Enable TDMR Gain with Convolution Neural Network Based Machine Learning Algorithm

Abstract

Two-dimensional magnetic recording (TDMR) with laterally-displaced multiple read heads has been developed to alleviate the impact of inter-track interference (ITI) and track edge effects. However, the cross-track (XT) position of the displaced 2nd head varies significantly with the skew angle of head assembly from ID to OD across disk radius [1]. As a result, each disk platter is divided into hundreds of groups of sectors, or zones, to accommodate the skew angle variations. The 2-dimentional equalization technique used in conventional data detection channels requires different channel parameters at different skew angles/zones for channel optimization, and the number of different channel parameter sets across the disk platter are in hundreds. Furthermore, resulting performance gain is further diminished by random fluctuations such as non-repeatable disk runout.In this situation, if we employ machine learning technique on TDMR data detection, one could imagine the possibility of achieving optimal performance in resolving ITI and track edge effects across the entire disk surface while using “learning” to replace relatively complex parameterization of physical modeling. In our previous work, we have developed deep neural network (DNN) for disk drive data detection channels with performance (detection error-rate vs. signal-to-noise ratio) slightly more superior than the state-of-the-art disk drive channels [2,3]. In the work presented here, we train a convolution neural network (CNN) in TDMR system to detect data under a wide range of ITI from different head skew angles. We demonstrate that in the medium noise dominant situation, CNN can learn all ITI in the training dataset and completely recover ITI-free BER without any physical modeling. Our eventual goal is to explore if we can train a single CNN detector to learn and optimize for all head skew angles/zones automatically, thereby lifting the complexity of designing different channel parameter sets at each head skew angle.The recorded tracks and CNN detection scheme are shown in Fig. 1. Two adjacent data tracks are written with random data and arbitrary phase. The width of the track is denoted TW, and no erase-band nor transition curvature is included. Two XT overlapped heads of width greater than TW are used. Here we simplify the modeling and assume each head only captures signals under its XT coverage. Since the relative positions of head-1 and head-2 are locked, we characterize ITI of the reader array with a single variable: the ratio between head-1 coverage of Track-II (Δx1) versus the track width (TW): δ1=Δx1/TW. Here we assume the ITI stays fixed within each zone, and different zones will have different ITI. In this work, we only investigate the medium noise dominated situation: the jitter noise is modeled as the uncertainty of the transition locations following a Gaussian distribution. Since both heads “see” the same magnetic patterns from the two tracks, the jitter noise is correlated noise. A CNN is used as the data detector in this work. It receives signals from both heads as input, and we only train it to detect data for Track-I. The rest of the parameters are listed in Fig. 1 caption.To simulate reading data at different skew angles, we evaluate the CNN detection BER with signals of 11 different ITI from δ1=0.2 to 0.8 as in Fig. 2. The black dashed line benchmarks the ITI-free BER. For the blue curve, the CNN detector is trained using signals from one skew angle/zone with a fixed ITI, δ1=0.5 . In this case, CNN can only eliminate the trained ITI, and the detection accuracy degrades significantly whenever the testing ITI is different. In contrast, the red curve plots the BER when we train the CNN using signals with mixed ITI same as the testing dataset. This time CNN can successfully eliminate ITI for all testing head skew angles/zones.Our results suggest that CNN has the capacity to learn very complicated correlations and is solely data-driven. With the learning process, the CNN-based ML channel presented here is capable of resolving the interference from ITI infested signals, automatically rejecting them, and arriving at correct detection. This approach is fundamentally different from the conventional channel design that relies on sophisticated physical modeling and channel parametrization. In addition, the fact that a single CNN can learn a wide range of ITI states that it can optimize the channel performance for head skew angles using a single CNN eliminating the need to use multiple channel parameter sets. Not only this machine learning channel can eliminate the complex physical characterization process required for conventional approach, more importantly, the TDMR performance gain can be fully realized without compromising. The potential of fully eliminating ITI also presents the possibility to have reader width wider than track width which could be critically important as track pitches in HDDs continue to decrease. **