Millimeter-sized smart sensors reveal that a solar refuge protects tree snail Partula hyalina from extirpation

Smart solar sensor design

To prevent interference with the movements of the highly mobile E. rosea predators, we developed a custom smart solar sensor using the Michigan Micro Mote (M³) platform^27,36. The M³ platform consists of a family of chips that can be integrated together through die-stacking in various ways, allowing its functionality to be customized. M³ achieves this degree of miniaturization by directly stacking bare-die chips, thus avoiding individual chip packaging, and custom-designed low-power circuits, reducing consumption to only 228 nW. The resulting systems can be powered for >1 week by a chip-scale battery³⁶ measuring only 1.7 × 3.6 × 0.25 mm. For the solar sensor, we selected chips from this set with the following functionalities and stacked them as shown in Fig. 4: (1) two custom-designed thin-film lithium-chemistry batteries³⁷, each with 8-µAh capacity and 4.2-V battery voltage, connected in parallel; (2) a power management chip to generate and regulate the three supply voltages used by the M³ chips from the battery supply voltage; (3) a microprocessor chip containing an ARM Cortex-M0³⁸ processor that executes the program controlling the sensor and 8 kB of SRAM for storing program and sensor data; (4) a short-range (5 cm) radio chip with on-chip antenna for retrieving data from the sensor; (5) a decoupling capacitor chip for stabilizing supply voltages; (6) a harvester chip that up-converts the voltage from the photovoltaic (PV) cells to the battery voltage and regulates battery charging; (7) a temperature sensor chip; (8) an inactive spacer chip that provides physical separation between the PV cell, which is exposed to light, and the remainder of the chips below it, which must be blocked from light; and (9) a PV chip for harvesting solar energy, containing also a small PV cell for receiving optical communication.

The battery chips measured 1.7 × 3.6 × 0.25 mm while the remaining chips were 1.05 mm wide, 150 μm thick and varied in length from 1.33 to 2.08 mm. The chips were stacked in staircase fashion (Fig. 4a) using die-attach film and connected electrically using wire bonding with gold 18-μm diameter wire. The radio die extended beyond the other chips at the back to expose the antenna. The chips communicated using a common bus protocol, called M-bus³⁶. The final chip stack was encased in epoxy (Fig. 4b). The top portion of the sensor was encased with clear epoxy to allow light penetration, thereby enabling energy harvesting and optical communication. The bottom portion was encapsulated with black epoxy to protect the sensitive electronics from light. Finally, the entire sensor was coated with 4 μm of parylene. The sensor was tested to withstand immersion in brine at pressures up to 600 atm for 1 h and in saline solution for 2 weeks.

The principal approach to reduce the M³ sensor’s power consumption is to duty-cycle its operation, meaning the processor executes code briefly (ms range) every 10–60 min and is in “sleep mode” for the remainder of the time. Sleep power is highly optimized to ~100 of nW using a number of recently developed circuit techniques^39,40,41. In active mode, the processor is operating and obtains and stores sensor data. The processor clock frequency was set to 80 kHz, and at 0.6 V supply, the power consumption was 1.0 μW. In sleep mode, the processor and logic are power-gated⁴², and only the SRAM, timer, optical receiver, and power management remain on, reducing the power consumption to only 160 nW. The 10-min sleep mode period length was selected to amortize the high power in active mode and minimize overall power consumption while retaining a sufficiently small sensor acquisition interval for the proposed study. The average power consumption of the entire sensor including all peripherals was 228 nW, and in tests, it was able to operate on a battery charge alone for 1 week. With PV-based harvesting, the sensor becomes energy autonomous at light levels >850 lux. For this study, the sensors were retrieved and recharged using a light station after each deployment.

Although duty-cycling lowers the average current draw from the battery, it limits measurements to times when the sensor is awake. This raised a particular difficulty for measuring the solar ecology of snails where continuous light monitoring is essential, preventing the use of duty-cycling. Typical light-sensing circuits monitor the current from a photodiode and consume ~μW power^{<a data-track="click" data-track-action="reference anchor" data-track-label="link" data-test="citation-ref" aria-label="Reference 43" title="Alhawari, M., Albelooshi, N. & Perrott, M. H. A 0.5 V 43}, a load that would deplete the batteries in only a few hours. Hence the light intensity had to be monitored during sleep mode. To achieve this without substantially increasing the sleep mode current draw, we observed that the harvester circuit inherently integrates and quantizes the harvested energy from the photovoltaic (PV) cell in a manner proportion to the ambient light level and can be modified to function as a light sensor readout circuit.

To up-convert the output voltage of the PV cell (250–450 mV) to that of the battery (3.9–4.2 V), the harvester performs a series of voltage doublings⁴⁴ using the circuit shown in Supplementary Fig. 6. Each voltage doubling circuit consists of two chains of inverters, configured as a ring oscillator. The two oscillators are coupled through on-chip MIM capacitors and are connected to the supplies V_in and V_double, as shown. During one oscillation cycle, each capacitor experiences two different configurations. When the input to its driving inverters is high, a capacitor is placed between V_in and ground (GND), i.e., in parallel with the PV cell, which charges it with a finite amount of charge. When its driving inverter inputs are switched low, the capacitor is placed between V_in and V_double, and it delivers the received charge to V_double, thereby up-converting the voltage from the PV cell. The amount of charge that is transferred per cycle is kept constant by the frequency regulation circuit. If the PV cell is exposed to intense light and produces a high current, the regulation circuit increases the frequency by reducing the delay of the voltage-controlled delay element to maintain a constant charge transfer per cycle. Conversely, if the light level drops, the regulation circuit slows the oscillation frequency.

As a result, the frequency of oscillation is proportional to the PV current to the first order. And, because the current of the PV cell is proportional to the light intensity, the oscillation frequency is a measure of the instantaneous ambient light level. To obtain the light dose over a sleep mode time period, we added a low-power counter (shown in Supplementary Fig. 6), which records the number of oscillations during this period, thereby integrating its total light dose. Each active-mode period, the microprocessor reads the counter value, resulting in a light sensor code, and resets the counter. The counter operates at a low supply voltage of 0.6 V, which reduces its power consumption by ~9× compared to a standard supply of 1.8 V. This allowed us to implement a 24 bit counter with negligible power consumption (5 nW or 2.2% of total average power). The resulting sensors continuously monitor the light level and record a light-dose code for every 10 min interval. The addition of the counter constitutes a relatively small change in the harvester circuit and allows light monitoring without additional chips or an increase in battery capacity or sensor size.

Sensor testing and calibration

Because the harvester oscillation frequency is dependent on temperature and battery voltage, these parameters are stored by the processor in SRAM along with the light sensor code. After data retrieval, the code is then converted to light intensity using a model that accounts for the temperature and battery voltage dependency. To construct this calibration model, four sensor nodes were measured at six light levels (0.5, 1, 5, 10, 50, and 100 klux) and four temperatures (25, 35, 45, and 55 °C), and four battery voltages (3.9, 4.0, 4.1, and 4.2 V); a total of 96 measurements were made for each sensor. After averaging the light sensor codes across the four sensors, a multidimensional, piecewise linear model was extracted to establish the relationship between the recorded digital code and the light intensity at a particular temperature and battery voltage (Supplementary Fig. 7). To calibrate the model for each fabricated sensor, we measured the light sensor code, temperature sensor code and battery voltage sensor code in controlled conditions (temperature: 25, 45, and 55 °C; light: 5 klux; battery voltage: 4.1 V) for each sensor. We then applied three-point calibration of the temperature sensor and one-point calibration of both the battery voltage and light sensors. The calibration conditions were selected based on the expected temperature and battery operating range in the field and on what the discriminating light intensity was expected to be. This was balanced with the time required to measure the 55 deployed systems in a controlled environment.

To verify the accuracy of the light readings, eight randomly selected sensor systems were tested at three light levels (0.5, 5, and 50 klux) and three temperatures (25, 30, and 35 °C), a total of nine conditions each. These testing conditions were selected to match the conditions that sensors experienced during the field testing and are representative of the error in light readings for the collected data. Figure 4c shows the resulting measurements after calibration was applied. The x-axis is the reported light level, and the y-axis is the actual light level the sensor was exposed to. The worst-case variation in reported light measurement was sigma/mean = 28%, at 5 klux, showing acceptable stability.

Nonlinearity was worse with a sensor light reading to actual controlled light intensity ratio ranging from −37 to +14%. However, because this is a comparative study of prey and predator species, and the same individual sensors were reused for both the prey and the predators, nonlinearity was judged to be less important than sensor-to-sensor variation and variation resulting from temperature change.

We manufactured 201 smart solar sensor systems, most of which were used for bench top testing and green house testing at the University of Michigan using locally caught specimens of Cepaea nemoralis land snails (Supplementary Fig. 8). A total of 55 tested units were taken to Tahiti and were reused in multiple deployments while there. Our small batch production cost for these sensors was ~$500 US per unit (including wafer fabrication, wafer dicing, system assembly, encapsulation, and yield loss); however, for large volume (>200 units) production, this was reduced to ~$150/unit.

Field methods

Two field populations of E. rosea and three of Partula hyalina located in five northern valleys of Tahiti-Nui, the main Tahitian peninsula, were investigated in August 2017 (Fig. 1a). These locations were selected by T. Coote, who had conducted extensive field surveys on Tahiti since 2004, as being the most accessible populations of both species then available.

Although E. rosea remains widely distributed throughout Tahiti, it has become less numerous in many valleys in recent years, possibly because of the introduction of another snail predator, the New Guinea flatworm Platydemus manokwari^12,35. Dead E. rosea shells were much more common than live specimens at our three Partula hyalina study locations, so we focused instead on the robust predator populations present in the nearby main Fautaua Valley and in its side-valley Fautaua-Iti. In both locations, we picked sites where foraging E. rosea had ready access to both shaded and open habitats. The Fautaua-Iti Valley location consisted of an open sunlit trail through the rainforest (Supplementary Fig. 1d), and the solar ecologies of nine predators were monitored here on two days: 5 on August 8 and 4 on August 11. The Fautaua Valley location consisted of a forest-edge adjoining an open grassy area (Supplementary Fig. 1e), and 29 predators were monitored here over two days: 12 on August 12 and 16 on August 14.

All three of our Partula hyalina study sites (Fig. 1a) consisted of discrete patches of vegetation between the edge of the forest and the primary stream, or captage, within each valley. The Tahitian valley of Tipaerui encompasses a small side valley, Tipaerui-Iti, which contained the most robust known surviving population of P. hyalina on Tahiti, consisting of hundreds of individuals. They were restricted to a linear stand of Etlingera cevuga extending for 60–70 m (Supplementary Fig. 1a). The solar ecologies of 28 aestivating Tipaerui-Iti Partula hyalina individuals were recorded over two days: 12 on August 10 and 16 on August 15. Partula hyalina population sizes were much smaller in the other two valleys, Faarapa, and Matatia (Fig. 1a), requiring us to monitor all of the individuals we encountered. The Faarapa Valley site consisted of a mixed stand of Barringtonia asiatica, Alocasia macrorrhiza, and Pisona umbellifera (Supplementary Fig. 1b). We detected six individuals at this site, and their solar environments were monitored on August 5. Our remaining Partula hyalina study site was in Matatia Valley (Fig. 1a), where a small, low-density population occurred in scrubby habitat attached to the foliage of Z. officinale, Pisona umbellifera, and Inocarpus fagifer (Supplementary Fig. 1c). A total of seven individuals were detected and assayed on August 7.

Each working day, we entered the study valley in the early morning between 8 and 9 a.m., prior to the appearance of the sun above the valley walls; and searched systematically for our respective target species. Euglandina rosea individuals were found foraging actively, either on the ground or climbing on vegetation, and they typically maintained this searching activity throughout the day. In contrast, Partula hyalina individuals were aestivating attached to the underside of leaves, and specimens typically remained in situ on the same leaf during the observation period.

To track the solar ecology of each predator, a smart solar sensor was reversibly attached to the dorsal surface of each E. rosea shell using a nut and screw method. The nut (McMaster-Carr, Brass Hex Nut, narrow, 0–80 thread size) was glued (Loctite, Super Glue) directly on the predator’s shell, and after allowing 10 min for bonding, a sensor, preglued to a compatible screw (McMaster-Carr, 18–8 Stainless Steel Socket Head Screw 0–80 thread size, 1/16” long), was attached mechanically. Each predator was numerically labeled using nail polish and released at the exact spot it had been discovered. For the rest of the study period, each predator was visually tracked as it continued its foraging until mid-afternoon, when the sun descended below the valley walls, and the snails and sensors were recovered. These invasive predators were then euthanized.

Aestivating Partula hyalina attach to the underside of leaves. Because our permit did not allow the direct attachment of light sensors to this endangered species, we deployed under-leaf sensors next to the aestivating snails using a nut/screw/magnets combination. This involved gluing, in advance, the screw to the sensor base and the nut to a round magnet (Radial Magnet Inc., Magnet Neodymium Iron Boron (NdFeB) N35, 4.78 mm diameter, 1.60 mm thickness). In the field, these components were assembled and held in place using another magnet positioned on the upper leaf surface. In addition to recording the under-leaf light intensities experienced by the aestivating Partula hyalina specimens, we also recorded the ambient light intensity by attaching a sensor to the upper surface of the leaves harboring the aestivating specimens.

Each working day, the data recording function of the smart sensors was activated before going into the field and was terminated after returning from the field, and the data were then retrieved via the sensors’ wireless communication link. For each sensor, the recording start time, meaningful time of the measurement start time, meaningful measurement end time, and sensor recording end time were recorded to properly calibrate the time of the recorded samples. The received raw data in digital format were then translated to time and light intensity information using a MATLAB program and the calibration data specific for that sensor.

Statistics and reproducibility

Recordings from each of the three categories (Partula hyalina leaf top, P. hyalina under leaf, and Euglandina rosea) over the 8 days of field recording were aggregated into their respective 10-min time intervals from 9:30 to 16:00 h. This recording time window avoided the early morning handling period when sensors were attached to the predator, spanned the midday period of peak solar irradiation (Figs. 2, 3), and enabled us to recover the visually tracked predators before losing them in the gathering darkness of the late afternoon valley forests. We collected light intensity measurements for 40 leaf top sensors, 41 under leaf P. hyalina, and 37 foraging E. rosea snails over the 9:30–16:00 h time period. Most aestivating P. hyalina (N = 26/41) had two under-leaf sensors bracketing the snails to record their immediate light environment (Fig. 1b) and for these individuals we used the mean light intensity of the two sensors to compare to the other two categories.

We compared the three categories (leaf top, P. hyalina under leaf, and E. rosea) for the 40 timepoints over the 9:30–16:00 h time period using a repeated measures analysis of variance (ANOVA) in the nlme⁴⁵ and car⁴⁶ packages in R v.3.5.0⁴⁷. We first tested the light intensity measurements for conformance to a normal distribution using the R code shapiro.test, with the result being a highly skewed distribution. We thus LOG transformed the measurement data. We specified the following linear mixed model for the 9:30–16:00 time interval using the nlme package in R:

Where LOGfullmean = the LOG transformed light intensity readings, group = leaf top, P. hyalina under leaf, or E. rosea, time = the 40 10-min time intervals from the 9:30–16:00 time period. We considered each individual as a random block and included the correlation between time and individual. The repeated measures ANOVA utilized the restricted loglikelihood (REML) method and excluded any missing timepoint measurements (na.action = na.exclude) from the 9:30–16:00 h time period. After running the linear mixed model in R, we then used the Anova command from the R package car followed by a post-hoc Tukey’s test to determine which categories significantly differed in their light ecologies.

Reporting summary

Further information on research design is available in the Nature Research Reporting Summary linked to this article.

Source: Ecology - nature.com