CRII: CIF: Machine Learning Based Equalization Towards Multitrack Synchronization and Detection in Two-Dimensional Magnetic Recording

该项目涉及二维磁记录（TDMR），这是一种用于硬盘驱动器的新型记录技术，可以大幅增加数据密度，达到每平方英寸10太比特。TDMR的收益来自两个方向，即(i)瓦片写入机制，其中相邻的数据磁道部分重叠写入，如屋顶瓦片，以便在磁盘上压缩更多磁道并增加数据密度，以及（ii）强大的信号处理算法，能够在数据磁道内部和跨数据磁道存在高水平干扰的情况下，从嘈杂的读回信号中有效地恢复数据。机器学习（ML）技术将用于在存在二维干扰、数据依赖和彩色媒体噪声的情况下开发此类数据恢复算法。提出的工作旨在实现TDMR的重大改进，最终允许在更少的高容量磁盘驱动器上存储指数级增长的数据量。该奖项部分支持博士生在TDMR读取通道设计方面的培训，最终为学生在数据存储行业创造就业机会。研究目标是开发高效的基于ML的均衡算法，该算法优于传统的高密度TDMR通信理论均衡。TDMR信道是高度非线性的，机器学习方法有望更好地学习其特性，与传统的线性通信理论方案相比，可能导致更高的误码率。期望的神经网络均衡方案寻求(i)结合媒体噪声的预测和消除，以及（ii）与研究者开发的新颖读取通道架构兼容，该架构将部分响应范式扩展到异步轨道的多轨检测情况。为了实现这一读信道，所开发的均衡器之后将采用旋转目标（ROTAR）算法，这是一种异步轨道的多轨道检测器，也是由研究者开发的。由此产生的读取通道预计将在面密度和吞吐量方面比目前工业上使用的通信理论和单轨道检测方案有所提高。开发算法的性能将与使用国际合作者提供的现实波形的传统算法进行比较。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

This project concerns two-dimensional magnetic recording (TDMR), a novel recording technology for hard disk drives that allows for a drastic increase in data density, up to 10 terabits per square inch. Gains from TDMR come from two directions, namely (i) the shingled writing mechanism whereby adjacent data tracks are written with partial overlap, like roof shingles, in order to squeeze many more tracks on the disk and increase data density, and (ii) powerful signal processing algorithms that enable efficient data recovery from noisy readback signals in the presence of high levels of interference both within and across data tracks. Techniques from machine learning (ML) will be used in developing such data recovery algorithms in the presence of two-dimensional interference, and data-dependent and colored media noise. The proposed work aims to achieve significant improvements in TDMR, eventually allowing exponentially increasing volumes of data to be stored on fewer disk drives with higher capacities. This award partially supports a PhD student to be trained in TDMR read channel design, ultimately creating career opportunities for the student in the data storage industry. The research objective is the development of efficient ML based equalization algorithms that outperform conventional communication-theoretic equalization for high density TDMR. The TDMR channel being highly nonlinear, ML approaches are expected to better learn its characteristics, potentially leading to higher bit-error rates when compared to conventional linear communication-theoretic schemes. The desired neural network equalization schemes seek to (i) incorporate the prediction and cancellation of the media noise, and (ii) be compatible with a novel read channel architecture, developed by the investigator, that extends the partial-response paradigm to the case of multitrack detection of asynchronous tracks. To realize this read channel, the developed equalizers will be followed by the rotating-target (ROTAR) algorithm, a multitrack detector of asynchronous tracks, also developed by the investigator. The resulting read channel is expected to yield gains in areal density and throughput over the communication-theoretic and single-track detection schemes currently used in the industry. The performance of the developed algorithms will be compared against that of conventional algorithms using realistic waveforms provided by international collaborators.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.