Piezoelectric Materials for Autonomous Internet of Things and Edge Intelligence: From Fundamental Physics to Self-Powered Intelligent Systems

The rapid proliferation of Internet of Things (IoT) technologies is transforming modern infrastructure, healthcare, transportation, and environmental monitoring into interconnected cyber-physical ecosystems. Billions of distributed sensor nodes are expected to operate in locations where wired power is unavailable and battery replacement is impractical, making energy autonomy a primary barrier to large-scale deployment. Conventional sensing platforms remain constrained by the energy required for acquisition, processing, and wireless communication, motivating sensing mechanisms capable of both detecting events and powering their own electronics.

Among ambient energy sources, mechanical energy is particularly attractive because it is ubiquitous in natural and built environments: structural vibrations, human motion, airflow, acoustic waves, and pressure fluctuations are continuously present across multiple scales. Piezoelectric materials are uniquely suited to exploit this resource because they inherently couple mechanical and electrical domains, allowing direct conversion of mechanical deformation into electrical charge while simultaneously acting as sensing elements. This dual functionality distinguishes piezoelectric systems from other energy harvesters and positions them as a key enabling technology for self-powered sensing architectures.

The understanding of piezoelectricity evolved from early electromechanical observations in quartz crystals [1, 2] to engineered sonar transducers [3], and later to optimized ferroelectric ceramics forming the basis of modern electromechanical devices [4, 5]. Ferroelectric ceramics such as barium titanate and lead zirconate titanate (PZT) enabled high coupling coefficients and industrial reliability [6], while the discovery of piezoelectric polymers such as polyvinylidene fluoride (PVDF) introduced flexibility and large-area conformability [7]. More recently, nanoscale and two-dimensional materials including ZnO nanowires and halide perovskites have demonstrated electromechanical coupling at extremely small thicknesses [8], alongside the development of environmentally friendly lead-free compositions. As a result, piezoelectric technology has evolved from simple transducers to multifunctional platforms capable of sensing, actuation, and energy harvesting.

Mechanical energy harvesting using piezoelectric structures has been demonstrated for powering wireless and mobile electronics [9, 10]. Nonlinear and bi-stable piezoelectric configurations have also been investigated to improve energy-harvesting performance over varying excitation conditions [11], while electromechanical architectures integrating piezoelectric elements with vibration and wave-control structures provide another route for converting captured mechanical energy into electrical power [12]. Biomedical systems that harvest energy from cardiac and respiratory motion have further shown the feasibility of extending implant lifetime [13]. However, increasing material performance alone does not guarantee autonomous operation. In practical IoT nodes, wireless communication and data processing often dominate the energy budget, making system architecture as important as the material itself. Recent advances in ultra-low-power computing and tiny machine learning (TinyML) enable local interpretation of sensor signals, dramatically reducing transmission requirements [14-16].

At the same time, scalable deployment requires compatible fabrication processes and long-term durability. Manufacturing pathways compatible with flexible substrates and integrated electronics are actively being explored [6], while long-term reliability under cyclic loading and environmental exposure remains a major challenge requiring fatigue-aware design and encapsulation strategies [17]. Hybrid and stacked harvesting architectures have been proposed to extend operation under variable excitation conditions [18], and recent demonstrations confirm that piezoelectric devices can function simultaneously as sensors and energy sources in realistic IoT platforms [19].

Despite rapid progress across materials, fabrication, power electronics, and machine learning, these advances are typically optimized in isolation. As a result, reported improvements in piezoelectric coefficients or harvested power do not necessarily translate into longer operational lifetime. Autonomous sensing is instead governed by the balance between generated energy, conversion efficiency, and algorithmic cost—a balance that remains insufficiently addressed in existing literature. While recent reviews comprehensively cover either material synthesis and harvesting circuits, or energy-harvesting technologies more broadly, they do not explicitly formulate the energy-information co-design condition linking material, circuit, and algorithmic layers that this work addresses.

Accordingly, this review provides an integrated perspective linking material physics, fabrication constraints, power-management efficiency, and embedded intelligence within a unified energy-autonomy framework. The energy-balance relation used in this framework is not proposed as a new conservation law. Its contribution is a design-oriented cross-layer interpretation that expresses material-dependent energy generation, interface losses, sensing and inference costs, communication frequency, and standby losses using a common autonomy criterion. This formulation enables interventions at different system layers—for example, increasing piezoelectric output or reducing radio activity—to be compared quantitatively rather than discussed independently. Specifically, the work (i) connects intrinsic coefficients (e.g., $d_{33}$, $g_{33}$) to extractable electrical energy, (ii) classifies materials by functional suitability for autonomous nodes, (iii) analyzes harvesting and duty-cycled operation in the context of system power budgets, and (iv) consolidates TinyML-based sensing strategies into deployable architectures.

This review adopts a narrative approach to synthesize interdisciplinary advances spanning piezoelectric materials, energy harvesting circuits, and embedded machine learning for autonomous IoT systems. The literature search was conducted primarily through Scopus, Web of Science, and IEEE Xplore, supplemented by Google Scholar for cross-referencing and grey literature identification. The search employed combinations of the following keyword strings: "piezoelectric energy harvesting," "self-powered sensor," "piezoelectric IoT," "TinyML," "edge AI," "lead-free piezoceramics," "nanogenerator," "PVDF sensor," "power management harvesting," and "autonomous sensing". The search covered publications from 2000 to 2026, with priority given to peer-reviewed journal articles, foundational textbooks, and high-impact conference proceedings. Earlier seminal works (pre-2000) were included where necessary to establish historical context and foundational principles. No formal inclusion/exclusion scoring rubric was applied, consistent with the narrative review methodology; however, source selection prioritized: (a) studies reporting quantitative electromechanical or energy-harvesting performance data, (b) works demonstrating system-level integration of sensing and intelligence, and (c) recent comprehensive reviews used for cross-validation and positioning. Reference lists of key review articles were manually screened (backward citation tracking) to identify additional relevant studies. In total, approximately 250 sources were screened, of which 80 are cited in this work.

The remainder of the paper first reviews the physical principles governing piezoelectric coupling, followed by the evolution of material systems from ceramics to nanostructures. Fabrication and characterization methods are then discussed, after which system-level integration with IoT architectures and edge intelligence is analyzed. Finally, challenges and future research directions are presented, outlining the pathway toward scalable, long-term, self-powered sensing networks.

The evolution of piezoelectric materials reflects a progressive shift from maximizing intrinsic electromechanical coefficients toward optimizing materials for system-level integration and energy-autonomous operation. Early research focused on demonstrating electromechanical coupling in crystalline materials, followed by maximizing output using engineered ceramics. Later developments prioritized mechanical compatibility through polymers and composites, while recent advances emphasize miniaturization and autonomous system integration. Figure 1 summarizes this transition, illustrating how the field evolved from material discovery toward functional IoT-oriented systems rather than purely improving intrinsic material properties.

Figure 1 emphasizes that material innovation has progressively aligned with application-driven constraints rather than purely crystallographic performance metrics.

image034.png

Figure 1. Timeline of piezoelectric material development and functional evolution

3.1 From intrinsic crystals to engineered ceramics

The earliest piezoelectric materials were naturally occurring crystals such as quartz, Rochelle salt, and tourmaline. These materials offered excellent chemical stability and reproducibility but produced relatively small electromechanical responses, restricting their use mainly to precision sensing and frequency-control applications.

The demand for stronger actuation and higher electrical output led to the development of ferroelectric ceramics. The discovery of barium titanate demonstrated that domain reorientation could significantly amplify polarization compared with intrinsic lattice distortion [31]. This development culminated in PZT, whose morphotropic phase boundary enables enhanced domain mobility and high electromechanical coupling [32]. Consequently, PZT became an industrial standard for actuators, ultrasonic transducers, and high-power electromechanical devices.

Subsequent advances introduced relaxor single crystals such as PMN-PT and PZN-PT, which exhibit exceptionally high piezoelectric coefficients suitable for medical imaging and precision-positioning systems [36]. However, their complex fabrication and lead content have motivated the development of environmentally safer alternatives. Lead-free materials, including potassium sodium niobate and modified barium titanate ceramics, have therefore been investigated through compositional engineering, phase-boundary control, and crystallographic texturing to approach the performance of lead-based systems [37, 38].

Recent KNN ceramics research increasingly focuses on multiscale phase engineering, crystallographic texturing, grain-orientation control, and defect regulation to enhance domain rotation and electromechanical response while retaining a lead-free composition [37]. Data-driven approaches have also been introduced to identify composition-structure descriptors associated with high $d_{33}$ in KNN-based ceramics [39]. These developments indicate that the performance gap between lead-free and lead-based ceramics is increasingly being addressed through coordinated composition-microstructure design rather than composition modification alone.

Despite their high electromechanical performance, ceramic materials remain mechanically rigid and brittle, limiting their use in flexible electronics, wearable sensing, and systems subjected to repeated large deformation.

3.2 Polymeric piezoelectrics: Flexibility and conformability

The discovery of piezoelectricity in PVDF introduced mechanically compliant materials capable of conformable sensing [7]. In polymers, polarization originates primarily from aligned molecular dipoles rather than from conventional ceramic domain switching. Although this mechanism produces lower charge coefficients, the low dielectric permittivity of PVDF results in high voltage sensitivity. Polymer piezoelectrics are therefore particularly suitable for high-impedance sensing interfaces, wearable devices, and low-power IoT nodes.

Mechanical stretching and electrical poling promote formation of the polar β-phase responsible for the piezoelectric response [33]. Copolymers and electrospun nanofibers can further improve dipole alignment, electroactive-phase content, and mechanical flexibility [40].

Recent research has extended these approaches toward microstructure-engineered nanofibers, porous films, conductive composites, and mechanically compliant sensor architectures. These strategies aim to improve strain transfer, increase the electroactive β-phase fraction, and maintain sensitivity under repeated deformation [41]. Recent reviews of flexible piezoelectric sensors emphasize filler engineering, fabrication optimization, microstructure control, cycle life, and integration with wearable intelligent systems.

Such materials are increasingly used in wearable sensors, biomedical patches, physiological monitoring, and human-motion recognition. However, polymeric piezoelectrics generally provide lower charge generation and electromechanical coupling than high-performance ceramics. Applications requiring larger harvested power or stronger actuation therefore commonly employ composite or hybrid material strategies.

3.3 Composite and hybrid architectures for property co-optimization

Composite piezoelectrics combine ceramic inclusions with polymer matrices to simultaneously achieve flexibility and sensitivity. Connectivity models such as particulate (0-3), fiber-reinforced (1-3), and layered (2-2) structures allow tuning of stiffness, permittivity, and coupling efficiency [42].

Nanoscale fillers such as ZnO nanowires, carbon nanotubes, and graphene can enhance interfacial polarization, charge transport, and local stress transfer [43]. Additive-manufacturing and printing techniques further enable scalable fabrication of conformable devices, patterned sensors, and piezoelectric structures embedded directly within host components [44].

Recent composite designs increasingly exploit porous or fibrous architectures, multiscale filler networks, and controlled ceramic-polymer interfaces to improve electromechanical response without sacrificing flexibility. Their performance nevertheless depends strongly on filler dispersion, interface adhesion, dielectric contrast, percolation behavior, and the efficiency with which mechanical strain is transferred to the active phase [45-48].

3.4 Nanostructured and two-dimensional piezoelectrics

The demonstration of piezoelectricity in ZnO nanowires showed that electromechanical coupling persists at nanoscale dimensions, enabling nanogenerators capable of converting small mechanical deformations into electrical signals and energy [34]. ZnO nanowires are particularly attractive because they combine nanoscale dimensions, semiconductor functionality, and compatibility with flexible substrates.

Atomically thin materials such as monolayer MoS₂ were subsequently shown to exhibit intrinsic piezoelectricity because of broken inversion symmetry [49]. Their atomic thickness and mechanical flexibility make them promising for ultrathin sensors and flexible electronic platforms. However, the coefficients reported for two-dimensional materials are often defined per unit length or through effective device parameters and should not be interpreted as directly equivalent to bulk $d_{33}$.

Data-driven and high-throughput computational methods are increasingly being applied to nanoscale materials. Machine-learning models can assist in identifying relationships among composition, symmetry, bonding, electronic structure, and functional response, thereby reducing the number of candidates requiring full first-principles or experimental evaluation. However, screening results should be distinguished carefully according to whether the predicted property is piezoelectricity, mechanical flexibility, bandgap, or resistive behavior.

Hybrid nanosheet-polymer and nanowire–polymer structures can improve mechanical durability, strain transfer, and electrical output while retaining flexibility [45]. Such architectures are particularly useful when atomically thin or nanoscale active materials require a mechanically robust support, improved interfacial coupling, or protection from environmental degradation.

Nanostructured and two-dimensional materials therefore represent a transition from conventional actuator-dominated design toward multifunctional platforms optimized for miniaturization, flexibility, sensing, and integration with energy-constrained electronics.

3.5 Functional classification for autonomous systems

While Table 1 compares intrinsic electromechanical coefficients, practical material selection also depends on mechanical compliance, durability, operating mode, fabrication compatibility, and system-level energy requirements. Different applications may prioritize flexibility, voltage sensitivity, fatigue resistance, or output power rather than maximizing a single material coefficient. Table 2 therefore classifies representative piezoelectric materials according to their functional roles in autonomous sensing systems.

Table 2 demonstrates that optimal material selection is application-driven rather than coefficient-driven. High-coupling ceramics such as PZT remain suitable for actuation and high-power harvesting, whereas polymers provide voltage sensitivity and mechanical compatibility for wearable and distributed sensing. Composite and nanoscale systems enable co-optimization of flexibility, miniaturization, and electromechanical response. Modern piezoelectric-material development is therefore increasingly guided by system integration, durability, and energy-autonomy requirements rather than by maximization of a single intrinsic coefficient.

Table 2. Functional classification and representative electromechanical properties of piezoelectric materials for autonomous Internet of Things (IoT) systems

Note: Values are representative of the cited literature and may vary with composition, dimensionality, processing, poling, measurement direction, substrate constraint, frequency, and characterization method. Values reported for nanoscale and two-dimensional materials should not be directly compared with bulk coefficients unless the coefficient definitions, normalization methods, and mechanical boundary conditions are equivalent.

Although piezoelectric materials directly convert mechanical deformation into electrical charge, the generated signal is typically intermittent, high-impedance, and poorly matched to electronic loads. Without proper conditioning, most of the harvested energy is lost. Efficient energy extraction and management are therefore essential for enabling autonomous sensing systems. This section is limited to the electrical-energy pathway from transducer output to regulated and stored power, including rectification, extraction, storage, duty cycling, and the system-level autonomy condition. Embedded inference methods and model-design considerations are discussed separately in Section 6.

5.1 Electrical nature of piezoelectric output

A piezoelectric device behaves electrically as a current source in parallel with an internal capacitance. Mechanical excitation produces alternating charge, resulting in an AC voltage whose amplitude depends on load impedance and excitation frequency [21]. Because the output impedance is typically very high, directly powering electronics leads to severe energy loss. As a result, the effective power delivered to electronics is often orders of magnitude lower than the theoretical electromechanical conversion predicted by material coefficients.

Maximum power transfer occurs only when the electrical load is matched to the internal capacitive reactance, which varies with frequency. Consequently, practical harvesting circuits must adapt to changing mechanical conditions rather than relying on fixed resistive loads [57].

5.2 Rectification and energy extraction circuits

The simplest interface consists of a diode bridge rectifier followed by a storage capacitor. However, voltage drops across diodes significantly reduce efficiency for low-amplitude signals [58]. To overcome this limitation, synchronized switching techniques were developed to actively control charge transfer.

Synchronized Switch Harvesting on Inductor (SSHI) circuits invert the voltage across the piezoelectric element at displacement extrema, dramatically increasing extracted power [57, 59]. Synchronous Electric Charge Extraction (SECE) circuits transfer energy directly into storage elements, making them suitable for irregular mechanical excitations [60]. These techniques can improve harvested power several times compared with passive rectification. However, their implementation increases circuit complexity and standby consumption, which must be carefully balanced against net energy gain in low-power IoT deployments.

Recent approaches employ adaptive and nonlinear interfaces capable of tracking optimal extraction conditions under variable vibration environments [61].

5.3 Energy storage and power regulation

Because mechanical excitation is intermittent, harvested energy must be accumulated before use. Storage elements typically include capacitors for short-term buffering and rechargeable batteries or supercapacitors for long-term operation [62]. Power management integrated circuits regulate voltage and enable cold-start operation when initial energy levels are extremely low. The electro-thermal reliability of the semiconductor switches used within these power-management circuits has also been examined using coupled Multiphysics simulation approaches, providing insight into thermal limits relevant to compact IoT power stages [63].

Efficient regulation is critical since IoT electronics often require stable DC voltages while the harvester provides irregular pulses. Duty-cycled operation is therefore commonly used, where sensing and communication occur only after sufficient energy has been accumulated [41].

5.4 System-level power budget and operation modes

The feasibility of a self-powered sensor depends on the balance between harvested energy and total system consumption [9]. In many low-power IoT platforms, wireless transmission requires substantially more energy per event than sensing or local computation, although the exact ratio depends on the radio protocol, payload, transmission power, and hardware platform.

Modern self-powered nodes therefore commonly employ event-driven and duty-cycled operation. The system remains in a low-power state until sufficient energy has been accumulated or a relevant mechanical event is detected, after which sensing, processing, and communication are performed according to the available energy budget. Reducing the frequency and duration of radio activity is particularly important in communication-dominated systems because it can significantly lower average power demand.

Hybrid harvesters combining piezoelectric, triboelectric, and photovoltaic mechanisms can further stabilize the energy supply under variable environmental conditions [64]. The present section focuses on the energy-management and operating-mode implications of these architectures. The signal-processing and machine-learning methods used to determine which events or features should be transmitted are discussed separately in Section 6.

5.5 Energy budget for autonomous inference

The feasibility of self-powered intelligent sensing depends on whether the usable harvested electrical energy exceeds the total energy required for sensing, computation, and communication during an operational cycle. The net operational energy can be expressed by Eq. (3):

$\eta_{\mathrm{PM}} E_{\text {harv}} \geq E_{\text {sense}}+E_{\text {infer}}+E_{\mathrm{tx}}$   (3)

where, $E_{\text {harv}}$ is the electrical energy generated by the piezoelectric transducer before power-management losses, $\eta_{\mathrm{PM}}$ is the efficiency of the power-management interface, and $E_{\text {sense}}$, $E_{\text {infer}}$, and $E_{\mathrm{tx}}$ are the energies required for data acquisition, embedded inference, and wireless transmission, respectively. Eq. (3) applies to one completed operational cycle and assumes that sufficient stored energy is available to execute that cycle; it does not describe the preceding cold-start or energy-accumulation transient. This cycle-level formulation also does not explicitly include continuous leakage and standby losses, which are incorporated in the time-averaged autonomy condition presented in Eq. (4).

Table 3 summarizes representative order-of-magnitude energy costs for low-power IoT operations. These values are intended for preliminary energy budgeting rather than as universal limits because consumption varies considerably with hardware, firmware, measurement boundaries, payload size, and operating conditions.

Table 3. Representative order-of-magnitude energy costs reported for low-power Internet of Things (IoT) operations

Note: The listed values represent approximate orders of magnitude compiled from the cited literature. Sensor and ADC energy may include different portions of the analog front end; tiny machine learning (TinyML) energy depends on the benchmark task and processing platform; and bluetooth low energy (BLE) energy depends on whether wake-up, channel switching, protocol overhead, and retransmission are included. Storage leakage is a continuous loss and is therefore more appropriately expressed as leakage current or power than as energy per operation.

Table 3 illustrates that wireless communication may remain the dominant energy term even when sensing and local computation are reduced to the microjoule range. The energy budget therefore favors architectures that reduce transmission frequency, radio-on time, and payload size. Improvements in power-management efficiency and reductions in storage or standby leakage may also produce substantial gains in the autonomy margin.

The detailed signal-processing and TinyML methods used to convert sensor waveforms into compact features, classifications, or event flags are discussed in Section 6. The present section is limited to quantifying the energy available for these operations and establishing the conditions under which they can be supported.

5.6 System-level autonomy condition

While previous sections discussed harvesting circuits and duty-cycled operation qualitatively, autonomous sensing ultimately depends on a simple quantitative requirement: the time-averaged usable harvested power must exceed the total average power required for sensing, computation, communication, and standby operation. This condition can be expressed as:

$\eta_{\mathrm{PM}} P_{\text {harv}}>f_s E_{\text {sense}}+f_i E_{\text {infer}}+f_t E_{\mathrm{tx}}+P_{\text {leak}}$    (4)

where, $P_{\text {harv}}$ is the average electrical power available at the output of the piezoelectric transducer before powermanagement losses, $\eta_{\mathrm{PM}}$ is the efficiency of the powermanagement interface (commonly implemented as a powermanagement integrated circuit, PMIC), $E_{\text {sense}}$, $E_{\text {infer}}$, and $E_{\mathrm{tx}}$ represent the energy required for sampling, local inference, and wireless transmission, respectively, and $f_s$, $f_i$, and $f_t$ denote the corresponding operation frequencies. $P_{\text {leak}}$ accounts for storage, standby, and other time-averaged parasitic losses.

Eq. (4) is a time-averaged rather than an instantaneous power balance. It is evaluated over an averaging interval sufficiently long to contain representative sensing, inference, communication, and idle periods. The operation frequencies $f_s$, $f_i$, and $f_t$ therefore represent average event rates over this interval. The formulation assumes statistically stationary or periodically repeatable excitation and operating conditions during the averaging period.

Frequency effects are included at the system level through the excitation-dependent value of $P_{\text {harv}}$ and through the operation frequencies multiplying the energy cost of each task. However, Eq. (4) does not explicitly resolve waveform-level dynamics, rectifier transients, storage-voltage ripple, cold-start behavior, or the initial charging of an empty storage element. These transient conditions require a time-dependent circuit and storage model. Accordingly, satisfying Eq. (4) indicates long-term average energy feasibility but does not, by itself, guarantee successful startup or uninterrupted operation during short-term excitation deficits.

As discussed in Section 2.4, $P_{\text {harv}}$ depends not only on the intrinsic piezoelectric response but also on dielectric permittivity, device capacitance, geometry, excitation frequency, mechanical loading, stiffness, damping, straintransfer efficiency, and boundary conditions. The factor $\eta_{\mathrm{PM}}$ subsequently accounts for losses associated with rectification, energy extraction, regulation, and storage interfacing.

For design comparison, the inequality in Eq. (4) can be expressed through the dimensionless autonomy ratio $\mathcal{A}$ (Eq. (5)):

$\mathcal{A}=\frac{P_{\text {usable}}}{P_{\text {demand}}}=\frac{\eta_{\mathrm{PM}} P_{\text {harv}}}{E_{\text {sense}} f_s+E_{\text {infer}} f_i+E_{\mathrm{tx}} f_t+P_{\text {leak}}}$   (5)

where, $P_{\text {usable}}=\eta_{\mathrm{PM}} P_{\text {harv}}$, and the denominator represents the total time-averaged power demand. A value of $\mathcal{A}<1$ indicates an average energy deficit, $\mathcal{A}=1$ represents the break-even condition, and $\mathcal{A}>1$ indicates an average energy surplus under the stated operating conditions. This ratio does not replace detailed transducer or circuit models; rather, it provides a common metric for comparing interventions made at different design layers.

This relation highlights that system autonomy is not governed by material performance alone. Increasing the effective piezoelectric response can raise $P_{\text {harv}}$ under appropriate mechanical and electrical conditions, but comparable or larger gains may be obtained by reducing the communication frequency $f_t$, lowering the inference energy $E_{\text {infer}}$, improving $\eta_{\mathrm{PM}}$, or minimizing $P_{\text {leak}}$. In communication-dominated deployments, reducing wireless transmissions may improve the autonomy margin more strongly than a modest increase in a single intrinsic piezoelectric coefficient.

Eq. (5) therefore links the principal design layers as follows:

•Material and device design influence $P_{\text {harv}}$.

•Circuit architecture determines $\eta_{\mathrm{PM}}$.

•Embedded intelligence determines $E_{\text {infer}}$ and $f_i$.

•Communication strategy determines $E_{\mathrm{tx}}$ and $f_t$.

•Storage and standby behavior determines $P_{\text {leak}}$.

Consequently, achieving sustained autonomous operation requires co-optimization across these layers rather than maximization of a single performance metric. This framework explains why high-output laboratory harvesters may still fail to support practical deployments: improvements in intrinsic material coefficients alone may not compensate for inefficient communication, poor interface efficiency, or excessive standby and storage losses.

Eq. (5) therefore establishes the energy envelope within which embedded inference must operate, while Section 6 examines how the available computational and communication budgets can be used efficiently.

5.7 Illustrative cross-layer comparison

To demonstrate how the autonomy ratio introduced in Eq. (5) can be used as a design tool, an illustrative comparison is performed between a material-level intervention and a communication-level intervention. The purpose of this example is not to reproduce a specific experimental device, but to show how changes made at different system layers can be evaluated using the same autonomy criterion.

Representative energy values within the ranges reported in Table 3 are considered:

$E_{\text {sense}}=2 ~\mu \mathrm{J}, E_{\text {infer}}=20 ~\mu \mathrm{J}, E_{\mathrm{tx}}=200 ~\mu \mathrm{J}$

with storage and standby losses represented by

$P_{\text {leak}}=5 ~\mu \mathrm{W}$

Sensing and local inference are assumed to occur at 1 Hz. A baseline piezoelectric device with $d_{33}=300$ pC/N is assumed to provide $100 \mu\mathrm{W}$ of usable average post-interface power under the specified mechanical excitation. At a wireless transmission frequency of 1 Hz, the corresponding average power demand is:

$P_{\text {demand}}=E_{\text {sense}} f_s+E_{\text {infer}} f_i+E_{\mathrm{tx}} f_t+P_{\text {leak}}=227 ~\mu \mathrm{W}$

Using Eq. (5), the baseline autonomy ratio is $\mathcal{A}_0=\frac{100}{227}=0.44$,  indicating an average energy deficit.

Using the material-to-power scaling established in Section 2.4, a material-level intervention can be estimated for a force-driven piezoelectric element. Holding force, geometry, capacitance, excitation frequency, and interface efficiency constant, usable power scales approximately as in Eq. (6):

$\frac{P_{\text {usable}, 2}}{P_{\text {usable}, 1}} \approx\left(\frac{d_{33,2}}{d_{33,1}}\right)^2$   (6)

This scaling relation assumes force, geometry, capacitance, excitation frequency, and interface efficiency remain constant, and is intended only as a first-order estimate within a single device design—it should not be used to compare different material systems or device architectures, since permittivity, capacitance, elastic compliance, electromechanical coupling, and interface losses can vary simultaneously across materials or geometries. Whenever device-level data are available, experimentally measured usable harvested power should be substituted directly into Eq. (5).

Increasing $d_{33}$ from 300 to 500 pC/N therefore gives an idealized scaling factor of $\left(\frac{500}{300}\right)^2 \approx 2.78$.

Accordingly, the usable average power increases from 100 $\mu \mathrm{W}$ to approximately 278 $\mu \mathrm{W}$. Since the operating frequencies and energy costs remain unchanged, the average demand remains 227 $\mu\mathrm{W}$, giving $\mathcal{A}_{d_{33}}=\frac{278}{227}=1.22$.

The material-level intervention therefore moves the node slightly above the break-even autonomy threshold.

A system-level intervention can instead retain the original $d_{33}=300$ pC/N while reducing the wireless transmission frequency from 1 to 0.1 Hz . Sensing and local inference are maintained at 1 Hz, but only one out of every ten inference outcomes is transmitted. The resulting average demand becomes:

$\begin{gathered}P_{\text {demand,event}}=(2)(1)+(20)(1)+(200)(0.1)+5 = 47 ~\mu \mathrm{W}\end{gathered}$

With the usable harvested power remaining at 100 μW, the corresponding autonomy ratio becomes $\mathcal{A}_{\text {event}}=\frac{100}{47}=2.13$.  

Table 4 summarizes the resulting autonomy ratios and illustrates, under the selected operating conditions, how a material-level increase in piezoelectric response compares with a communication-level reduction in transmission frequency.

Table 4. Illustrative comparison of material-level and communication-level interventions using the autonomy ratio

Note: The material-level power scaling is an idealized force-driven sensitivity estimate assuming unchanged force transfer, device capacitance, geometry, excitation frequency, mechanical boundary conditions, and interface efficiency. It should not be interpreted as a universal comparison between different materials or device structures.

Under the selected conditions, both interventions move the sensing node above the break-even threshold. However, reducing the transmission frequency produces the larger autonomy margin because wireless communication is the dominant energy-consumption term in the assumed power budget. The material-level intervention increases usable harvested power by approximately 178%, whereas reducing the transmission frequency by 90% lowers the total average demand by approximately 79%.

This comparison is a controlled sensitivity analysis rather than a universal ranking of materials or design strategies. The approximate $d_{33}^2$ dependence applies only when the applied force, geometry, effective capacitance, excitation frequency, mechanical boundary conditions, and interface efficiency are held constant. When different materials or device structures are compared, dielectric permittivity, capacitance, elastic compliance, electromechanical coupling, dielectric loss, resonance behavior, strain-transfer efficiency, and interface losses may change simultaneously. Therefore, experimentally measured usable harvested power should be inserted into Eq. (5) whenever device-level data are available, rather than estimating system feasibility from $d_{33}$ alone [21].

The example demonstrates the practical role of the cross-layer framework. Material improvements, interface optimization, inference reduction, communication duty cycling, and leakage mitigation can be compared using a common system-level metric. The dominant intervention depends on the application-specific energy budget: communication reduction is most effective in radio-dominated systems, whereas transducer and interface improvements become more important when mechanical input is weak, or communication is already infrequent.

5.8 Design-oriented architecture selection

The autonomy condition can be translated into practical design choices by identifying the term that dominates the energy budget in a given application. Rather than attempting to maximize all subsystem performance simultaneously, the design process should prioritize the principal material, circuit, computational, or communication constraint. Table 5 summarizes representative architecture choices for several operating scenarios. These recommendations are indicative rather than prescriptive and should be validated using application-specific excitation spectra, device measurements, and energy budgets.

Table 5. Representative architecture choices derived from the system-level autonomy condition

Note: Polyvinylidene fluoride (PVDF), lead zirconate titanate (PZT), synchronous electric charge extraction (SECE), synchronized switch harvesting on inductor (SSHI), bluetooth low energy (BLE).

Table 5 illustrates that architecture selection should be guided by the dominant energy constraint rather than by absolute material performance alone. In low-frequency, communication-dominated applications, reducing transmission activity may improve the autonomy margin more effectively than increasing a single piezoelectric coefficient. In continuously excited systems, stable energy extraction and impedance matching may become more important, whereas long-duration deployments may be limited primarily by material fatigue, interface degradation, and storage leakage. Engineering effort should therefore focus on the limiting term identified through Eq. (5), while recognizing that the preferred material, circuit, inference model, and communication protocol remain application-dependent.

espite rapid advances, long-term maintenance-free systems remain rare in real-world deployment.

The convergence of advanced piezoelectric materials, low-power electronics, and embedded machine learning has enabled the emergence of autonomous sensing nodes. However, large-scale deployment of self-powered intelligent IoT systems remains limited by interdisciplinary challenges spanning materials science, device engineering, circuit design, and data processing. Addressing these challenges requires coordinated optimization across all system layers rather than isolated improvements in individual components.

A recurring gap across the literature is the mismatch between laboratory demonstrations and field deployment constraints. Many prototypes report high peak voltages or power densities under controlled excitation yet omit long-duration fatigue, storage leakage, and realistic mounting variability—factors that can reduce net usable energy by orders of magnitude. Similarly, ML-enabled sensing studies often report accuracy without reporting inference energy, memory footprint, or performance drift under changing boundary conditions. Addressing these reporting gaps is essential for translating piezoelectric Artificial Intelligence of Things (AIoT) concepts into scalable and maintainable sensor networks.

7.1 Materials and mechanical reliability

Although high-performance ceramics provide strong electromechanical coupling, their brittleness limits durability under repeated deformation and impact loading. Flexible polymers improve mechanical compliance but may exhibit lower charge sensitivity, polarization relaxation, and temperature-dependent degradation. In composite and flexible devices, interface delamination, electrode cracking, changes in contact resistance, and inefficient strain transfer may degrade performance even when the intrinsic piezoelectric phase remains functional [17, 29, 41].

Reliability results must be interpreted according to the applied loading mode. Electrical fatigue refers to degradation during repeated electric-field cycling and is commonly quantified through changes in remanent polarization, dielectric response, strain, or piezoelectric coefficients. Mechanical fatigue concerns cyclic stress, strain, bending, vibration, or impact and may lead to cracking, depolarization, or delamination. Device-level durability testing additionally includes electrodes, bonding layers, substrates, encapsulation, and electrical interconnections. Consequently, cycle counts obtained under different loading modes should not be treated as directly equivalent lifetime measures.

Table 6. Representative fatigue and durability results reported for piezoelectric materials and devices

Note: Polyvinylidene fluoride (PVDF), lead zirconate titanate (PZT).

Table 6 summarizes representative fatigue and durability results across major piezoelectric material and device classes. The reported values are illustrative rather than universal because loading amplitude, cycling frequency, temperature, geometry, mounting conditions, electrode design, and failure criteria vary considerably among studies.

Table 6 shows that reported durability spans several orders of magnitude and cannot be interpreted independently of loading mode. Bulk PZT studies report electrical cycling up to approximately 10⁶ cycles, while selected polymer-composite systems have maintained sensing performance after 10⁶ mechanical cycles. By contrast, many nanogenerator demonstrations remain limited to shorter cycling intervals. Flexible devices may additionally fail through electrode cracking, interfacial delamination, or contact-resistance drift before substantial degradation of the active piezoelectric phase occurs [17, 29, 41, 74-77].

The severity of the validation gap can be illustrated by converting laboratory cycle counts into equivalent continuous operating time. A device excited continuously at 1 Hz experiences approximately $3.15 \times 10^7$ cycles per year, whereas operation at 10 Hz corresponds to approximately $3.15 \times 10^8$ cycles per year. Accordingly, $10^5$ cycles represent only approximately 27.8 h at 1 Hz and 2.78 h at 10 Hz. A 10-cycle test corresponds to approximately 11.6 days at 1Hz and 1.16 days at 10 Hz. Therefore, even apparently extensive laboratory tests may cover only a small fraction of the accumulated loading expected in continuously vibrating machinery or infrastructure applications.

The implications differ for event-driven wearables and sporadically activated sensors, for which the accumulated number of cycles may be considerably lower. Nevertheless, flexible devices introduce additional degradation mechanisms associated with bending radius, interfacial shear, electrode cracking, sweat and humidity exposure, temperature variation, washing, and packaging [17, 41]. In such systems, electrode and interface failure may dominate before a measurable loss of intrinsic piezoelectric polarization occurs.

A meaningful reliability assessment for autonomous piezoelectric IoT devices should therefore report the loading waveform and amplitude, cycling frequency, total cycle count, temperature, humidity, mounting and packaging conditions, electrode architecture, electrical load, and the criterion used to define failure. Output voltage alone is insufficient because it depends strongly on measurement impedance and device capacitance. Where applicable, studies should also report charge, power under a fixed electrical load, capacitance, leakage current, polarization or $d_{33}$, and contact resistance before and after cycling.

Recent materials development increasingly targets not only peak $d_{33}$, but also fatigue resistance, thermal and humidity stability, electrode-material interface integrity, and polarization retention under realistic cyclic loading. Future work should combine mechanically compatible electrodes, improved ceramic-polymer adhesion, robust encapsulation, accelerated aging, and multiphysics lifetime modeling. Such validation is necessary before short laboratory demonstrations can be extrapolated to deployment-relevant service periods.

7.2 Energy density and power stability

Ambient mechanical energy is inherently irregular and application-dependent. Even with efficient power-management circuits, insufficient excitation may interrupt sensing or computation. Hybrid harvesting approaches combining piezoelectric, triboelectric, and photovoltaic mechanisms offer a promising pathway to stabilize energy availability across varying environments. Another challenge lies in cold-start operation, where the system must accumulate enough energy to initiate computation. In low-frequency or sporadic excitation environments, the time-averaged harvested power may fall below the standby leakage power of storage elements, leading to a net energy deficit. Ultra-low-leakage storage elements and adaptive duty-cycling strategies will be required for reliable operation under weak mechanical inputs. Long-term deployments of vibration-powered sensor nodes show that harvested energy strongly depends on environmental variability and mounting conditions, reinforcing the need for adaptive harvesting strategies and hybrid energy management approaches [74, 75].

7.3 Energy-aware embedded intelligence

Machine-learning algorithms designed for conventional embedded systems often assume stable power supply and memory availability. In energy-harvesting devices, intermittent power may interrupt processing, requiring non-volatile memory and checkpointing strategies. Model compression, quantization, and event-driven inference must therefore be co-designed with sensing hardware.

Recent demonstrations of ultra-low-power inference hardware and energy-adaptive neural network architectures further confirm that embedded intelligence can operate within harvested energy budgets. Future TinyML frameworks may incorporate adaptive inference, where model complexity dynamically adjusts to available energy. Neuromorphic and spiking neural networks also offer potential for extremely low-power processing compatible with harvested energy levels.

7.4 System integration and standardization

Currently, piezoelectric IoT research lacks standardized datasets and performance metrics. Reported power density, sensitivity, and accuracy values are often measured under incomparable conditions, making fair evaluation difficult. Establishing benchmark excitation profiles and unified reporting protocols would significantly accelerate technological maturity. Reporting should include excitation spectrum, mechanical boundary conditions, load matching, inference energy per decision, and storage leakage over time. Recent efforts toward shared datasets and benchmarking methodologies for self-powered sensing and edge intelligence aim to improve reproducibility and fair comparison across reported systems [78, 79].

Integration with wireless protocols and edge computing architectures also remains challenging. Reliable large-scale deployment requires interoperability between sensing nodes, edge processors, and cloud analytics while maintaining strict energy budgets.

7.5 Sustainable manufacturing and future outlook

Environmental considerations increasingly influence material selection and device fabrication. Environmental impact must be evaluated not only at the material level but across the lifecycle, including encapsulation polymers, electronics disposal, and recyclability of composite architectures [80]. Lead-free piezoelectric ceramics, recyclable substrates, and low-temperature processing methods will be essential for sustainable mass production [81]. Emerging research also highlights recyclable substrates, environmentally friendly processing, and scalable printed electronics as key elements for sustainable large-scale deployment [82].

Looking forward, further progress in piezoelectric IoT systems will likely depend on the coordinated development of smart materials, adaptive power management, and embedded intelligence. As these layers mature, sensors are expected to move beyond passive measurement toward greater on-node data interpretation and decision support, provided that reliability, benchmarking, and system-level validation keep pace. Achieving this requires holistic design approaches in which materials, electronics, and algorithms are optimized jointly rather than independently.

The developments discussed above demonstrate that piezoelectric sensing is transitioning from component-level functionality to system-level autonomy. The following section summarizes the key contributions of this review and outlines its implications for future research.