Toward Smarter SSDs: A Comprehensive Survey of Machine Learning Techniques for Internal Optimization
Taeseok Kim · IEEE Access · 2025
Solid-State Drives (SSDs) have become a foundational component of modern storage infrastructures, offering superior latency, throughput, and energy efficiency compared to traditional spinning disks. However, as computing systems continue to scale in complexity and workload diversity, SSDs face mounting challenges in maintaining consistent performance, endurance, and resource utilization. Traditional rule-based management techniques, while efficient under static conditions, struggle to adapt to dynamic I/O patterns. This limitation leads to inefficiencies in core operations such as address translation, garbage collection, wear leveling, and internal scheduling. In this survey, we examine the growing body of research that leverages machine learning (ML) for SSD-internal optimization. Although this area holds significant potential, it has received comparatively less attention than host- or system-level approaches. We present a systematic taxonomy of over 40 representative studies and categorize them based on optimization targets, ML methodologies, evaluation strategies, and implementation maturity.We also highlight critical challenges including computational overhead, dataset limitations, and real-time adaptability, all of which impact the feasibility of deploying ML within resource-constrained SSD controllers. Beyond synthesizing the current state of the art, we identify key trends such as the adoption of deep and reinforcement learning, the emergence of host-assisted learning architectures, and increasing efforts toward prototype-level implementation. Furthermore, we outline future research directions that emphasize holistic optimization across SSD subsystems, cross-layer collaboration, and energy-aware intelligent control. By focusing on the architectural integration of ML directly within SSDs, this survey aims to guide the development of next-generation storage systems that are not only faster and more durable but also inherently self-optimizing and context-aware.