Periodic Reporting for period 3 - SpinENGINE (Harnessing the Emergent Properties of Nanomagnet Ensembles for Massively Parallel Data Analysis)
Período documentado: 2023-01-01 hasta 2024-06-30
Computing power struggles to keep up with analysis demands as data generation keep increasing exponentially. Realizing new platforms for massively parallel data processing, where large amounts of data are processed 'all at once' rather than piece-by-piece, is key to closing this gap. Now, SpinENGINE is combining two cutting-edge concepts, reservoir computing and nanomagnet ensemble dynamics, to realise this vision. Reservoir computing utilises a reservoir with highly nonlinear dynamics that projects input signals onto high-dimensional spaces and use simple linear processing techniques to extract an output. SpinENGINE is using the emergent and tuneable nonlinear interactions in nanomagnet ensembles as the reservoir to create a new massively parallel, computational device.
The SpinENGINE project will lay the foundations for a new, massively parallel, computational platform based on emergent behaviour in large nanomagnet ensembles. The project will develop an efficient, highly scalable, and easily reproducible platform meeting the data analysis challenges in our increasingly data-rich society. We will build upon our recent discoveries and use complex, nonlinear, and highly tunable interactions in such ensembles to realize a hardware platform for “Reservoir Computing”, a biologically-inspired computational approach. Our critical hypothesis is that the synergies between the inherent properties of nanomagnet ensembles and those required for reservoir computing will enable the efficient creation of a highly adaptive computational platform for the analysis of complex, dynamic data sets. This has the potential to greatly outperform current approaches using conventional CMOS hardware. We have defined four key objectives, each of individual scientific merit that will contribute essential functionalities to the computing platform as a whole:
Objective 1—To establish protocols for reservoir computing based on nanomagnet ensemble behaviour, i.e. representing linear independent reservoir transfer functions (x_i (t)) with emergent ensemble properties and memory of past states through ensemble non-volatility.
Objective 2—To identify the most suitable approach for tuning interactions between nanomagnets, i.e. determine the effectiveness of different external stimuli in tuning inter-element coupling with optimized element geometry, arrangement, and materials.
Objective 3—To understand and demonstrate tuneable emergent behaviour in nanomagnet ensembles, i.e. determine the sensitivity of emergent properties of large nanomagnet ensembles to external stimuli as well as lattice geometries, material compositions, local inhomogeneities, and boundary effects.
Objective 4—To demonstrate computational functionality in nanomagnet ensembles. This will require demonstration of data read/write methods and integration of connections for read, write, and tuning operations on a large ensemble.
SpinENGINE brings together a multidisciplinary team of researchers with expertise in computer science, condensed matter physics, material science, computational modelling, and high-resolution microscopy. This enable us to simultaneously explore the fundamental behaviours of nanomagnet ensembles and understand how these can be harnessed for useful computation.
By the end of the SpinENGINE project we have demonstrated that coupled nanomagnet ensembles and their emergent properties may serve as a promising platform for reservoir computing. We have developed simulator frameworks suitable for exploring such systems in simulation, methodologies for fabrication of physical implementations and promoted the use of task-independent metrics for evaluation of their performance as reservoirs. Finally, we have implemented demonstrator implementations, two lab-based and on palm sized implementation with all-electrical input and readout.
The SpinENGINE project will lay the foundations for a new, massively parallel, computational platform based on emergent behaviour in large nanomagnet ensembles. The project will develop an efficient, highly scalable, and easily reproducible platform meeting the data analysis challenges in our increasingly data-rich society. We will build upon our recent discoveries and use complex, nonlinear, and highly tunable interactions in such ensembles to realize a hardware platform for “Reservoir Computing”, a biologically-inspired computational approach. Our critical hypothesis is that the synergies between the inherent properties of nanomagnet ensembles and those required for reservoir computing will enable the efficient creation of a highly adaptive computational platform for the analysis of complex, dynamic data sets. This has the potential to greatly outperform current approaches using conventional CMOS hardware. We have defined four key objectives, each of individual scientific merit that will contribute essential functionalities to the computing platform as a whole:
Objective 1—To establish protocols for reservoir computing based on nanomagnet ensemble behaviour, i.e. representing linear independent reservoir transfer functions (x_i (t)) with emergent ensemble properties and memory of past states through ensemble non-volatility.
Objective 2—To identify the most suitable approach for tuning interactions between nanomagnets, i.e. determine the effectiveness of different external stimuli in tuning inter-element coupling with optimized element geometry, arrangement, and materials.
Objective 3—To understand and demonstrate tuneable emergent behaviour in nanomagnet ensembles, i.e. determine the sensitivity of emergent properties of large nanomagnet ensembles to external stimuli as well as lattice geometries, material compositions, local inhomogeneities, and boundary effects.
Objective 4—To demonstrate computational functionality in nanomagnet ensembles. This will require demonstration of data read/write methods and integration of connections for read, write, and tuning operations on a large ensemble.
SpinENGINE brings together a multidisciplinary team of researchers with expertise in computer science, condensed matter physics, material science, computational modelling, and high-resolution microscopy. This enable us to simultaneously explore the fundamental behaviours of nanomagnet ensembles and understand how these can be harnessed for useful computation.
By the end of the SpinENGINE project we have demonstrated that coupled nanomagnet ensembles and their emergent properties may serve as a promising platform for reservoir computing. We have developed simulator frameworks suitable for exploring such systems in simulation, methodologies for fabrication of physical implementations and promoted the use of task-independent metrics for evaluation of their performance as reservoirs. Finally, we have implemented demonstrator implementations, two lab-based and on palm sized implementation with all-electrical input and readout.
During the first reporting period of the SpinENGINE project, we established the basic concepts and methodologies necessary to carry out the work planned in this project. Most importantly we developed simulators specifically designed to model the nanomagnetic systems under investigation in SpinENGINE and established the methodology needed for physical realization of the proposed nanomagnetic systems. These tools and methodologies have in turn allowed us to carry out the first investigations of tunability of artificial spin systems for reservoir computing.
The team has advanced the development of specialized simulators (Flatspin, RingSim and Hotspice), which now enable efficient design space exploration, e.g. using evolutionary approaches to discover new artificial spin system geometries optimized for different transfer functions. This not only reduces design time but also allows for direct fabrication of optimized designs in the lab.
Additionally, the simulators are powerful enough to evaluate implementations against established, task-independent metrics for reservoir computing such as kernel quality and generalization capability. The project has also successfully demonstrated reservoir computing experimentally, using a single node nanoring reservoir and completed computational tasks with satisfactory outcomes.
Furthermore, the team has demonstrated the ability to read and write data into the two types of nanomagnet ensembles studied and to tune the response of the nanomagnet ensembles using various stimuli. Notably, reproducible, clocked magnetization reversal in artificial spin systems has been demonstrated, which is considered a crucial step in the design of reservoirs with linear memory.
Finally, proof-of-concept demonstrators (TRL4) has been developed and evaluated against tasks such as Mackey-Glass time series prediction and pattern recognition. Demonstrators based on nanoring ensembles as well as out of plane ASI systems have been demonstrated. A palm-sized demonstrator with all-electrical input and readout has been made.
The team has advanced the development of specialized simulators (Flatspin, RingSim and Hotspice), which now enable efficient design space exploration, e.g. using evolutionary approaches to discover new artificial spin system geometries optimized for different transfer functions. This not only reduces design time but also allows for direct fabrication of optimized designs in the lab.
Additionally, the simulators are powerful enough to evaluate implementations against established, task-independent metrics for reservoir computing such as kernel quality and generalization capability. The project has also successfully demonstrated reservoir computing experimentally, using a single node nanoring reservoir and completed computational tasks with satisfactory outcomes.
Furthermore, the team has demonstrated the ability to read and write data into the two types of nanomagnet ensembles studied and to tune the response of the nanomagnet ensembles using various stimuli. Notably, reproducible, clocked magnetization reversal in artificial spin systems has been demonstrated, which is considered a crucial step in the design of reservoirs with linear memory.
Finally, proof-of-concept demonstrators (TRL4) has been developed and evaluated against tasks such as Mackey-Glass time series prediction and pattern recognition. Demonstrators based on nanoring ensembles as well as out of plane ASI systems have been demonstrated. A palm-sized demonstrator with all-electrical input and readout has been made.
The SpinENGINE project has already produced original research beyond the state of the art. We have developed dedicated simulators for dipolar coupled artificial spin ice systems (flatspin and Hotspice) and for nanoring ensembles (RingSim). These models enable efficient simulation of large-scale interacting spin ensembles beyond what has previously been possible and are currently used for efficient design space exploration.
Experimental demonstration of reservoir computing has also been demonstrated using a single node nanoring reservoir and computational tasks such as spoken digit recognition, TI-46, has been completed with good results.
We have shown that we can reliably read and write data into the two types of nanomagnet ensembles that are studied in this project. For the ring arrays, the reading signal is based on the AMR response and for the ASI, a Hall bar is added under the array which exploits the AHE to probe the magnetization states.
We have demonstrated tuning of nanomagnet ensembles using a variety of stimuli, perhaps most notably, we have demonstrated reproducible, clocked magnetization reversal in ASI systems. Unlocking the discrete dynamics of these dynamical systems is believed to be a crucial door-opener for the design of reservoirs with linear memory.
Finally, we have shown, by the realization of proof-of-concept demonstrators with all-electrical input and readout, that it is possible to move these systems up the TRL ladder and that commercially viable devices is within reach.
Experimental demonstration of reservoir computing has also been demonstrated using a single node nanoring reservoir and computational tasks such as spoken digit recognition, TI-46, has been completed with good results.
We have shown that we can reliably read and write data into the two types of nanomagnet ensembles that are studied in this project. For the ring arrays, the reading signal is based on the AMR response and for the ASI, a Hall bar is added under the array which exploits the AHE to probe the magnetization states.
We have demonstrated tuning of nanomagnet ensembles using a variety of stimuli, perhaps most notably, we have demonstrated reproducible, clocked magnetization reversal in ASI systems. Unlocking the discrete dynamics of these dynamical systems is believed to be a crucial door-opener for the design of reservoirs with linear memory.
Finally, we have shown, by the realization of proof-of-concept demonstrators with all-electrical input and readout, that it is possible to move these systems up the TRL ladder and that commercially viable devices is within reach.