Energy Harvesting for Battery-Free IoT Sensors: Solar, Thermoelectric, Vibration and RF Power Budgets and PMIC Design (2026)

Energy Harvesting for Battery-Free IoT Sensors: Solar, Thermoelectric, Vibration and RF Power Budgets and PMIC Design (2026)

Energy Harvesting IoT: Power Budgets and PMIC Design for Battery-Free Sensors (2026)

A 25 cm² indoor solar cell under office light delivers somewhere in the tens to low hundreds of microwatts. That is roughly a thousand times less than the same cell in sunshine, yet it can run a Bluetooth sensor forever. It can also fail to start at all on a dim Monday morning. Energy harvesting IoT works or fails on that narrow margin. The harvester is rarely the hard part. The storage element, the PMIC cold-start threshold and the firmware’s reaction to a sagging rail decide whether a device runs for ten years or dies in its first winter.

This matters now. Batteries are the dominant maintenance cost of large sensor fleets. Meanwhile, 3GPP has finished the core Release 19 work on Ambient IoT, and silicon vendors ship radios with cold-start budgets in the low hundreds of microjoules. This post gives you the math and the design rules to decide whether a node can be battery-free, and how to build it.

What this covers: power densities by source, the harvester-to-radio chain, PMIC selection, storage sizing for dark periods, worked BLE and LoRaWAN budgets, energy-aware firmware, failure modes and a decision matrix by environment.

Context and Background

Battery-free sensing is older than the IoT label. EnOcean’s kinetic light switches have shipped for two decades, and the wireless protocol behind them was ratified as ISO/IEC 14543-3-10 in 2012. Its full title is “Wireless Short-Packet (WSP) protocol optimised for energy harvesting”, and it covers OSI layers 1 to 3. Passive RFID and NFC tags are battery-free by design: the reader’s field powers them. What changed in the 2020s is that ordinary sensing nodes, the kind that measure temperature, humidity, occupancy, vibration or CO₂, can now run on ambient energy instead of on a coin cell.

Three trends pushed this forward. First, sleep currents fell. Silicon Labs’ xG22E family quotes 1.40 µA in EM2 deep sleep with 32 kB RAM retained and 0.17 µA in EM4. It also boots from a zero-energy state in about 8 ms using roughly 150 µJ. Second, dedicated harvesting PMICs matured. Parts from TI, Analog Devices and e-peas now cold-start from a few hundred millivolts and a few microwatts. Third, indoor photovoltaics tuned for LED spectra moved from labs into products, from dye-sensitized cells (Exeger Powerfoyle) to organic cells (Epishine).

Standards bodies followed. The Ambient IoT Alliance launched in February 2025 with founding members including Atmosic, Infineon, Intel, PepsiCo, Qualcomm, VusionGroup and Wiliot. 3GPP defined Ambient IoT as a new device class in Release 19, based on the TR 38.769 study. The Bluetooth SIG is still exploring energy-harvesting features rather than shipping a dedicated specification, according to ABI Research’s 2025 review.

The incumbents are unchanged: primary lithium cells. A well-designed cellular node using PSM and eDRX can already reach multi-year life, as covered in our guide to cellular IoT power saving with PSM, eDRX and RAI. So harvesting must justify itself. It wins on total cost of ownership when nodes are numerous, hard to reach or expected to outlive any battery. For foundational IoT context, see our complete technical guide to IoT. For the primary standard text, the IEC preview of ISO/IEC 14543-3-10 is the authoritative reference.

The Energy Harvesting IoT Reference Architecture

An energy harvesting IoT node is a chain of five stages: a transducer turns ambient energy into electricity, a PMIC cold-starts and charges storage at the transducer’s maximum power point, a storage element buffers energy across dark or idle periods, a regulator feeds the load, and firmware throttles work to match what the chain can sustain.

Energy harvesting IoT reference architecture from ambient source through PMIC and storage to MCU and radio

Figure 1: The energy harvesting IoT signal and power chain, with the feedback path from storage voltage to the firmware scheduler.

The diagram reads left to right. Ambient light, heat, motion or RF hits a transducer: a PV cell, thermoelectric generator (TEG), piezoelectric or electromagnetic harvester, or rectenna. The PMIC’s cold-start circuit bootstraps from an empty state. A boost charger with maximum power point tracking (MPPT) then fills the storage element. A buck converter or LDO feeds the MCU, sensor and radio. The crucial path is the feedback loop: a voltage monitor reports storage state to firmware, and an energy-aware scheduler decides how much work the radio may do.

Power density sets the ceiling, not the design

Start with physics, because it bounds everything downstream. Standard outdoor sunlight (AM1.5G) is 100 mW/cm² of incident power. Typical indoor artificial light is about 0.05 to 1 mW/cm², or 50 to 1,000 µW/cm², spanning roughly 200 to 2,000 lux, according to Jagadamma and Wang’s 2021 review in Frontiers in Chemistry. A 2026 arXiv review of self-powered microsystems puts incident indoor PV power at roughly 55 to 300 µW/cm² under 200 to 1,000 lux, depending on spectrum. Indoor light is about a thousand times weaker than sunlight.

Conversion efficiency then cuts that further. Lab perovskite cells have reported about 36% under white LED light, per the Frontiers review. Commercial products sit lower. Epishine’s first production organic cell was specified at 15% efficiency for 20 to 1,000 lux in 2021. Multiply 55 to 300 µW/cm² incident by 15% to 25% efficiency and you get roughly 8 to 75 µW/cm² of electrical output. That is a derived estimate, not a datasheet figure, and it assumes the cell faces the light.

Other sources are harder to normalise. TEG output depends on the temperature difference across the module, which is always smaller than the difference between heat source and air. Vibration harvesters can produce milliwatts on a tuned shaker table but almost nothing off-resonance. RF harvesting falls with the square of distance. The 2026 review warns that power figures across modalities are not directly comparable, because each paper defines efficiency differently. Treat any single “µW/cm²” number as a starting point for your own measurement, not as a design input.

The PMIC decides whether the node ever wakes up

A harvesting PMIC solves three problems a normal regulator ignores. It must start from a fully discharged state with a tiny, low-voltage source. It must hold the source at its maximum power point. And it must protect storage from over-charge and over-discharge, while drawing only a few hundred nanoamps itself.

Cold start is the first gate. TI’s BQ25570 needs 600 mV typical (700 mV maximum) and about 15 µW at its input to cold-start from a storage voltage below 100 mV. Once running, it harvests down to 100 mV. The e-peas AEM00940/AEM00941 cold-starts at about 380 mV and 3 µW. The newer e-peas AEM13920, a dual-source part, cold-starts from 275 mV and at least 1.5 µW. Its datasheet notes a typical cold-start time of about three minutes at that minimum. For TEGs, Analog Devices’ LTC3108 uses a 1:100 step-up transformer to start from inputs as low as 20 mV.

MPPT is the second gate. The common method is fractional open-circuit voltage. The PMIC briefly disconnects the source, samples its open-circuit voltage (Voc) and regulates the input to a fixed fraction of it. The BQ25570 samples for 256 ms every 16 s. It regulates to 80% of Voc for solar and 50% for TEGs, or to a resistor-set ratio. The AEM13920 offers ratios from 35% to 85% with sampling periods from 128 ms to 16.38 s. The 50% figure for TEGs is not arbitrary: a TEG behaves like a voltage source with series resistance, so peak power transfer happens at half its open-circuit voltage.

Quiescent current is the third gate, and it is easy to underestimate. The BQ25570 draws 488 nA typical in full operation. The AEM13920 draws 645 nA in supply state with its buck enabled, and about 10 nA in shipping mode. At 3 V, 500 nA is 1.5 µW. That is a meaningful fraction of a 10 µW harvest budget, and it never sleeps.

Not every “low-power PMIC” is a harvesting PMIC

Read the input specification before shortlisting a part. Nordic’s nPM2100, announced in January 2025, is an excellent PMIC with a 150 nA quiescent current and up to 95% efficiency at 50 mA. But it is a boost regulator for primary batteries, such as AA, AAA, LiMnO₂ coin and zinc-air cells, with a 0.7 to 3.4 V input. It has no cold-start-from-microwatts circuit and no MPPT. It belongs in the “primary cell done well” column of the decision matrix, not the harvesting column. The distinction matters because many hybrid designs pair a harvester PMIC with a primary-cell fallback. The e-peas AEM00940 has a dedicated primary-battery input for exactly that case.

Walk-through: Building an Energy-Neutral Power Budget

Energy neutrality means that over a full cycle (usually a day, sometimes a week) energy harvested equals or exceeds energy consumed. Storage only needs to bridge the gap between the two curves. The budget below walks the steps in Figure 2. Every number in the worked examples is illustrative unless it is attributed to a datasheet. Replace the numbers with your own measurements before committing a design.

Energy-neutral power budget flow for battery-free IoT sensors from harvest to storage sizing

Figure 2: The energy-neutral budgeting loop. Harvest minus PMIC drain and storage leakage gives usable energy. That must cover the sleep floor plus tasks, then storage is sized for the longest dark period, and cold start is checked at worst-case light.

The loop has two exits that most first designs miss. The first is storage leakage. It is subtracted before you know whether tasks fit, because it runs continuously. The second is the cold-start check at worst-case light. A node that is energy-neutral on average can still deadlock if it ever fully discharges and the morning light is below the PMIC’s cold-start threshold.

Step 1: Daily harvest for indoor PV

Take a 50 × 50 mm cell (25 cm²) on an office wall. Assume 300 lux, ten lit hours on weekdays and darkness at night and on weekends. Interpolating the 55 to 300 µW/cm² range above gives about 80 µW/cm² incident at 300 lux. At an assumed 15% cell efficiency, that is 12 µW/cm², or 300 µW across the cell. Derate for angle, dust and spectrum mismatch by 50%, and by an assumed 80% PMIC efficiency at this low input. The result is about 120 µW while the lights are on (illustrative).

Over ten hours that is 120 µW × 36,000 s ≈ 4.3 J per weekday. Averaged across 24 hours, it is about 50 µW. Averaged across a seven-day week with dark weekends, it is about 36 µW. Use the weekly figure, because weekends are real.

Step 2: Subtract the always-on floor

Next, list everything that draws current all the time. Take a PMIC quiescent of about 0.5 µA, an MCU in deep sleep with RTC of about 1.5 µA, and a sensor in standby of about 0.3 µA. At 3 V, that is about 7 µW. The MCU figure is consistent with the xG22E EM2 number above; the others are illustrative. Then add storage leakage, which can easily be the largest single term. We return to that in Step 4.

Step 3: Price each task in microjoules

Energy per task, not average current, is the right unit. A sensor read and conversion might cost on the order of 100 µJ (illustrative). A cold boot of the xG22E from power-on reset costs about 150 µJ, versus about 17 µJ from EM4, per Silicon Labs’ documentation. That ninefold gap is why harvesting firmware avoids full power-off between events when wake-ups are frequent.

For the radio, compute energy from time-on-air and transmit power. A Bluetooth LE advertising event on three channels typically costs a few tens of microjoules on current SoCs. We use 50 µJ below as an illustrative round number; measure your own with a power profiler. LoRa is different, because airtime depends heavily on the spreading factor (SF). Using the standard Semtech time-on-air formula for a 12-byte application payload (25 bytes on air with LoRaWAN MAC overhead), at 125 kHz bandwidth, coding rate 4/5 and an 8-symbol preamble:

Spreading factor Time on air (computed) TX energy at ~130 mW draw (illustrative)
SF7 61.7 ms ≈ 8 mJ
SF9 205.8 ms ≈ 27 mJ
SF10 411.6 ms ≈ 54 mJ
SF12 1,482.8 ms ≈ 193 mJ

The airtimes come from the formula. The energy column assumes roughly 40 mA at 3.3 V during transmission. That is a typical order of magnitude for a sub-GHz LoRa transceiver at about +14 dBm, not a verified datasheet value. Receive windows add a few more millijoules per uplink.

Step 4: Run the two budgets

The BLE node advertises every 10 s and reads its sensor every 60 s:

Item Energy per event Rate Average power
Always-on floor — continuous ≈ 7 µW
BLE advertising 50 µJ every 10 s 5 µW
Sensor read 100 µJ every 60 s ≈ 1.7 µW
Load total ≈ 14 µW

The LoRaWAN node sends once every 15 minutes:

Item SF7 SF12
Always-on floor ≈ 7 µW ≈ 7 µW
Uplink plus RX windows (≈ 10 mJ vs ≈ 200 mJ per 900 s) ≈ 11 µW ≈ 222 µW
Sensor reads every 60 s ≈ 1.7 µW ≈ 1.7 µW
Load total ≈ 20 µW ≈ 231 µW

Against a weekly harvest of about 36 µW, the BLE node and the SF7 LoRaWAN node both close with margin. The SF12 node does not close at all. The spreading factor alone swings the budget by more than an order of magnitude. A battery-free LoRaWAN node is therefore only as good as its link margin and the network’s adaptive data rate behaviour. Place it where ADR can hold SF7 to SF9, or do not make it battery-free.

Now add storage leakage. A small supercapacitor’s leakage current is often specified in the single-digit microamp range after the cell has conditioned (illustrative; check your part’s datasheet and its measurement conditions). At 3.6 V, 5 µA is 18 µW. That is more than the entire BLE load. This is the single most counter-intuitive result in indoor harvesting: the storage element, not the harvester or the radio, often sets the budget.

Step 5: Size storage for the longest dark period

A weekend is roughly 62 hours without office light, from Friday evening to Monday morning. At a 14 µW load, that is 14 µW × 223,000 s ≈ 3.1 J. A capacitor’s usable energy between two voltages is E = ½C(V₁² − V₂²). Between 3.6 V and 2.2 V, that is about 4.06 × C joules. So 3.1 J needs about 0.77 F before leakage. Add the leakage term and the requirement climbs further, because leakage itself drains the cap over 62 hours. This is why lithium-ion capacitors (LICs), small rechargeable lithium cells and solid-state thin-film batteries dominate indoor designs that must bridge weekends. A supercapacitor alone suits outdoor solar, with short nights, or event-driven nodes that tolerate going dark.

Energy-aware firmware state machine for energy harvesting IoT nodes using storage voltage thresholds

Figure 3: An energy-aware firmware state machine. Storage voltage selects between full operation, log-only, checkpoint-and-stretch, and hibernate, and the wake period adapts to the voltage trend.

The state machine in Figure 3 is the firmware half of the budget. Above V_HIGH the node senses, transmits and flushes any backlog. Between V_LOW and V_HIGH it senses and logs locally but does not spend energy on the radio. Between V_CRIT and V_LOW it checkpoints state to non-volatile memory and stretches its wake period. Below V_CRIT it hibernates and the PMIC disconnects the load. In every lit state, it adjusts the next wake interval from the voltage slope, not just the voltage level.

Step 6: Encode the policy in firmware

The pseudo-code below implements Figure 3 in C. It assumes a PMIC that exposes storage voltage through an ADC channel or a status pin, FRAM or flash for checkpointing, and an RTC wake source. Threshold values are placeholders: derive yours from the storage chemistry’s discharge curve and the radio’s peak current.

/* Energy-aware scheduler for a battery-free sensor node.
 * Thresholds are illustrative placeholders; derive from your
 * storage discharge curve, ESR and radio peak current. */
#include <stdint.h>
#include <stdbool.h>

#define V_HIGH_MV   3300  /* enough for TX burst + margin        */
#define V_LOW_MV    2900  /* sense/log only, no radio            */
#define V_CRIT_MV   2500  /* checkpoint, then stretch interval   */
#define T_MIN_S       10
#define T_MAX_S     3600

typedef struct {
    uint32_t seq;
    uint16_t backlog;          /* readings stored, not yet sent */
    uint32_t period_s;
    uint16_t last_mv;
} nv_state_t;

extern nv_state_t *nv;          /* lives in FRAM / NVM page      */
extern uint16_t read_vstore_mv(void);
extern int16_t  sensor_read(void);
extern void     log_reading(int16_t v);   /* FRAM ring buffer    */
extern bool     radio_send_batch(uint16_t max_items);
extern void     rtc_sleep_s(uint32_t s);  /* EM2/EM4-type sleep  */

static uint32_t clamp(uint32_t v, uint32_t lo, uint32_t hi) {
    return v < lo ? lo : (v > hi ? hi : v);
}

void scheduler_tick(void) {
    uint16_t mv = read_vstore_mv();      /* sample under light load */
    int32_t slope = (int32_t)mv - (int32_t)nv->last_mv;
    nv->last_mv = mv;

    if (mv >= V_HIGH_MV) {
        log_reading(sensor_read());
        nv->backlog++;
        /* send at most what the reserve above V_LOW can pay for */
        if (radio_send_batch(nv->backlog > 8 ? 8 : nv->backlog))
            nv->backlog = 0;
    } else if (mv >= V_LOW_MV) {
        log_reading(sensor_read());       /* no radio */
        nv->backlog++;
    } else if (mv >= V_CRIT_MV) {
        nv->seq++;                        /* checkpoint already in NV */
        nv->period_s *= 2;                /* back off hard */
    } else {
        rtc_sleep_s(T_MAX_S);             /* PMIC may cut load anyway */
        return;
    }

    /* Adapt: shorten period when charging, lengthen when draining */
    if (slope > 20)       nv->period_s = nv->period_s * 3 / 4;
    else if (slope < -20) nv->period_s = nv->period_s * 3 / 2;
    nv->period_s = clamp(nv->period_s, T_MIN_S, T_MAX_S);

    rtc_sleep_s(nv->period_s);
}

Three details matter more than the structure. First, sample the storage voltage under a known, light load. A supercapacitor’s ESR makes the voltage drop during a radio burst, and a reading taken mid-burst will trigger false brownouts. Second, the state struct lives in non-volatile memory, so a total power loss resumes from the last checkpoint rather than from zero. Third, the hysteresis between thresholds prevents the node from oscillating at a boundary. That oscillation is a classic way to burn the reserve with repeated boots. Pair this scheduler with hardware brownout detection and a watchdog, using the patterns in our post on watchdog timers, brownout detection and fault recovery.

Intermittent computing: when power loss is normal

Some designs go further and accept power loss as routine. With tiny storage, such as a few hundred microfarads on an RF-powered tag, the node may lose power mid-task many times a minute. Intermittent computing techniques checkpoint volatile state to non-volatile memory just before the supply collapses, then restore it on the next boot. FRAM is popular here because writes are fast and need no erase cycle. Silicon Labs notes, by contrast, that a flash page erase on its parts takes about 12.9 ms. That is an eternity when the rail is collapsing. The rule is simple: make every task idempotent or resumable, and never let an erase-heavy operation start below V_LOW.

The Other Sources: Thermoelectric, Vibration, Kinetic and RF

Indoor PV dominates commercial battery-free IoT sensors because light is common, predictable and cheap to harvest. The other modalities win in narrower niches, and each has its own physics trap.

Thermoelectric generators

A thermoelectric generator for IoT converts a temperature difference into voltage through the Seebeck effect. The open-circuit voltage is roughly V = N × S × ΔT, where N is the number of thermocouples, S the Seebeck coefficient per couple and ΔT the temperature difference across the module. The module has internal resistance R, so maximum power into a matched load is V²/4R. That is why MPPT regulates TEGs to half their open-circuit voltage.

The trap is thermal, not electrical. ΔT across the module is a fraction of the difference between the hot surface and ambient air. The heat must flow through the contact interface, the module and the heat sink into the air. The heat sink’s thermal resistance to still air usually dominates. Halve the effective ΔT and you quarter the power. A pipe at 60 °C in a 25 °C room may leave only a few kelvin across the module with a small passive heat sink. Because power scales with ΔT², designers spend as much effort on the heat sink and mounting as on electronics.

Voltages are low as a consequence. The LTC3108 datasheet’s example uses a 30 mm, 127-couple module with 2 Ω resistance and rates its typical circuit for ΔT of 1 to 20 °C. At the low end, the TEG produces tens of millivolts. That is why the part uses a 1:100 step-up transformer to start from 20 mV. Note its trade-off: the LTC3108 draws 3 mA from the source at 20 mV input with no load, and about 6 µA of quiescent current on its auxiliary rail. That is acceptable for a steam pipe and unacceptable for a body-worn device.

TEGs suit steam and hot-water lines, motor housings, district-heating valves and radiator valves: anywhere a stable gradient persists for hours. They are poor where the gradient reverses or vanishes, such as outdoor equipment that tracks air temperature.

Vibration harvesters

Piezoelectric and electromagnetic vibration harvesters are resonant. A proof mass on a spring couples to the machine. Power peaks sharply at the resonant frequency and falls off within a bandwidth of roughly f₀/Q. Classic linear-oscillator analysis shows output scaling with proof mass and with the square of acceleration amplitude, and falling with frequency for a given acceleration. High Q gives high peak power but a narrow band. A harvester tuned to 50 Hz mains-synchronous vibration may produce little if the motor runs on a variable-frequency drive.

Published lab results look impressive. The 2026 arXiv review cites designs reaching tens of milliwatts at 0.5 g. But those are absolute powers from specific, often bulky prototypes under controlled excitation. They do not transfer to a random bearing housing. In practice, vibration harvesting fits fixed-speed rotating machinery, where the frequency is known and stable. It pairs naturally with condition monitoring: the same vibration you harvest is the signal you are measuring. Bistable and multi-magnet designs widen the band, at the cost of complexity.

Kinetic harvesting: one action, one telegram

Kinetic harvesters convert a single mechanical action into a burst of energy. EnOcean’s PTM 215 pushbutton module uses an electro-dynamic generator driven by an energy bow with 1.8 mm travel and a typical operating force of 8 N. On the 868.3 MHz version, one press-and-release sends three subtelegrams at 125 kbps. The user manual limits operation to 6,000 actuations per hour and a 1% duty cycle.

The design principle is the opposite of PV. There is no storage and no budget across time. Each event must pay for its own transmission, so the protocol is built around very short packets with redundancy. That is what ISO/IEC 14543-3-10 standardises. Kinetic harvesting fits switches, door contacts and anything driven by a human or a mechanism. It cannot support periodic sensing.

RF harvesting and backscatter

RF harvesting looks attractive because radio is everywhere. The Friis equation shows why it rarely powers a sensor at a useful distance. In the US 902 to 928 MHz band, FCC §15.247 limits a transmitter to 1 W conducted with up to 6 dBi of antenna gain. That is about 4 W EIRP (36 dBm). Free-space path loss at 915 MHz is about 31.7 dB at 1 m, 41.2 dB at 3 m and 51.7 dB at 10 m.

With a 0 dBi receive antenna, received power is therefore about +4.3 dBm (2.7 mW) at 1 m, −5.2 dBm (0.3 mW) at 3 m and −15.7 dBm (27 µW) at 10 m. Those are computed best-case free-space values before rectifier losses. Powercast’s P2110B, a 902 to 928 MHz harvester, operates down to −12 dBm input. So a dedicated, regulation-maximum transmitter powers it only within several metres. Ambient Wi-Fi or cellular signals, at much lower levels, deliver far less.

This is why the most successful RF-powered IoT does not try to be a normal radio. Wiliot’s IoT Pixels, at 2.8 × 4.4 cm and 0.2 mm thick, harvest RF from nearby devices and speak Bluetooth. 3GPP’s Ambient IoT work centres on very low-power devices, including backscatter designs. Those reflect a reader’s carrier instead of generating their own, which cuts transmit energy by orders of magnitude. According to the 3GPP work plan as summarised in August 2026, the Release 19 core and performance work is complete. Conformance testing was targeted for December 2026, and Release 20 “phase 2” items run into 2027. Treat ambient-IoT cellular as infrastructure that operators must deploy, not something a device maker can rely on today.

NFC belongs in the same family. A passive NFC tag is powered by the reader’s 13.56 MHz field during a tap. It suits commissioning, logging cold-chain data and reading out a dead node, but not unattended periodic telemetry.

Choosing an Energy Harvesting IoT Source: Decision Matrix by Environment

Pick the energy source from the site, not from the datasheet. Figure 4 encodes the selection order that most deployments follow.

Decision tree for choosing an energy harvesting IoT source by deployment environment

Figure 4: Source selection by environment. Reliable light favours indoor PV. Stable gradients favour TEGs, fixed-frequency vibration favours tuned harvesters, discrete actions favour kinetic switches, and reader infrastructure enables RF or backscatter. Otherwise, use a primary battery done well.

Environment Best source Typical storage Radio fit Main risk
Lit office, retail, hospital Indoor PV (a-Si, DSSC, OPV, perovskite) LIC, thin-film or small Li cell BLE, Zigbee GP, sub-GHz SF7 Weekends, desk drawers, dimming
Outdoor, sunny Outdoor PV Supercap or LiFePO4 LoRaWAN, NB-IoT, LTE-M Winter, snow, low-temperature charging
Pipes, motors, HVAC valves TEG Supercap or Li cell BLE, sub-GHz Heat sink sizing, gradient loss
Fixed-speed rotating machinery Tuned piezo or EM Supercap BLE, proprietary Frequency drift, VFDs
Switches, doors, handles Kinetic None EnOcean, Zigbee GP Event-only, no periodic data
Warehouse, retail with readers RF or backscatter Small cap BLE tags, RFID, future 3GPP A-IoT Reader density, regulation
Dark, sealed, remote None: primary cell Li-SOCl₂ or LiMnO₂ Any, with PSM or eDRX Passivation, replacement cost

The indoor PV technology column deserves nuance. Amorphous silicon has long been the cheap default for calculators and indoor products. Dye-sensitized cells, like Exeger’s Powerfoyle, and organic cells, like Epishine’s, are flexible and tuned for LED spectra. Exeger reports production capacity of 2.5 million m² per year across two Stockholm factories. Perovskites hold the lab records but face stability and lead-content questions.

Supply risk is real, too. Ambient Photonics, which had partnered with Google on a solar-powered remote and with Lenovo on a self-charging keyboard, wound down operations and auctioned its equipment in early 2026. Dual-source your cell, or design the mechanical outline so a second vendor’s cell fits.

Trade-offs, Gotchas, and What Goes Wrong

Cold-start deadlock. This is the most common energy harvesting IoT field failure. The node runs for months, a long dark spell fully discharges storage, and the morning light is just below the PMIC’s cold-start threshold. The harvester produces power the PMIC can no longer use. Mitigations include a lower-threshold PMIC, a hibernate state that disconnects the load well before empty, and a small primary cell as a cold-start backstop.

Supercapacitor leakage and ageing. Leakage is highest right after charging and depends strongly on temperature and voltage. Datasheet figures are often measured after hours of conditioning at room temperature. Capacitance and ESR also degrade over years, especially near the rated voltage and at high temperature. Derate the voltage and measure leakage at your real operating point.

Temperature. Lithium cells must not be charged below their specified low-temperature limit. A PMIC that keeps charging a frozen Li cell damages it. Supercaps and LICs tolerate cold better, but their ESR rises, which deepens the voltage dip during a radio burst. Outdoor designs need temperature-aware charge cut-off.

ESR and peak current. A radio transmit burst of tens of milliamps through a high-ESR storage element causes an instant voltage drop. That can trip brownout even when stored energy is ample. Add a low-ESR buffer capacitor at the load, or start transmissions only above a voltage that includes the expected ESR drop.

MPPT overhead at very low light. Fractional-Voc sampling disconnects the source, and the sampling logic draws current. At the lowest light levels, the MPPT overhead can exceed the gain from tracking. Some PMICs let you lengthen the sampling period for this reason.

Optimistic lab budgets. Bench tests under a desk lamp at 1,000 lux overstate real installations by several times. Measure lux at the actual mounting points across a full week, including weekends and holidays, before committing a cell size.

Firmware that ignores energy. A standard RTOS application that retries failed uplinks immediately will drain a harvesting node. Bounded retries, batching and backoff tied to storage voltage are mandatory, not optional.

Practical Recommendations

Begin every battery-free IoT sensor project with a site survey, not a schematic. Log lux, temperature gradients or vibration spectra at the real mounting points for at least a week. Build the energy budget in microjoules per task, include PMIC quiescent and storage leakage as first-class terms, and size storage for the longest realistic dark period with margin. Choose the radio configuration for the budget: BLE advertising and LoRaWAN at low spreading factors fit indoor PV; SF12 and frequent cellular attaches usually do not.

Choose the PMIC by cold-start voltage and power at your worst-case input, not by peak efficiency. A part that starts from a few microwatts at under 400 mV, such as the e-peas AEM00940/AEM00941 or AEM13920, protects you against deadlock better than a more efficient part with a higher threshold. For TEG designs below about 100 mV, consider a transformer-based starter like the LTC3108. Keep primary-battery PMICs such as the nPM2100 for nodes that cannot harvest at all.

Then make firmware energy-aware from day one. Use voltage-gated states, adaptive intervals and non-volatile checkpoints. Treat the first energy harvesting IoT field pilot as a measurement exercise. Log storage voltage with every uplink, so the fleet tells you where the budget is tight.

  • [ ] Week-long site survey at real mounting points, weekends included
  • [ ] Budget in µJ per task, including PMIC Iq and storage leakage
  • [ ] Storage sized for the longest dark period at worst-case temperature
  • [ ] PMIC cold-start verified at minimum light or ΔT, with a timed test
  • [ ] Voltage-gated firmware states with hysteresis and NV checkpoints
  • [ ] Load-side low-ESR buffer for radio bursts
  • [ ] Storage voltage reported in every uplink for fleet telemetry
  • [ ] Second-source option for the harvester cell

Frequently Asked Questions

What is energy harvesting in IoT?

Energy harvesting in IoT means powering sensors from ambient energy instead of replaceable batteries. The main sources are light (indoor or outdoor photovoltaics), heat gradients (thermoelectric generators), vibration (piezoelectric or electromagnetic harvesters), mechanical actions (kinetic switches) and radio waves (RF harvesting or backscatter). A harvesting PMIC cold-starts from the tiny source, tracks its maximum power point and charges a storage element such as a supercapacitor or small lithium cell. Firmware then scales its activity to the energy available, so the node stays energy-neutral over a day or week.

How much power can an indoor solar cell produce?

Typical indoor lighting delivers roughly 50 to 300 µW/cm² of incident light power at 200 to 1,000 lux, depending on the spectrum. After cell efficiency, typically 15% to 25% for commercial indoor cells and higher in lab records, electrical output is roughly 8 to 75 µW/cm² while lights are on. A 25 cm² cell at 300 lux might average a few tens of microwatts over a week, once nights and weekends are counted. That is enough for BLE advertising every few seconds or a low-spreading-factor LoRaWAN uplink every 15 minutes. These are estimates; measure your site.

Can a LoRaWAN sensor run without a battery?

Yes, if the link stays at low spreading factors. A 12-byte LoRaWAN uplink takes about 62 ms on air at SF7 but about 1.48 s at SF12, using the standard time-on-air formula at 125 kHz. Transmit energy scales roughly with airtime, so an SF12 node needs more than twenty times the energy per uplink. Indoor PV can support SF7 to SF9 at intervals of minutes. SF12 nodes typically need outdoor solar, a larger harvester or a battery. Placing gateways so ADR can hold a low SF is the key design lever.

Are supercapacitors better than batteries for energy harvesting?

Neither is universally better. Supercapacitors tolerate hundreds of thousands of cycles, charge at very low currents and handle cold well. But their energy density is low and their leakage can exceed a sensor’s entire load. Small lithium cells, lithium-ion capacitors and thin-film batteries hold far more energy with lower self-discharge. That makes them better for bridging nights and weekends indoors. Lithium cells cannot be charged below their low-temperature limit, however. Use supercaps for short dark periods and event-driven nodes, and lithium-based storage for multi-day autonomy.

What does an energy harvesting PMIC do?

An energy harvesting PMIC does four jobs a normal regulator does not. It cold-starts from an empty state using a tiny, low-voltage source, from 275 to 600 mV depending on the part. It holds the source near its maximum power point, usually by sampling open-circuit voltage and regulating to a fraction of it. It charges and protects the storage element with over-charge and over-discharge thresholds. And it does all this while drawing only hundreds of nanoamps. Examples include TI’s BQ25570, Analog Devices’ LTC3108 and the e-peas AEM family.

What is 3GPP Ambient IoT?

3GPP Ambient IoT is a new class of extremely low-power cellular devices, including battery-free backscatter devices. It was defined in Release 19, building on the TR 38.769 study on solutions in NR. According to the 3GPP work plan, the Release 19 core radio work completed in 2025. Performance requirements completed in March 2026, and conformance testing was targeted for late 2026. Release 20 phase-two work items run into 2027. Ambient IoT needs network-side support from operators, so it is not yet a deployable option for most products.

Further Reading

References

  1. Texas Instruments, BQ25570 Ultra Low Power Harvester Power Management IC with Boost Charger and Nanopower Buck Converter, datasheet. https://ti.com/lit/gpn/BQ25570
  2. e-peas, AEM00940/AEM00941 datasheet. https://e-peas.com/aem0094x_datasheet
  3. e-peas, AEM13920 datasheet and launch announcement (29 August 2023). https://oldsite.e-peas.com/aem13920_datasheet ; https://e-peas.com/news/e-peas-introduces-first-energy-harvesting-pmic-to-handle-two-independent-energy-sources-at-the-same-time/
  4. Analog Devices, LTC3108 Ultralow Voltage Step-Up Converter and Power Manager, datasheet. https://www.analog.com/media/en/technical-documentation/data-sheets/3108fc.pdf
  5. Nordic Semiconductor, “nPM2100 PMIC extends battery life of primary cell powered Bluetooth LE products”, January 2025. https://nordicsemi.com/Nordic-news/2025/01/nPM2100-PMIC-extends-battery-life-of-primary-cell-powered-Bluetooth-LE-products
  6. Silicon Labs, Energy Harvesting Applications with EFR32xG22E. https://docs.silabs.com/energy-harvesting/1.0/energy-harvesting-start
  7. Jagadamma, L. K. and Wang, S., “Wide-Bandgap Halide Perovskites for Indoor Photovoltaics”, Frontiers in Chemistry 9:632021, 2021. https://www.frontiersin.org/articles/10.3389/fchem.2021.632021/full
  8. Lin, L. et al., “Energy Harvesting for Self-Powered Microsystems: A Critical Review of Materials, Power Management, and System Integration”, arXiv:2609.20157. https://arxiv.org/abs/2609.20157
  9. US eCFR, 47 CFR §15.247, Operation within the bands 902–928 MHz, 2400–2483.5 MHz and 5725–5850 MHz. https://www.law.cornell.edu/cfr/text/47/15.247
  10. Powercast, P2110B Powerharvester Receiver datasheet. https://www.mouser.com/datasheet/2/329/P2110B_Datasheet_Rev_3-3159474.pdf
  11. EnOcean, PTM 210 / PTM 215 User Manual. https://www.rohm.com/documents/11401/2139988/PTM+210+PTM+215+PTM+210U+PTM+210J+User+Manual.pdf
  12. IEC, ISO/IEC 14543-3-10 Wireless Short-Packet (WSP) protocol optimised for energy harvesting, preview. https://webstore.iec.ch/p-preview/info_isoiec14543-3-10%7Bed1.0%7Den.pdf
  13. 3GPP, TR 38.769, Study on solutions for Ambient IoT in NR (Release 19). https://3gpp.org/dynareport/38769.htm ; work-plan status summary: https://whatthespec.net/sos/c/aiot-detail/where-the-work-stands
  14. CNX Software, “The Ambient IoT Alliance aims to promote and develop standards for batteryless IoT devices”, February 2026. https://www.cnx-software.com/2026/02/23/ambient-iot-alliance-promote-develop-standards-for-batteryless-iot-devices-energy-harvesting/
  15. Lookout Santa Cruz, “Lights out for Ambient Photonics”, January 2026. https://lookout.co/lights-out-for-ambient-photonics-pioneering-solar-cell-technology-amid-funding-shortfalls/story
  16. Exeger, Powerfoyle for sensors. https://www.exeger.com/powerfoyle/applications/sensors/
  17. Wiliot, What is energy harvesting for Bluetooth devices. https://www.wiliot.com/what-is-energy-harvesting-for-bluetooth-devices
  18. ABI Research, “Energy Harvesting with Ambient IoT: Progress in Connectivity Standardization”, 2025. https://www.abiresearch.com/research-highlight/ambient-iot-connectivity-update-2025

By Riju — about

Comments

No comments yet. Why don’t you start the discussion?

Leave a Reply

Your email address will not be published. Required fields are marked *