Machine Learning for Intelligent Fiber Sensors: From Physical Design Optimization to Advanced Sensing Applications

Fiber sensors encompass optical-fiber platforms and electrically functional fibers, yarns, and textiles that produce spectral, temporal, and spatial data across wearable, biomedical, robotic, infrastructure, and environmental applications. Their responses are shaped by nonlinear transduction, cross-sensitivity, conformal contact, manufacturing variation, drift, and limited calibration coverage. This review examines how machine learning (ML) can be used across the fiber-sensing pipeline, from algorithm-assisted physical design and fabrication-aware optimization to denoising, demodulation, representation learning, multimodal inference, and application-level decision making. Direct demonstrations in optical and electrically functional fiber systems are distinguished from transferable methods developed in adjacent photonic, additive manufacturing, tactile, and soft-sensor platforms. The available evidence is strong for calibration, signal recovery, and task inference within a defined device and operating domain. By contrast, experimental evidence for ML-guided physical co-design, cross-device and cross-batch transfer, long-term stability, and hardware-constrained deployment remains limited. Accordingly, ML is most promising as part of a coupled sensor–model–hardware system that preserves task-relevant information and reports when an input lies outside its validated operating range. This review therefore emphasizes validation boundaries over headline accuracy alone and identifies uncertainty quantification, physics-guided learning, fabrication metadata, and edge-aware co-design as priorities.