| Popis: |
The technology using densely arranged discrete magnetic dots on disk surface to store binary data is referred as bit patterned magnetic recording (BPMR). The research work presented in this thesis covers two important aspects of the technology: The development of a novel method for experimentally fabricating the bit patterned media, referred to as template two-phase growth and the development of a machine-learning based data recovery scheme when a magnetic read back sensor is significantly wider than the discrete magnetic bits. This thesis, therefore, is divided into two parts: The first part involves experimental fabrication and materials development and the second part uses modeling and computer simulation approaches. The conventional way of producing bit patterned media is the lithography-based subtractive technique that uses etching to create patterned magnetic dots from a uniform magnetic sheet film. At required area densities, in the order of 1013 dots per square inch area, the pitch distance between adjacent dots is less than 10nm. This requirement makes the lithography-based subtractive method extremely challenge if not impossible. With an approach in contrary to the etching-based subtractive techniques, we worked on an additive method based-on two phase template growth: by lightly patterning an underlayer followed by co-growth of two different types of materials one could generates well-arranged dense magnetic dots with thin non-magnetic boundaries separating them. In this thesis, we focused on ordered FePt-L10 as the magnetic material and oxide as the boundary material. In the second part of this research work, we present a study that uses the machine learning-based data detection channel consisting of a convolutional neuron network (CNN) for data recovery in BPMR with a relatively wide reader covering two data tracks. We demonstrate that in noise-less conditions, CNN can fully resolve both inter-symbol interference (ISI) and inter-track interference (ITI) by sufficient training. Additive white Gaussian electronic noise causes the performance of CNN detection channel to degrade. However, by doubling the sampling frequency to take the advantage of staggered interference sidetrack, the detection performance significantly improves with much-enhanced tolerance of white Gaussian noise. It is concluded that a machine learning channel with a convolution neural network could enable the application wider reader in BPMR due to the learning ability in eliminating inter-track interference. This data recovery ability in the presence of severe inter-track interference is associated with the fact that bits' position in adjacent tracks is fixed. With the machine learning channel, no additional reader would be necessary as in the case of two-dimensional magnetic recording (TDMR). Furthermore, we explore the CNN detection channel performance when track misregistration (TMR) is introduced into the readback signal. Due to the TMR effect (or say, reader offsets), the sampled readback signals may experience a variable inter-track interference instead of a fixed one. By training with dataset that contains a few selected TMR levels and testing with more different TMR levels, CNN can still resolve inter-track interference and maintain optimal performance under electronic noise. We concluded that the CNN-based detection channel has the potential to resolve inter-track interference for different reader positions to the target track by training with signals from a few sparse ITI levels/reader positions. The key is to cover the extreme conditions and carefully choose the sparse training point in between to reach the desired bit-error rate level. |
