequency (rf) signals will be used to excite SWs and to encode a specific topology of interconnections.
In order to turn our vision into reality, the consortium has identified four objectives:
1. Create the neurons, by defining the appropriate physical properties of the magnetic system;
2. Control the synapses, by tuning nonlinear interactions of SWs via external means;
3. Train the system, by developing protocols to achieve desired computational functions;
4. Validate the technology, by building a working proof of concept to test real cases of object classifications.
Compared with other NN, the challenge of k-NET does not lie in building the physical system and interconnections, but to address and control the ensemble of SW modes. This approach has a high flexibility as the programmed function (i.e. the synaptic weights) is encoded in the time and frequency modulation of rf control signals and is therefore hardware independent. This potentially simplifies the fabrication process and minimizes device-to-device variation, which is a key factor limiting reliability of nano-electronic circuits and allows a single magnetic element to be used to realize a large number of functions.
A successful implementation of k-NET will impact NC by establishing a new field, by developing the science and technology of k-space based neuromorphic computing (k-NC) that has the potential to overcome many hurdles of present hardware-based NC and to allow them to compete with software-NC when a large (speed, energy-efficiency) product of the computation is necessary.