Abstract
<title>Abstract</title> <p>Plasma etching is a critical but difficult-to-monitor process in semiconductor manufacturing because wafer quality is governed by nonlinear plasma chemistry, sheath dynamics, surface reactions, chamber conditions, and tool drift. Physical metrology provides accurate measurements of quantities such as critical dimension, etch depth, etch rate, profile shape, and uniformity, but it is usually sparse, delayed, and costly. Virtual metrology (VM) addresses this limitation by estimating wafer quality from in situ sensor and equipment data, including optical emission spectroscopy (OES), radio-frequency (RF) signals, and equipment engineering system (EES) variables. This review analyzes 105 plasma-etch VM studies published between 2012 and 2025 and organizes the field around a two-stage framework. Stage 1 covers feature selection, feature transformation, spectral and temporal compression, clustering, deep representation learning, and hybrid compression strategies. Stage 2 covers regression, classification, endpoint detection, fault detection and classification, and uncertainty-aware and interpretable prediction, with multi-task extensions in which several etch targets are predicted from a shared representation. The review shows that the main challenge in plasma-etch VM is not only prediction accuracy, but also the construction of compact, robust, and physically meaningful representations from high-dimensional, noisy, and drift-prone process data. Key open issues include sparse labels, model drift, chamber transferability, benchmarking, uncertainty reporting, latency, and integration with advanced process control. The review highlights knowledge-guided compression, adaptive learning, shared representations, and uncertainty-aware deployment as important directions for industrial VM systems.</p>