Machine Learning-based Data Detection Channel for Two-Dimensional Magnetic Recording System in Hard Disk Drives

Hard disk drive (HDD) has been one of the key technology enabling data centric life we are enjoying today. Currently, over 90% data worldwide are stored in HDDs. Inside a high capacity HDD, currently capable of storing 20TB of data, there are about 10 disk platters. Binary bits are stored as the magnetic moment orientation of tiny magnets in the magnetic layer on each surface of a platter with a storage density exceeding 1 Tbits/in2. Magnetoresistive read sensor flying closely over the surface as the disk rotates converts the magnetic fields from the recorded bits to electric signal for data detection. Present data detection channel, an elaborated signal processing system, performs elaborate data recovery work since the signal doubtlessly contains severe interference and also corrupted by both magnetic and electronic noise. The highly spatially packed bits on a disk surface exasperate this situation. In recent years, multiple spatially displaced read sensors have replaced the single read sensor in an attempt to resolve interference from adjacent data tracks. However, the scheme suffers from diminished gain returns due the complexity and inflexibility of physical models needed. 

In this thesis work, a completely new type of data detection channel based on machine learning (ML) neural networks has been developed through a systematic and thorough modeling study followed by actual ASIC hardware realization. It is found that the novel data detection channel constructed using optimized neural networks can perform superior data detection when the noise in the read back signal is mainly caused by the magnetic interactions among the bits. The ML channels are also capable of completely eliminate the interference from adjacent track in the multi-reader setting without any needs for physical modeling of constructing the signals from different readers. The autonomous performance optimization in the ML channels for correlated magnetic noise and intertrack interference substantially complements the performance weakness in conventional signal processing based channels. 

As a part of this thesis work, the developed neural network based ML data detection channel via computer simulation has been implemented a ASIC integrated circuit tape out using TSMC 28nm technology. Systolic array based parallel architecture have been utilized for the high throughput along with various circuit innovations for detection accuracy and power efficiency. The chip testing with data from a state-of-the-art commercial hard disk drive demonstrates a 30% more accurate data detection comparing to the on-drive data detection channel. 

This thesis work is first of its kind in the field of hard disk drive technology. This trailblazing research provides the very first set of systematic analysis along with hardware data, illuminating new directions for the technology advancement. We believe this research enlightens the future use of artificial intelligence (AI) for hard disk drive technology. It is our hope that more significant research and development will follow and hard disk drives will have ML data detection channels with significantly enhanced storage capacity in the very near future.