PIML review of 212 studies shows gains in machine health, but gaps remain
Physics-informed ML beats baselines in 212 studies, yet focuses mostly on batteries and bearings.
A systematic literature review from Christopher Braun, Julian Raible, and Marco F. Huber (Fraunhofer IIS) analyzed 212 studies on physics-informed machine learning (PIML) applied to prognostics and health management (PHM)—the discipline of predicting when industrial equipment will fail. Published on arXiv (2608.10047) and in the Journal of Intelligent Manufacturing, the review categorizes PIML approaches into four classes: observational bias, inductive bias, learning bias, and hybrid approaches. This taxonomy helps researchers understand how physical knowledge is injected into ML pipelines, whether through loss functions, architectural constraints, or pre-processed data.
Across all four classes, the studies consistently reported better predictive performance than purely data-driven baselines, especially for predicting remaining useful life and detecting faults in rotating machinery and energy storage systems. The literature, however, is heavily skewed toward lithium-ion batteries and bearings, limiting generalizability. The authors also found that while PIML improves accuracy, many claims about enhanced interpretability and causal reasoning lack rigorous evidence. They call for transferable design patterns, standardized benchmarks comparing integration strategies, and lightweight, uncertainty-aware models suitable for real-time industrial deployment.
- Review of 212 papers maps PIML into 4 bias classes: observational, inductive, learning, hybrid
- PIML consistently beats conventional ML baselines for PHM tasks like RUL prediction and fault detection
- Research is concentrated on lithium-ion batteries and bearings; interpretability gains are not robustly proven
Why It Matters
PIML promises safer, cheaper predictive maintenance, but needs balanced benchmarks and real-time models to scale beyond labs.