Physics-Informed Machine Learning in Prognostics and Health Management: A Systematic Literature Review

This paper presents a systematic literature review of Physics-Informed Machine Learning (PIML) applied to Prognostics and Health Management (PHM), covering 212 studies. The review introduces a four-class classification scheme for integrating physics into machine learning: observational bias, inductive bias, learning bias, and hybrid approaches. It further categorizes studies by PHM task (fault detection, diagnosis, health assessment, and prognosis) and by asset type.

The review finds that lithium-ion batteries and bearings are the most studied assets, accounting for approximately half of the studies. The literature is dominated by problem-specific solutions rather than transferable design patterns. Across all four classes, the reviewed studies consistently demonstrate improved predictive performance over conventional baselines across a broad range of assets.

The paper addresses three main research questions: (1) what types of prior physical knowledge are leveraged and in what forms, (2) how this knowledge is incorporated and how the form of representation influences feasibility, and (3) how incorporating physics helps overcome limitations of purely data-driven methods and what the primary challenges are.

Regarding the types of prior knowledge, the field relies predominantly on mechanistic models and explicit degradation laws, while structural and causal relationships and qualitative degradation properties see less frequent application. The forms of representation span from highly formalized expressions (e.g., ODEs, PDEs), through empirical and phenomenological formulations, down to implicit forms that capture principled assumptions.

Regarding incorporation, three overarching pathways constitute distinct yet complementary classes: observational bias (introducing physics via additional training data), inductive bias (embedding physics into the model design), and learning bias (imposing physics through the learning algorithm, typically via composite loss functions). Hybrid approaches form a conceptually distinct fourth class, where a physics-based model and an ML model are coupled either in parallel or in series. Learning-bias approaches are the most widely adopted, followed by inductive-bias and observational-bias approaches. Hybrid approaches constitute the least frequently adopted class.

The form of representation is the primary design lever determining which pathways are viable. Mechanistic models are flexible enough to support all four classes, while structural and topological information naturally maps to inductive bias via graph-based approaches, and qualitative properties are predominantly expressed as learning bias or simple architectural constraints. Richer, more formal representations permit closer alignment with physics but require substantial modeling effort and are often asset-specific, whereas qualitative properties are inherently limited in physical fidelity yet typically retain broad applicability.

Regarding practice, the review finds that PIML already provides tangible benefits, particularly in predictive performance. However, claims of improvements concerning generalizability, interpretability, robustness, and physical consistency lack sufficient supporting evidence. The review notes that observational-bias methods can plausibly tackle data scarcity and improve in-distribution generalization, but they leave the hypothesis space unchanged and are poorly equipped to enforce physical consistency or principled extrapolation. Hybrid approaches that employ in-series coupling largely inherit the shortcomings of purely data-driven methods.

The primary challenges identified include: the field's fragmentation and lack of standardized design patterns; the entanglement between embedded physics and specific assets, limiting transferability; the absence of uncertainty quantification in most studies; the lack of evidence for interpretability; and the lack of online capability and closed-loop integration with existing monitoring and maintenance workflows. Most methods remain at the level of offline prototypes rather than operational tools, corresponding broadly to early Technology Readiness Levels (approximately TRL 3–4).

The review also highlights several methodological gaps: the absence of ablation experiments in many studies, preventing isolation of the contribution of incorporated physics; the risk of evaluation leakage when physics-model parameters are calibrated on data overlapping with the test set; and the lack of quantification of computational costs and trade-offs. Physical consistency lacks a shared definition and corresponding metrics, leaving reported improvements largely qualitative.

Future research directions proposed include: benchmark studies systematically comparing integration strategies under controlled conditions; moving from problem-specific designs toward modular building blocks for inductive-bias approaches; developing adaptive loss-balancing methods and physics-informed optimizers; establishing formal definitions and metrics for physical consistency; prioritizing generalizable solutions over asset-specific studies; developing physics-informed foundational degradation models; automating methods for selecting, calibrating, and embedding prior knowledge; integrating uncertainty quantification as an integral design objective; investigating lightweight architectures for edge hardware; and demonstrating closed-loop integration with industrial infrastructure.

Improvements to AI Systems:

1. Physics-constrained neural architectures – Embed ODE/PDE constraints directly into layer design (inductive bias) so the model inherently satisfies physical laws (e.g., conservation of mass, energy) without needing extra loss terms.

Resulting capability: The AI can predict system degradation (e.g., bearing wear) with physically consistent trajectories even under sparse or noisy sensor data.

2. Hybrid physics-ML coupling with uncertainty quantification – Implement a serial/parallel coupling where a mechanistic model (e.g., battery capacity fade) and a neural network share latent states, with Bayesian layers to output prediction intervals.

Resulting capability: The AI can provide reliable remaining-useful-life (RUL) estimates with calibrated confidence bounds, enabling risk-aware maintenance decisions.

3. Adaptive physics-informed loss balancing – Use a meta-learning controller that dynamically weights the data-loss vs. physics-loss terms during training, based on gradient conflict or validation performance.

Resulting capability: The AI automatically adjusts to changing operating conditions (e.g., temperature shifts) without manual re-tuning, maintaining accuracy and physical plausibility.

4. Modular physics-aware building blocks – Create reusable neural components (e.g., graph-based structural encoders, differentiable physics simulators) that can be plugged into different PHM tasks (fault detection, prognosis) across asset types.

Resulting capability: The AI can transfer learned degradation patterns from one asset (e.g., bearings) to another (e.g., gears) with minimal retraining, reducing the need for asset-specific data.

5. Ablation-aware training pipeline – Automatically generate counterfactual models (with/without physics) during training and report performance deltas, preventing evaluation leakage by calibrating physics parameters only on training folds.

Resulting capability: The AI system can self-audit its own improvements, clearly isolating the contribution of embedded physics and flagging overfitting to test data.

6. Physics-informed data augmentation – Generate synthetic training samples by solving mechanistic models with perturbed parameters, then feed them alongside real data (observational bias).

Resulting capability: The AI can learn robust degradation patterns from limited real failure data, improving fault detection accuracy in rare-event scenarios.

7. Edge-deployable lightweight physics encoders – Distill a full physics-informed model into a compact neural network (e.g., via knowledge distillation) that runs on low-power hardware while retaining key physical constraints.

Resulting capability: The AI can perform real-time health monitoring and anomaly detection on embedded sensors (e.g., in wind turbines or EVs) without cloud connectivity.

8. Closed-loop maintenance recommender – Integrate the PIML predictor with a decision module that outputs maintenance actions (e.g., replace, derate) based on predicted RUL and uncertainty, and updates the physics model from post-maintenance feedback.

Resulting capability: The AI can autonomously trigger maintenance, learn from actual failure outcomes, and improve its long-term predictive accuracy in operational settings.

In modern industry, keeping complex systems reliable, safe, and efficient hinges on Prognostics and Health Management (PHM). Machine Learning (ML) has largely driven advancements in diagnostics and prognostics, yet purely data-driven models face inherent limitations, such as poor generalization, an inability to infer causal relationships, and a lack of interpretability. Physics-Informed Machine Learning (PIML) helps mitigate these limitations by incorporating prior physical knowledge directly into the ML pipeline, thereby fostering growing interest in its application to PHM. This work investigates how PIML is being leveraged in the context of PHM through a systematic literature review of 212 studies. The review introduces a four-class classification scheme, consisting of observational bias, inductive bias, learning bias, and hybrid approaches, and further categorizes studies by PHM task. Across all four classes, the reviewed studies consistently demonstrate improved predictive performance over conventional baselines across a broad range of assets, although the literature is heavily skewed toward lithium-ion batteries and bearings, and dominated by problem-specific solutions. Overall, the review indicates that physics-informed approaches already provide tangible benefits, whereas claims of improvements concerning some of the aforementioned limitations lack sufficient supporting evidence. Future research should prioritize transferable design patterns, benchmarks comparing integration strategies, and uncertainty-aware models that are lightweight and robust enough for online deployment in real-world settings.