Selecting an appropriate energy storage device is the first step in designing an energy harvester circuit, and based on the selected energy storage device, the power management circuit should be designed with the target of optimizing the overall power efficiency. The systems targeted for micro-power energy harvesting applications typically have additional stringent constraints on die area and usage of external components. As energy harvested from the PV module is intermittent and the maximum power that it can provide may be much less than the needed peak power during sensing and data transmission, an energy storage device is needed. Although the proposed H₂ sensor may consume less than 20 μA on average for sensing and data transmission, the peak power consumption during wireless data transmission is much higher. The target wireless transceiver consumes 3.3 mA and 2.8 mA during data transmission and reception [3]. Miniaturized supercapacitors, NiMH microbatteries and lithium-based batteries are different solutions that have been used in wireless sensor nodes frequently. The battery should have enough capacity to provide enough power for a few hours’ operation of the sensor, even without energy harvesting. In addition to capacity, other characteristics, such as size, peak discharge current, nominal operating voltage, V_EOC, V_EOD, cycle life and leakage, are very important factors to design an energy-efficient sensor that relies on energy harvesting.

Although miniaturized supercapacitors, such as GZ115F of CAP-XX [10], can provide much higher peak charge and discharge currents and have much longer cycle lives in comparison with the rechargeable batteries, they have some disadvantages that make them inappropriate for our application. The main disadvantage of supercapacitors is their high leakage currents that can be even larger than the current provided by the PV cell under reduced illumination level. In addition, they have much lower energy capacity compared with rechargeable batteries of similar sizes. Among lithium batteries, state-of-the-art thin film lithium batteries, such as MEC125 of Infinite Power Solutions [11], are the best candidates for micro-power energy harvesting applications thanks to their low leakage currents, long cycle lives and high discharge currents. The main disadvantage of these batteries is their high nominal voltage levels. As thin film lithium batteries have nominal voltages of more than 3.8 V, additional DC-DC converters are required to use these batteries for ULP applications. The target V6HR microbattery has a nominal voltage of 1.2 V and can provide a peak discharge current of 18 mA that is much higher than the required current during data transmission. This battery has a large nominal capacity of 6.2 mAh, while its diameter and height are 6.8 mm and 2.15 mm, respectively. As in the proposed gas sensor, all electronic circuits, including the external wireless transceiver, can work with a sub-1.2 V supply voltage, so additional step-down or step-up circuits are not required. Table 1, compares the main characteristics of the proposed V6HR battery with GZ115F and MEC125. The main disadvantage of this NiMH battery is its lower cycle life compared to supercapacitors and thin film lithium batteries. In addition, this battery has a relatively low leakage current. It loses at most 20% of its charge during the first month, which is much lower than the discharge current of supercapacitors.

The block diagram of the target H₂ gas sensor, including the required external components and the main blocks of the proposed solar energy harvester, is depicted in Figure 1. Four miniaturized nanowire solar cells have been connected in series to be used as the energy source. The main tasks of a solar energy harvester block are battery charging, battery management and energy management. The direct charging method proposed in [6] has been used for charging the target NiMH microbattery with very high efficiency. In this approach, battery voltage (V_bat) is compared with open circuit voltage of the PV module (V_pv) using a dynamic comparator, and if it is lower, the battery is connected to the PV module to store harvested energy. If V_pv drops below V_bat, the switch between V_pv and V_bat is kept turned off to avoid battery discharge through the PV module. In the direct charging method that has been deployed for solar energy harvesting, the maximum power point voltage (V_mpp) of the PV module should be close to the V_EOC of the battery to achieve high efficiency. Although P_mpp and I_SC of the PV module changes considerably in different lightning conditions, V_oc and V_mpp of the PV module does not change significantly. In Figure 2, deliverable power of the PV module in different illumination levels has been simulated using the electrical model of the target PV module [12]. In AM1.5, solar intensity is almost 1 mW/mm². Although, in 10% of AM1.5, deliverable power is reduced by roughly the factor of 10; V_oc drops from 2.35 V to 1.95 V.

The second main task of the energy harvester circuit is battery management. As the charging current is limited by the PV module, the battery is charged in either standard or trickle charging modes, depending on the illumination conditions [2]. In trickle charging mode, the battery is charged by a small current, and it can be continuously charged after reaching V_EOC. However, in standard charging mode, the battery charging should be stopped after reaching V_EOC to avoid permanent damage to the battery. In addition, to avoid full discharge of the battery that reduces the cycle life of the battery, the battery should be disconnected from the sensor when it drops to V_EOD. In order to avoid overdischarge or overcharge of the battery in standard charging mode, V_EOC and V_EOD voltage levels are detected by this circuit.

Finally, the last task of the energy harvester circuit is energy management. Since, under reduced light intensity, the power delivered by the PV module may be lower than the average power consumption of the sensor, the energy harvester circuit should reduce power consumption of the microsystem to avoid full discharge of the battery during this period. Under reduced light intensity, the microsystem continues its autonomous operation at a lower speed and with a lower duty cycle. The proposed energy harvester circuit reliably measures energy stored in the target battery and scales the power consumption of the microsystem up or down according to the measurement results. In order to reduce the power consumption and speed of digital circuits, operating frequency is scaled down and data transceiver is activated by a lower duty cycle. Bias currents are also scaled down to reduce power consumption and the speed of analog circuits. In order to reliably estimate the energy stored in the battery, battery voltage is measured when the battery is discharged by a high discharge current.

The PV module, with the total area of 28 mm², can provide a maximum power P_mpp of 2.88 mW at its V_mpp in AM1.5 illumination [1]. However, as harvested solar energy may be as small as a few micro-watts under reduced light intensities, the power consumption of this circuit should not be higher than a few hundred nanowatts to have high efficiency during these periods. As neither V_pv nor V_bat changes rapidly, comparator and voltage level detector blocks are activated every few seconds to minimize average power consumption. This circuit consumes less than 350 nW, which is mainly due to the clock generator and digital control unit that are always active.

Pd nanowire grid fabricated on a silicon wafer has been used for gas sensing. Each grid consists of 14 Pd nanowires, half of which are covered with a passivation layer to prevent hydrogen from reaching the nanowire [8]. These coated nanowires, which are only sensitive to temperature, have been used as reference nanowires, while the remaining nanowires, which are sensitive to both temperature and H₂ concentration, have been used as sensing nanowires. This circuit measures the conductance change of the sensing nanowires in comparison to the reference nanowires. In addition to eliminating the effects of temperature by first order, the measurement is only sensitive to the ratio of conductance of nanowires, instead of their absolute values, and as a result, a higher accuracy is achievable. Although the temperature effect is eliminated by first order, there are still second order effects, as the temperature coefficient of nanowire resistance changes according to H₂ concentration. In order to compensate this second order effect, temperature is measured using an integrated temperature sensor [9]. H₂ sensing accuracy can be increased by incorporating the measured temperature during sensor calibration. An incremental analog to digital converter (ADC) [13] converts the measured temperature and H₂ concentration to 12-bit digital values. Power consumption and conversion time of this ADC can be reconfigured according to the energy stored in the battery to match the amount of available power, resulting in an adaptive, autonomous sensor. The power management circuit dynamically reduces the operating frequency of digital circuits and the bias currents of analog circuits in this sensor interface circuit under reduced light intensity. However, ADC conversion time increases as a result.

Emerging wireless sensors that are powered by micro-power energy harvesting sources have more stringent energy requirements compared to traditional wireless sensors. Apart from minimizing total power consumption of the sensor, which is mainly achieved by duty cycling operation of the sensor, they should have low peak power and ultra-low standby current. Low peak power ensures that miniaturized batteries with limited peak discharge currents can be used to power up the circuit. Ultra-low standby current guarantees that the average power consumption of the sensor can be minimized by heavily duty cycling sensing and data transmission. The main factors impacting power consumption of a wireless transceiver are supply voltage, carrier frequency and receiver sensitivity. The power consumption of transceiver can be reduced by operating at lower supply voltages. Although most of the wireless transceivers work with at least 1.8 V supply voltage, ultra-low power wireless transceivers with sub-1.2 V voltage, such as TZ1053 [3] or ZL70250 [14], are more suited to our application, thanks to their lower power consumption during data transmission and reception. These transceivers operate at sub-1 GHz frequency bands and have much lower peak-power and standby power, compared to state-of-the-art 2.4 GHz transceivers.

The second important factor is the carrier frequency. Some of the factors that affect choice of carrier frequency are operation range, power consumption, transmission data rates and antenna size. Although 2.4 GHz protocols, such as Zigbee, have been used extensively for wireless sensing applications, sub-1 GHz wireless systems offer several advantages for ultra-low power, low data rate applications. These transceivers have less power consumption for the same operating range, thanks to reduced attenuation rates and blocking effects at lower frequencies. Besides requiring higher power for the same link budget, higher traffic in the 2.4 GHz frequency band increases the interference in this frequency band.

Finally, receiver sensitivity also affects the power efficiency. A narrower bandwidth creates higher receiver sensitivity and allows efficient operation at lower transmission rates. Overall, all radio circuits running at higher frequencies, including low-noise amplifiers, power amplifiers, mixers and synthesizers, need more current to achieve the same performance as lower frequencies.

Although these sub-1 GHz transceivers have some disadvantages, such as larger size of antennas and lower data rates, overall, they are more suited for our application. A higher data rate can improve the energy efficiency (energy per bit) of the transceiver for high bandwidth applications with large data payloads. In fact, overall power consumption of a wireless transceiver is not only a factor of physical layer items, such as radio architecture, carrier frequency and antenna choice, but is also a function of the amount of time that the radio needs to run in order to transport the payload data over the air. Transmission time depends on the data rate and the protocol overhead to establish and maintain the communication link. For example, a Zigbee transceiver that has a 250 kbps data rate can send 1 MB of data almost five-times faster than a TZ1053 that has a 50 kbps data rate. However, in ultra-low power applications, typically, data payloads are small and a 50 kbps data rate is more than enough. In addition, as standard protocols, such as Zigbee or Bluetooth, offer highly sophisticated link and network layers and have a large protocol overhead, they are not very efficient for sending small data payloads.

The proposed energy harvester is based on the direct charging scheme [6], in which a dynamic comparator compares V_bat with V_pv in a timely manner. The dynamic comparator and required control signals can be seen in Figure 3.

After turning off the switch between V_pv and V_bat, the first inverting gain stage gets reset by activating the reset control signal, to establish a common mode voltage of V_cm at the V_{C_Top} node. Meanwhile, V_{C_Bot} is connected to V_pv by activating the Sample_V_pv control signal, and the difference between V_pv and V_cm is sampled on the sampling capacitor. In the next step, by deactivating the reset and activating Sample_V_bat control signal, V_{C_Bot} is connected to V_bat, and as a result, V_{C_Top} changes to V_cm + (V_bat – V_pv). Outputs of the first and second gain stages (V_o1 and V_o2) change according to the new V_{C_Top}, and V_o2 is latched and stored as V_Comp, by enabling the latch_enable control signal. If V_Comp gets 1, it means that V_bat is higher than V_pv, and the switch should be kept turned off to avoid battery discharge, but if V_Comp gets 0, the PV module starts to charge the battery by turning on the switch. After the switch is turned on, V_pv will follow V_bat during battery charging.

As the operating voltage of the PV module is determined by the battery, end-to-end efficiency from the PV module to the battery is reduced when battery voltage diverges from the V_mpp of the PV module during battery charging. Different maximum power point tracking (MPPT) strategies have been deployed in energy harvesting applications to maximize the amount of harvested energy. As the V_mpp of the PV module varies with incident light conditions, an efficient MPPT scheme ensures that the maximum power is extracted from a PV module at any given time. Many complex and accurate MPPT schemes have been investigated for large solar harvesting systems [15]; however, when the PV module is small and can only provide a few micro-watts under reduced light intensity, only low-overhead schemes that incur very little power overhead can be good candidates. In [16], several low-overhead MPPT approaches, including fractional open-circuit voltage (FOC), fractional short circuit current (FSC) and hill-climbing techniques, have been discussed that can be deployed in inductive and SC DC-DC converters. Although in direct charging V_pv always follows V_bat and no MPPT scheme can be deployed, nevertheless, it achieves even higher efficiency than competing SC and inductive DC-DC converters [6]. In addition, the die area is much smaller than SC and inductive DC-DC converters, as large passive components, such as pumping capacitors, or inductors have not been used in this architecture.

When the target V6HR NiMH microbattery is fully charged, the battery voltage is close to V_EOC; as the battery is discharged by a low current, V_bat drops to its nominal value and remains almost constant, up to getting close to the fully discharged state. However, if the battery is discharged by a high current, as can be seen in the discharge curve of the battery in Figure 4 [2], V_bat drops immediately and voltage drop depends on the remaining charge of the battery. During wireless data transmission, the battery is discharged by a high current, close to 5 CA (C being the 1 hour charge or discharge current). As can be seen in Figure 4, if discharge capacity is less than 20%, V_bat is higher than 1.1 V during this period; however, if discharge capacity is higher than 20%, V_bat drops from 1.1 V down to 900 mV, depending on the remaining charge of the battery.

In our circuit, battery voltage is detected during this period to accurately estimate the energy stored in the battery. Depending on the detected V_bat, the power-performance of the integrated electronic circuits and operation duty cycle of the wireless transceiver are reconfigured to guarantee autonomous operation of the sensor. If V_bat is high enough, the sensor can operate at its highest performance: solar energy harvester and sensor interface circuits work at their highest speed, and measurement results are sent to the base station every 10 seconds. If V_bat is not high enough, the sensor interface circuit and wireless transceiver are activated with a lower duty cycle to minimize total power consumption of the sensor. If V_bat is close to V_EOD, these blocks should get deactivated temporarily to avoid full discharge of the battery.

The voltage level detector (LD) in Figure 5 is used to determine the battery voltage level [6]. After circuit startup, battery voltage is checked to make sure that it is more than the end of discharge threshold voltage (V_EOD) and can power up the circuit. If V_bat is less than V_EOD, it means that the battery is fully discharged and cannot provide enough power for the electronic circuits. In this situation, the PV module continuously charges the battery by keeping the switch closed. As soon as V_bat passes V_EOD, LD starts its normal operation, comparing V_bat with specified threshold voltages in a timely manner and updating the V_{bat_level} as a result. In order to generate a bandgap reference voltage, 25 substrate PNP transistors have been used as Q₁ and Q₂ in a common-centroid layout [6]. These transistors are biased with a 100 nA current source to generate V_BE1 and V_BE2 in non-overlapping Φ₁ and Φ₂ phases, and the SC circuit sums up V_BE1 and (V_BE1 – V_BE2) with appropriate coefficients. When V_bat reaches V_EOC, the switch between the PV module and the battery is turned off to avoid overcharge of the battery.

This SC circuit detects when V_bat passes the V_L specified in Equation (1) by setting the V_{bat_level} output. By using a variable γC capacitor, different voltage levels between V_EOD and V_EOC can be detected to estimate the remaining charge of the battery. In Equation (1), α, β and γ coefficients are the ratios of tunable αC, βC and γC capacitors. These variable capacitors have been implemented using a matrix of 10 fF metal-insulator-metal (MIM) capacitors that have been used as unity capacitors.

As V_BE1 is complementary to absolute temperature (CTAT) and (V_BE1V_BE2) is proportional to absolute temperature (PTAT), different CTAT, PTAT or temperature-independent reference voltages (V_ref) can be built by proper selection of αC and βC capacitors. After that, by modifying the γC capacitor, target voltage levels can be detected. As can be seen in Table 2, by modifying βC and γC capacitors, different battery voltages, starting from V_EOD of 0.9 V up to V_EOC of 1.5 V, can be detected. In order to have a temperature-independent voltage reference, αC should be modified according to the selected βC capacitor. By detecting the battery voltage between 900 mV and 1.1 V, the remaining charge of the battery can be estimated according to the discharge curve of the battery in Figure 4.

After measuring V_bat and estimating the energy stored in the battery, the operating frequency of digital circuits and bias currents of analog circuits are reconfigured according to the remaining charge of the battery. In the current-starved ring oscillator in Figure 6a, the frequency is determined by R_L and C_L and can be specified as Equation (2). In Equation (2), K₁ is a constant value that depends on the number of inverter stages in the ring oscillator [17]. By using a digital resistive trimming network to modify R_L, F_osc is reconfigured by an energy harvesting circuit, according to the measured V_bat. When V_bat is low, R_L is increased to decrease F_osc. Four target operating frequencies have been considered, corresponding to four voltage levels that are detected by the LD block, ranging from 125 kHz, corresponding to V_L0, to 1 MHz corresponding to V_L3.

In addition to a fixed 10 nA beta-multiplier (BM) current reference that has been used to provide the required bias current for a solar energy harvester, a switched-capacitor beta-multiplier (SCBM) current source has been designed to generate frequency-proportional bias currents. In this circuit, which can be seen in Figure 6b, SC resistors have been used instead of regular resistors to generate frequency-proportional bias currents [18]. For example, I_ADC can be specified as in (3): (3)

In Equation (3), K₁ and K₂ are constant values that depend on the ratio between the widths of transistors in Figure 6b. K₁ depends on the ratio between M₇ and M₄, and K₂ depends on the ratio between M₅ and M₆. The current ripple of I_bias is minimized by using a large decoupling capacitor, C₂, and two complementary branches charging and discharging C₃ and C₄ capacitors in non-overlapping Ø₁ and Ø₂ clock phases with F_osc frequency [18]. As F_osc is determined by the oscillator, I_bias scales dynamically by changing the frequency of the oscillator. In addition, this current source can be easily deactivated by turning off the Ø₁ and Ø₂ clocks.

In ultra-low power analog circuits, such as the discrete time incremental ADC that has been used here, bias currents should be high enough to guarantee correct operation at the target operating frequency. Generated bias currents (I_ADC and I_Temp) have been used for the sensor interface circuit. If V_bat is low, F_osc is decreased to reduce the power consumption of digital circuits in the ADC. When the system is operating at a lower frequency, lower bias currents can be used to reduce the power consumption of analog circuits in the sensor interface circuit. By using these frequency-proportional bias currents, power consumption of the sensor can be scaled down dynamically by reducing F_osc.

The digital control unit (DCU) activates the comparator and LD blocks in a timely manner, to check the charging status of battery. As neither V_bat nor V_pv change rapidly, these blocks are activated every few seconds to minimize their average power consumption. After activating each block, DCU generates the required control signals for related SC circuits. Standard complementary metal-oxide-semiconductor (CMOS) logic has been used to implement DCU. The power consumption of DCU is mainly determined by the low frequency 20-bits counter that activates the comparator and LD in a timely manner. Table 3 presents die area and power consumption of individual blocks, operating at 1 MHz as its highest clock frequency. As can be seen, the simulated total power consumption of the circuit is mainly determined by the clock generator, bias circuit and DCU blocks, which are always active. Although comparator and LD blocks consume considerable power during their active time, nevertheless, as they are activated every 10 seconds, their average power consumption is less than 1 nW. When V_bat is low, the average power consumption of the clock generator and DCU blocks is reduced by operating at a lower clock frequency. By reducing the operating clock frequency, the simulated power consumption of clock generator and DCU is reduced roughly linearly, reaching to 23 nW and 13 nW, respectively, at 125 kHz.

The proposed sensor interface circuit is depicted in Figure 7. An incremental ADC [13] measures H₂ concentration and temperature at different times. Individual Pd nanowires that have between 7 kΩ and 9 kΩ resistance should be biased, with a minimum bias voltage of 50 mV for 10 seconds. These nanowires have been represented by NW_ref and NW_sense in Figure 7. The conductance of NW_sense nanowires may change by 20% in comparison with NW_ref nanowires as H₂ concentration varies from zero to 30% [7]. The voltage around the reference nanowire, V_R = (V₁ – V₂), is used as the reference voltage, while the voltage of the sensing nanowire (V₂) is used as the input voltage for the following incremental ADC in consecutive non-overlapping clock phases. This ADC can achieve 12 bits resolution by using 200 fF metal-insulator-metal (MIM) capacitors, as C₃, C₄ and C₅, and a conventional two-stage amplifier with at least 75 dB gain for the integrator [9]. As a 1 MHz clock has been used for ADC and since this ADC needs 2^N cycles for N-bits conversion, a 12-bit conversion takes nearly 4 ms. After finishing gas sensing, temperature is measured to further increase the accuracy of gas sensing using an integrated temperature sensor. In the proposed integrated temperature sensor, substrate PNP transistors have been used to generate proportional to absolute temperature (V_ptat) and temperature-independent reference (V_ref) voltages, and the ADC converts (V_ptat/V_ref) to a 12-bit digital value. In Figure 7, by using a 50 fF MIM capacitor as C₁ and a 360 fF MIM capacitor as C₂, a temperature-independent reference voltage of approximately 310 mV has been generated. Similar to gas sensing, n-bit digital representation of (V_ptat/V_ref) is stored in the ADC counter and sent to a wireless transceiver after temperature sensing. The accuracy of temperature sensing is mainly limited by mismatch between Q₁ and Q₂ and nonlinearity in temperature dependence of V_BE1 and (V_BE1V_BE2). Although these errors can be minimized by using dynamic methods presented in [19] to reach ±0.1 °C accuracy, such power-consuming techniques are not needed here. The proposed low power temperature sensor can achieve ±1 °C accuracy by only calibrating C₂ and I_Temp at room temperature [9]. When the sensor interface circuit operates at a lower frequency, Q₁ and Q₂ transistors are biased with a lower I_Temp bias current to reduce the average power consumption of the circuit.