betaFeedback

WorksPhysics and Astronomy

Temperature Flips the Sign of the Hydrogen Response in Nanogranular Palladium Networks

Abstract

In nanogranular palladium networks, hydrogen uptake has two opposing effects on electrical resistance. Absorbed hydrogen scatters electrons within the particles, raising their resistance, while particle expansion narrows the gaps between them and lowers the junction resistance. Such networks are widely studied as resistive hydrogen sensors, and the sign of their net response is usually treated as a fixed property of the architecture. In networks grown by atomic layer deposition on macroporous silicon, we show that changing the operating temperature can reverse this sign at a fixed hydrogen concentration. A thermodynamic model indicates that scattering contributes less as temperature rises, allowing sub-angstroem changes in gap width to dominate. Where the two effects cancel, the network absorbs hydrogen yet gives the same reading as in hydrogen-free gas, so a leak could escape detection. Between 25 and 40 degrees Celsius, this cancellation point is predicted to rise over tenfold, crossing reported leak-alarm setpoints. At 55 degrees Celsius, the measured response stays negative but passes through a minimum, so a concentration above a setpoint can mimic one below it. Annealing at 200 degrees Celsius alters the calibration and turns the room-temperature response negative at the lowest concentration tested. Quantitative sensing therefore requires a defined thermal pretreatment and an operating temperature giving distinct signals across the intended concentration range. Response polarity becomes an operating parameter rather than a fixed property. More broadly, any nanoparticle network whose particles deform on sorption may show the same temperature-dependent competition between transport within and between particles, with implications for sensors and switches beyond palladium-hydrogen.

Hydrogen sensors are needed wherever hydrogen is produced, stored, or converted, from leak detection in distribution infrastructure to process monitoring in chemical reactors. [1, 2, 3] Resistive hydrogen sensors translate gas concentration into a change in electrical resistance. Because this change also depends on operating temperature and device history, reliable operation requires a calibration that accounts for both. In leak detection, the sensor must respond unambiguously well below the lower flammability limit of hydrogen in air, which is approximately 4 vol % at room temperature. [4] Reported alarm setpoints for area monitors near indoor hydrogen dispensers are 0.4 and 1.0 vol % H₂, corresponding to 10 and 25 % of that limit. [5] In chemical processing, self-adapting reactors that adjust their operation to fluctuating renewable feedstocks depend on such local measurements. [6, 7]

Palladium is widely used for electrical hydrogen detection near room temperature. [3, 8, 9, 10] Molecular H2H_{2} dissociates on the Pd surface, and atomic hydrogen occupies interstitial sites to form PdHx\mathrm{PdH}_{x}. Dissolved hydrogen scatters conduction electrons and increases the resistivity of Pd. [11, 12] Hydrogen uptake also increases the lattice parameter by approximately 6 % per unit H/Pd. [11] In a continuous Pd film, the increase in resistivity generally dominates the electrical response. In a granular film, lattice expansion can bring neighboring particles closer together, close critical gaps, and open conductive pathways, thereby lowering the network resistance. [13, 14, 15, 16] In nanoparticle arrays and films near the percolation threshold, both mechanisms operate, and the measured response reflects their balance. [17, 18, 19, 20, 21] Reported Pd architectures are accordingly classified by the sign of their response (Table 2). For networks near the percolation threshold, this sign is generally attributed to morphology.

ArchitectureDominant mechanismResponse signRefs.
Continuous films and thick wireslattice resistivitypositive[3, 50]
Mesowire arrays with break junctionsgap closurenegative[13]
Ultrathin discontinuous filmsgap closurenegative[14, 15]
Nanogaps on elastomersgap closurenegative[16]
Nanoparticle arrays and near-percolation filmsbothmorphology dependent[17, 18, 19]
ALD network on macroporous Si (this work)bothswitched by TT and device history—

Table 2. Polarity of the resistive hydrogen response of Pd architectures near room temperature. A positive response denotes an increase in resistance during H2H_2 exposure.

The balance between these contributions is sensitive to changes in particle dimensions. A particle 30 nm in diameter expands by only 0.2 Å at a hydrogen loading of x=0.01x = 0.01, yet the tunneling conductance between neighboring particles depends exponentially on the gap width. [22] Sub-ångström expansion can therefore switch a critical contact, whereas the same loading changes the lattice resistivity by only approximately one percent. [12, 23] Hydrogen storage in nanoparticles also differs from that in bulk Pd. [24] Surface and subsurface sites can bind hydrogen more strongly than bulk interstitial sites. [9, 25] In the dilute α\alpha phase of nanocrystalline Pd, hydrogen also segregates to grain boundaries. [26, 27] Particle size and mechanical constraints narrow the miscibility gap between the dilute α\alpha phase and the hydride β\beta phase. [28, 29, 30, 31] These size and interface effects also change how hydrogen uptake varies with temperature. Whether changing the operating temperature can shift the balance between electron scattering and junction conduction far enough to reverse the response of the same network remains an open question.

Such a temperature-dependent polarity would undermine any calibration that assumes a fixed sign. At the null point, the concentration at which the two contributions cancel, hydrogen produces no net resistance change, and the sensor gives the same reading as in hydrogen-free gas. If temperature moves the null point toward an alarm setpoint, a leak near that concentration can go unreported. Where the calibration curve is nonmonotonic, one resistance value corresponds to two concentrations. Both ambiguities can be avoided only if the null point and its dependence on temperature and device history are known.

Here, we use atomic layer deposition (ALD) to grow Pd nanoparticle networks on macroporous silicon. Pd initially nucleates as islands that coalesce as the number of ALD cycles increases. [32, 33, 34, 35] The cycle count therefore controls the network’s proximity to the percolation threshold. Macroporous silicon provides a rigid three-dimensional scaffold with micrometer-scale pores, open gas access, and a large internal surface area. [36, 37, 38] Pd ALD coats both the external surface and the pore walls (Figure 1 and SF Figure 1). The availability of this scaffold as a free-standing membrane with through-pores also offers a route to flow-through sensors with Pd-coated pore walls. We measure the response of a network annealed at 200 °C to 0.05–5 vol % H2H_{2} in N2N_{2} at operating temperatures of 25, 40, and 55 °C. A thermodynamically informed two-channel model relates hydrogen uptake to electron scattering within the particles and improved conduction across their junctions. Fitted jointly to 300 exposure cycles, the model predicts how the null point shifts with operating temperature. To assess the effect of thermal history, we compare calibration curves before and after this anneal at each operating temperature and examine the room-temperature polarity pattern in an independently treated device. The measured curves identify the concentration ranges over which a resistance change gives an unambiguous reading. Because the null point shifts predictably with temperature, response polarity becomes an operating parameter. The operating temperature could then be chosen to place the null point at a target hydrogen concentration, such as an alarm setpoint.

Pd nanoparticle network grown on macroporous silicon

Figure 1. Pd nanoparticle network grown on macroporous silicon by 1100 ALD cycles. (a) Photograph of the 15 mm specimen. (b,c) Top-view SEM images at increasing magnification and (d) an image of the uncoated top surface. (e–g) Cross-sectional SEM images of the coated pore walls and (h) an image of an uncoated pore wall. Pink and green frames indicate successive magnifications. Orange outlines identify representative Pd features with diameters of 40 to 50 nm.

Results and Discussion

ALD-grown nanoparticle networks on macroporous silicon

Macroporous silicon with 6 μ\mum pores arranged on a 12 μ\mum square lattice and a pore depth of 215 μ\mum served as the scaffold. Palladium was deposited by thermal ALD from Pd(hfac)2\mathrm{Pd(hfac)}_{2} and ozone at 200 °C (Methods). Scanning electron microscopy of an 1100-cycle coating shows granular Pd on the planar top surface, around the pore openings, and on the pore walls, while the macropores remain open (Figure 1). Individual Pd features measure 40 to 50 nm in diameter, and an analysis of 83 features yields a mean size of 42.4±10.342.4 \pm10.3 nm (Figure SF Figure 2a,Figure 2c). Energy-dispersive X-ray maps of an as-deposited 580-cycle device confirm the presence of Pd in both top view and cross section (Figure SF Figure 1). The sensing devices studied below received 580 cycles. The treated device, used for the temperature study, was imaged after thermal treatment. The 118 Pd features analyzed have a mean projected-area-equivalent diameter of 29.8±5.729.8 \pm5.7 nm and a median of 29.0 nm (Figure SF Figure 2b,Figure 2d). This distribution provides the morphological input for the model.

Figure 2. Positive hydrogen response of the reference device (580 ALD cycles, as-deposited) at room temperature. (a) Resistance during five 180 s H2_2 exposures and 180 s N2_2 recovery periods at each concentration between 0.05 and 5 vol %. Blue and red indicate the gas phases. The transients in (a) were smoothed, which removes the short spikes visible in the raw data (Figure SF 3). We used the unsmoothed data to determine RlateR_{\mathrm{late}}, all response values, and t90t_{90}. The inset shows the cycles at 0.3 vol %. (b,c) Response and recovery times t90t_{90}, defined as the first crossing of 90 % of the signed change toward the late level (Methods). (d) Late response 100(R−R0)/R0100(R-R_0)/R_0, where R0R_0 is the local pre-exposure baseline. Error bars indicate cycle standard deviations (n=5n = 5). Blue, green, and red lines show linear fits over 0.05 to 1, 1 to 5, and 0.05 to 5 vol %, respectively. Dashed markers in (b,d) indicate the bulk Pd–H plateau concentration of 1.62 vol % at 25 °C.

As-deposited networks respond positively

Figure 2a shows the resistance of the reference device, an as-deposited 580-cycle network, during five 180 s exposures at each concentration between 0.05 and 5 vol % H2H_2 in N2N_2, each followed by 180 s of N2N_2. The resistance increases during every exposure and recovers in N2N_2, whereas the uncoated support exhibits no measurable response. The late response, defined as the relative resistance change averaged over 140 to 160 s after the gas switch (Methods), increases from 0.33 % at 0.05 vol % to 17.85 % at 5 vol % (Figure 2d). The calibration curve is approximately linear below 1 vol % and flattens above 3 vol %. The three linear fits compare the low- and high-concentration regimes with a fit over the full range; no single straight line captures both regimes and their crossover near the bulk plateau concentration. This behavior is consistent with competing contributions to the response, as represented by the two-channel model below. The response and recovery times t90t_{90} are shortest at 5 vol %, at 10 and 16 s, respectively (Figure 2b,Figure 2c). Both exhibit a local maximum near 2 vol %. The concentration of this maximum coincides with the bulk Pd–H plateau concentration at 25 °C, which is 1.62 vol % at 1 bar [11, 39] (Supporting Information, Section SS2). Slower kinetics during α−β\alpha-\beta conversion are well documented for Pd nanowires [40] and nanoparticles. [41] The baseline also drifts upward during cycling at concentrations above 1 to 2 vol %, and the offset persists when low concentrations are revisited (Figure SF Figure 3). The persistence of this offset over the experimental period is consistent with changes in the network during repeated hydride formation and decomposition. A second as-deposited device from a separate ALD run reproduces the positive calibration curve, with responses of 0.16 % at 0.05 vol % and 16.18 % at 5 vol % (Figure SF Figure 4). Its higher baseline resistance, 200 to 350 Ω compared with 39 to 46 Ω, may reflect a thinner coating deposited in that run.

Figure 3. Temperature-dependent response of the treated device (580 ALD cycles, after annealing) at three temperatures and the joint two-channel model. (a–c) Late responses at nominal temperatures of 25, 40, and 55 °C (points with cycle standard deviations; n=5n = 5, 10, and 5, respectively). Blue, red, and green curves show the model sum, scattering channel, and connectivity channel, respectively. Dashed markers indicate the bulk Pd–H plateau concentrations of 1.62, 3.40, and 6.66 vol %; the last lies outside the plotted range. (d–f) Mean transients at 40 °C for 0.1, 2, and 5 vol % H2_2, with standard-deviation bands (n=5n = 5) and model curves. (g) First run at 40 °C, comprising five 180 s exposure and 180 s recovery cycles per concentration, shown on a concatenated timeline. Shaded bands indicate H2_2 exposure. The model was fitted jointly to all 300 cycles at the three temperatures using nine parameters.

Figure 4. Effect of annealing on the hydrogen response of the same 580-cycle Pd network. Late response after 90 days of storage (red circles) and after subsequent annealing in air for 10 min at 200 °C (blue squares), measured at (a) room temperature, (b) 40 °C, and (c) 55 °C. The inset in (a) expands the low-concentration region around zero response. Each point represents the mean of 100(R−R0)/R0(R - R_0)/R_0 over 140 to 160 s after H2H_2 onset. Error bars indicate cycle standard deviations, with n=10n = 10 for the aged state at 55 °C and the treated state at 40 °C, and n=5n = 5 otherwise. Pale-red dashed curves distinguish the two runs for the aged state at 55 °C.

Temperature shifts the null point and makes the calibration nonmonotonic

Following the 10 min anneal in air at 200 °C, we measured the network at nominal operating temperatures of 25, 40, and 55 °C, with 75, 150, and 75 hydrogen exposure cycles, respectively. Figure 3a–Figure 3c shows the

late responses. At 25 °C, the response is positive above approximately 0.1 vol % and slightly negative below this concentration. At 40 °C, the response is negative up to approximately 1 vol % and positive at higher concentrations, reaching 3.8 % at 5 vol %. At 55 °C, the response is negative throughout the measured range, reaching −4.2 % at 0.3 vol % before becoming less negative as concentration increases. The calibration curve at 55 °C is therefore nonmonotonic. At this temperature, the same late-response value can correspond to two hydrogen concentrations, one on either side of the minimum.

The transient responses reflect the same changes. At 40 °C, exposure to 0.1 vol % produces a monotonic decrease in resistance (Figure 3d). Exposure to 2 vol % produces a rapid decrease followed by a slower increase, and the resistance undershoots the baseline during recovery (Figure 3e). At 5 vol %, the positive branch dominates, but the recovery undershoot persists (Figure 3f). These traces are consistent with two contributions of opposite sign and different kinetics. The thick-coating device (1100 ALD cycles, as-deposited) exhibits the same trend between room temperature and 30 °C, with pronounced overshoots at the lowest concentrations and a strongly reduced kinetic maximum (Figure SF 6).

Two electrical channels driven by one hydrogen loading

We describe all three temperature programs using a single model that separates hydrogen uptake from electrical transduction. The model describes the response of the treated device as the sum of two opposing electrical contributions from hydrogen absorption,

Smod(t,T)=Ss(t,T)+Sp(t,T).(1)S_{\mathrm{mod}}(t,T)=S_{\mathrm{s}}(t,T)+S_{\mathrm{p}}(t,T). \tag*{(1)}

SsS_{\mathrm{s}} is the positive scattering contribution (subscript s), and SpS_{\mathrm{p}} is the negative connectivity contribution (subscript p, for percolation). The latter represents improved conductive paths as expansion reduces the separation of critical junctions. Both contributions are fractional changes relative to the zero-loading resistance Rref(T)R_{\mathrm{ref}}(T), such that Rmod/Rref=1+SmodR_{\mathrm{mod}}/R_{\mathrm{ref}}=1+S_{\mathrm{mod}}. Their equilibrium targets depend on the same inferred hydrogen loading, but the channels evolve separately toward those targets. Their sum is compared with the measured response after cycle-local normalization. The individual channel curves represent components assigned by the model.

The equilibrium loading xx of a particle with diameter DD follows a Lacher-type isotherm, [42]

12ln⁡(pH2p0)=ln⁡(xDr−xD)−EeffxDRT+E0(T)+usurf/DRT,(2)\frac{1}{2}\ln\left(\frac{p_{\mathrm{H}_2}}{p_0}\right)=\ln\left(\frac{xD}{r-xD}\right)-\frac{E_{\mathrm{eff}}xD}{RT}+\frac{E_0(T)+u_{\mathrm{surf}}/D}{RT}, \tag*{(2)}

where r=0.6r=0.6 is the maximum site occupancy, EeffE_{\mathrm{eff}} is an effective H–H interaction, E0(T)E_0(T) is an insertion energy incorporating the Pd–H entropy of −94.5 J K−1 mol−1 H2-94.5\ \mathrm{J\,K^{-1}\,mol^{-1}}\ \mathrm{H}_2, and usurf/Du_{\mathrm{surf}}/D is a surface term accounting for the size dependence of nanoparticle isotherms. [28] The loading is averaged over the measured size distribution shown in Figure SF Figure 2d. This loading drives two electrical channels. The scattering channel follows the tabulated resistivity of bulk PdHx_x, [12]

Sseq=As(T)[ρ(x)ρ(0)−1].(3)S_{\mathrm{s}}^{\mathrm{eq}}=A_{\mathrm{s}}(T)\left[\frac{\rho(x)}{\rho(0)}-1\right]. \tag*{(3)}

Here, ρ(x)/ρ(0)\rho(x)/\rho(0) is the relative resistivity of bulk PdHx_x interpolated from tabulated data, As(T)A_{\mathrm{s}}(T) scales this intrinsic response to that of the network, and SseqS_{\mathrm{s}}^{\mathrm{eq}} is the equilibrium target of the scattering channel. The connectivity channel represents a finite population of critical junctions that close as the particles expand,

Speq=−Pmax[1−exp⁡(−ηHD‾d∗x)],(4)S_{\mathrm{p}}^{\mathrm{eq}}=-P_{\mathrm{max}}\left[1-\exp\left(-\frac{\eta_{\mathrm{H}}\overline{D}}{d^*}x\right)\right], \tag*{(4)}

where ηH=0.063\eta_{\mathrm{H}}=0.063 per H/Pd is the linear expansion coefficient, D‾\overline{D} is the mean diameter, and d∗d^* is an effective gap scale. The two channels relax toward their respective targets with first-order time constants τs(T)\tau_{\mathrm{s}}(T) and τp\tau_{\mathrm{p}}. These effective time constants summarize contributions from surface exchange, transport, and network rearrangement, processes that differ even between facets and vertices of individual Pd nanocrystals. [43] The measured signal is the sum of the two channels after applying the same cycle-local baseline normalization used for the data. Temperature enters through E0(T)E_0(T), As(T)A_{\mathrm{s}}(T), and τs(T)\tau_{\mathrm{s}}(T). All other parameters are shared

across the three temperatures. Nine parameters were fitted to the 300 cycles using robust nonlinear least squares. The full model specification, fitting procedure, cycle-bootstrap confidence intervals, and holdout tests are provided in Sections SS 5 to SS 7 of the Supporting Information.

Across all 300 exposure cycles at 25, 40, and 55 °C, the joint fit gives a pointwise root-mean-square error of 0.57 percentage points (Figure 3 and Figures SF 8 to SF 11). It captures the temperature-dependent response polarity and, at 40 °C, the initial resistance drop, slower rise, and recovery undershoot (Figure 3d–Figure 3f). Table 1 lists the fitted parameters. The connectivity channel saturates at a resistance decrease of 3.7 %, with a time constant of 17.8 s at all temperatures. The scattering amplitude decreases from As=2.56A_s = 2.56 at 25 °C to 1.06 at 40 °C and 0.19 at 55 °C, while the channel’s time constant decreases from 25.8 to 5.9 s. The equilibrium null point of the total response shifts from 0.12 vol % at 25 °C to 1.36 vol % at 40 °C and beyond the measured range at 55 °C. We conclude from the joint fit that the sign reversal results from a decrease in the positive scattering contribution rather than an increase in the negative connectivity contribution.

TT (°C)E0(T)E_0(T) (kJ mol−1^{-1} H)As(T)A_s(T)τs(T)\tau_s(T) (s)Null point (vol % H2_2)
250.942.5625.80.12
401.651.0611.91.36
552.360.195.926

Table 1. Parameters of the joint two-channel fit to the treated device. Temperature-dependent quantities are listed separately for each temperature. The shared parameters are Pmax⁡=3.70P_{\max} = 3.70 %, τp=17.8\tau_p = 17.8 s, d∗=0.113d^* = 0.113 Å, and Eeff=10E_{\mathrm{eff}} = 10 kJ mol−1^{-1} H (lower bound). The null point is the modeled equilibrium concentration at which the total response changes sign. The value at 55 °C is extrapolated beyond the measured range. Cycle-bootstrap intervals are provided in Table ST 2.

At 55 °C, the model captures the nonmonotonic response, with a fitted minimum of −3.2 % near 0.8 vol %. The measured minimum occurs at 0.3 vol % and is about one percentage point deeper (−4.2 %). The shared connectivity amplitude Pmax⁡P_{\max} limits the depth of the fitted minimum. The concurrent 6 % baseline increase (Figure SF 7) suggests that the contribution of individual junctions to the network resistance may change during cycling. At 25 °C, the measured calibration curve flattens above the bulk plateau pressure, and the model overestimates the response at 5 vol % by 1.75 percentage points. Partial conversion to the β\beta phase could contribute to this flattening through saturation of the resistivity increase.1

The scattering amplitude decreases faster than the hydrogen loading

The most pronounced parameter trend is the temperature dependence of the scattering amplitude. AsA_s decreases by a factor of 13.3 between 25 and 55 °C. Expressed as an Arrhenius slope, this decrease corresponds to an apparent energy of approximately 70 kJ mol−1^{-1}. The Pd–H thermodynamics in Equation 2 already reduce the equilibrium loading at fixed H2_2 pressure by a factor of 1.6 over the same temperature interval. The resistivity of bulk PdHx_x at fixed xx is nearly independent of temperature.2 The scattering channel of this network therefore loses amplitude much faster than can be explained by the dilute lattice solution of hydrogen in Pd. This temperature dependence is characteristic of hydrogen in strongly binding sites. Chemisorbed hydrogen on Pd surfaces has a binding energy of approximately 45 kJ mol−1^{-1} H,3 while hydrogen at grain boundaries and dislocations in nanocrystalline Pd has a trapping energy of 20 to 60 kJ mol−1^{-1} H relative to lattice sites.4 In a network of 30 nm particles, a large fraction of Pd atoms is located at surfaces and internal boundaries. Hydrogen at these sites occupies the interfaces where conduction electrons are scattered as the particle size approaches the electron mean free path. We therefore attribute the scattering channel to hydrogen at boundaries, whose population decreases steeply with increasing temperature, rather than to the lattice solution alone. The time constant of this channel decreases with an apparent activation energy of 40 kJ mol−1^{-1} (Table ST Table 2), within the range reported for hydrogen absorption and desorption in Pd nanoparticles.5

Sub-ångström gap closure and suppression of the miscibility gap

The connectivity channel converts hydrogen loading into a decrease in resistance on a length scale of d∗=0.11d^{*} = 0.11 Å. A particle of 30 nm expands by 0.19 Å at x=0.01x = 0.01 and by 1.9 Å at x=0.1x = 0.1. Within the model, this channel is therefore half-saturated at a loading of only approximately 0.004 and fully saturated well within the α\alpha phase. This model behavior explains why the negative response dominates at low concentrations and why the sign reversal occurs at the low-concentration end of the calibration curve. The value of d∗d^{*} is smaller than the decay length of tunneling conductance across a vacuum gap, approximately 0.5 Å for a Pd work function of 5 eV. [22] This effective scale favors an interpretation in terms of threshold switching rather than smooth changes in tunneling conductance. In this interpretation, the first increments of expansion close a threshold population of nearly touching junctions, as in nanogap and break-junction hydrogen switches. [13, 16, 47] Within the model, saturation at Pmax⁡=3.7P_{\max} = 3.7 % quantifies the fraction of the treated device’s network resistance assigned to these junctions. The invariance of this saturation amplitude between 25 and 55 °C reflects the model’s treatment of the junction population as a structural property, with temperature affecting the connectivity channel only through hydrogen loading.

The isotherm parameter provides further insight into nanoscale effects. In the Lacher description, the bulk critical temperature of approximately 270 °C corresponds to Eeff≈30E_{\mathrm{eff}} \approx30 kJ mol−1^{-1} H. The fit places EeffE_{\mathrm{eff}} at its lower bound of 10 kJ mol−1^{-1} H, where the miscibility gap closes far below room temperature. The network thus behaves as though α\alpha–β\beta coexistence were strongly suppressed. Reduced particle size, surface stress, and elastic clamping by a rigid substrate all narrow the miscibility gap of Pd nanoparticles. [24, 28, 29, 30, 31, 48] Particles grown by ALD on a silicon scaffold are clamped on one side, converting part of the hydrogen-induced strain into stress and raising the chemical potential of hydrogen. The absence of plateaus in the calibration curves at 40 and 55 °C is consistent with this interpretation. The kinetic maximum near 2 vol % at room temperature and the residual at 25 °C and 5 vol % are consistent with possible phase conversion in part of the particle population at the highest concentrations. A single-branch isotherm remains the simplest description that captures the three temperature programs, but its effective parameters do not explicitly resolve phase conversion.

Implications for calibration and operation

The two-channel description has direct implications for the use of nanogranular Pd sensors. First, the sign of the response depends on the operating point. A treated device of this type exhibits a negative response below and a positive response above a null point that shifts from 0.1 vol % at 25 °C to above 5 vol % at 55 °C. Second, the calibration curve at 55 °C is nonmonotonic. Each late response has the minimum at 0.3 vol % occurs at two concentrations, one on either side of the minimum. A single late-response value therefore cannot identify the concentration unless the monotonic branch is known. Third, the transient response resolves this ambiguity. The two channels have different time constants, 6 to 26 s for scattering and 18 s for connectivity, so the initial slope and recovery undershoot reflect the balance between the channels (Figure 3d–Figure 3f). Calibration using the full transient response is therefore more robust against sign ambiguity than calibration using a single late value. The physical principles apply to Pd networks near percolation in general; the porous scaffold adds a rigid three-dimensional support compatible with flow-through geometries.

Table 2 compares this work with Pd architectures of known response polarity. Continuous films exhibit positive responses. Break junctions, nanogaps, and discontinuous films exhibit negative responses and are often fast, although they exhibit hysteresis. Nanoparticle arrays and near-percolation films exhibit both signs, depending on morphology. In the ALD network on macroporous silicon, temperature shifts the balance between scattering and connectivity, and the two-channel model describes the resulting polarity changes between the measured concentration range. Measurements extend down to 0.05 vol %, and the response time is 10 s at 5 vol %, making the performance competitive with that of resistive Pd sensors without catalytic coatings. [9] Faster and more sensitive Pd sensors are available, particularly plasmonic and single-nanowire devices. [49, 50, 51, 52] The contribution of the present architecture lies in its scalable three-dimensional scaffold and an interpretable response that allows calibration under changing temperatures.

Effect of annealing on the hydrogen response

To determine how thermal history affects the calibration at a fixed operating temperature, we compared the same network at nominal temperatures of 25, 40, and 55 °C before and after annealing (Figure 4). The first measurements followed 90 days of storage after deposition. The subsequent anneal in air at 200 °C for 10 min followed a protocol adapted from Pd nanowire sensors. [53]

At room temperature, the response was positive throughout the measured range before annealing. At 0.05 vol % H2H_2, it changed from +0.51 ± 0.06 % before treatment to −0.46 ± 0.12 % afterward (Figure 4a, inset). At 5 vol %, it remained positive, increasing from 17.44 to 19.19 %. A second device subjected to the same anneal reproduced this polarity pattern in two successive runs. Its responses were −0.59 ± 0.08 and −0.19 ± 0.03 % at 0.05 vol % and 18.56 and 17.78 % at 5 vol %, respectively (Figure SF 5).

At 40 °C, both calibration curves crossed zero, but the crossing shifted to a higher concentration after annealing (Figure 4b). The measured crossing lay between 0.4 and 0.5 vol % before treatment and between 1 and 2 vol % afterward. At 0.1 vol %, the response changed from −0.44 ± 0.01 to −1.78 ± 0.04 %. At 55 °C, both states responded negatively throughout the measured range. Annealing deepened the minimum, and at 0.5 vol % the response changed from −3.05 to −4.04 % (Figure 4c).

Several processes could shift the balance between scattering and junction conduction after an air anneal. Particle coarsening or relaxation of growth stress could change the gaps along conducting paths and hence the resistance change produced by particle expansion. [15] Surface oxidation during the anneal and reduction during subsequent hydrogen exposure could also change the surface state, [54] while annealing of cycling-induced defects could alter the hydriding kinetics. [46, 55] The paired calibrations establish that the effect of treatment persists at the measurement temperatures. Distinguishing these mechanisms would require structural or surface measurements before and after treatment.

Conclusions and Perspective

Changing the operating temperature can reverse the sign of the resistance response to hydrogen in a nanogranular Pd network on macroporous silicon, even at a fixed hydrogen concentration. In the two-channel model, this reversal reflects the balance between two effects of hydrogen uptake. Electron scattering raises the resistance within the particles, while particle expansion narrows the gaps between them and lowers the resistance at their junctions. Between 25 and 55 °C, the fitted scattering amplitude decreases thirteenfold, allowing sub-ångström changes in gap width to dominate the response.

For the annealed network, the model predicts that the equilibrium null point shifts from 0.12 vol % H2H_2 at 25 °C to 1.4 vol % at 40 °C, crossing the reported alarm setpoints of 0.4 and 1.0 vol %. The 200 °C anneal also altered the calibration at each operating temperature. It reversed the low-concentration response at 25 °C, shifted the null point to a higher concentration at 40 °C, and deepened the negative response minimum at 55 °C.

At the equilibrium null point, the two contributions cancel and the sensor gives the same resistance reading as in hydrogen-free gas. When the contributions evolve on different time scales, the transient response can still indicate hydrogen uptake. At 55 °C, the measured response remained negative throughout the tested range but passed through a minimum near 0.3 vol %, so two concentrations on opposite sides of it could produce the same late resistance change. Quantitative sensing from the late response therefore requires a defined thermal pretreatment and an operating temperature at which the intended concentration range lies on a monotonic branch of the calibration curve. In such a network, response polarity is an operating parameter rather than a fixed property. Tuning the operating temperature could place the null point at a chosen hydrogen concentration, such as an alarm setpoint, so that exposures below and above that concentration produce resistance changes of opposite sign.

The model infers hydrogen loading from the electrical response. Measuring equilibrium uptake and resistance together on the annealed network under H2H_2 and D2D_2 at 40 °C would provide an independent test. If isotope substitution shifts the null point mainly through its effect on uptake, cancellation should occur at different partial pressures but at similar atomic loadings (H/Pd and D/Pd). Because the calibration was measured with N2N_2 as the carrier gas, its applicability to leak detection should next be tested in humid air. For reactor monitoring, Pd-coated silicon membranes with through-pores offer a way to test how liquid-phase transport and competing adsorbates affect the calibration under flow.

The same competition may arise in other nanoparticle networks where absorption increases particle resistivity while expansion improves conduction across interparticle junctions. In such materials, temperature could likewise select which effect dominates and thereby set the sign of the response. Particle deformation, moreover, does not require absorption into the bulk. Physisorption of gases and vapors, including water, induces reversible strains in nanoporous solids that are captured by surface-stress and solvation-pressure models. These strains can even change sign along the isotherm, with contraction during capillary condensation and expansion at higher filling. [56, 57] Adsorbates also modulate the resistance of nanoporous gold by surface scattering, for example upon water adsorption. [58, 59] In ligand-capped gold nanoparticle films, by contrast, vapor-induced swelling of the matrix between the particles widens the tunneling gaps and raises the resistance. [60] How adsorption-induced strain of the particles themselves, acting on the interparticle contacts, competes with surface scattering inside them has, to our knowledge, not yet been resolved. This question is the adsorption counterpart of the competition reported here. Resolving this interplay would link the mechanics of sorption at nanoscale interfaces to charge transport in nanostructured matter. It would also turn the coupling between sorption, strain, and conduction from a source of ambiguity into a design principle for sensors, switches, and adaptive nanomaterials well beyond palladium–hydrogen.

Methods

Substrates and atomic layer deposition

Macroporous silicon (SmartMembranes GmbH) consisted of n-doped single-crystal Si(100) with a square array of pores measuring 6 μm6\ \mu\mathrm{m} in diameter and 215 μm215\ \mu\mathrm{m} in depth, with a center-to-center spacing of 12 μm12\ \mu\mathrm{m}. Pieces measuring 15×15 mm215 \times15\ \mathrm{mm}^{2} were used as received. Palladium was deposited by thermal ALD in a home-built reactor using palladium(II) hexafluoroacetylacetonate (Pd(hfac)2_2, minimum purity 95 %, Strem Chemicals) and ozone generated from pure oxygen. [35, 61] The substrate and precursor temperatures were 200 and 65 ∘C65\ ^\circ\mathrm{C}, respectively. Each ALD cycle consisted of two 30 s Pd(hfac)2_2 pulses, each followed by a 90 s purge, and one 1 s ozone pulse followed by a 60 s purge. Five devices were examined. Four devices received 580 ALD cycles: the as-deposited reference device, the second as-deposited device from a separate run, the treated device used for the aging and temperature measurements, and the replication device. The as-deposited thick-coating device received 1100 ALD cycles. The treated device was first measured approximately 90 days after deposition and was subsequently heated on a hot plate at 200 ∘C200\ ^\circ\mathrm{C} for 10 min in ambient air. [53] The replication device received the same treatment.

Electron microscopy

SEM images and EDS maps were acquired at 15 kV. Feature sizes were determined by manually outlining Pd features. A total of 83 features were evaluated for the 1100-cycle coating. For the treated device, 118 features were outlined in an image in which the 200 nm scale bar spanned 151 pixels. Each feature was assigned the diameter of a circle with the same projected area. Further details, including the EDS acquisition, are provided in Section SS 1.

Hydrogen exposure and resistance measurement

Two mass-flow controllers (MC series, Alicat Scientific) mixed a 5 vol % H2_2 mixture in N2_2 with pure N2_2 to obtain H2_2 concentrations ranging from 0.05 to 5 vol %. The gas passed through a 7.05 mL chamber at a total flow rate of 1 standard liter per minute and a pressure near 1 bar (Figure SF 12). The ideal exchange time for a well-mixed chamber, t90=ln⁡(10)V/Qt_{90} = \ln(10) V / Q, is 1 s. Four tungsten-carbide pins contacted the sample. Resistance was measured in a four-point configuration using a Keithley 6220 current source supplying alternating ±1\pm1 mA and a Keithley 2182 nanovoltmeter. The median sampling intervals were 0.61 s for the initial calibration and 1.11 s for the post-treatment data. A PT100 located at a sample corner measured the temperature and provided feedback for a Kapton heater. A thermal-conductivity hydrogen sensor (Infineon TCIXTMA1) and a pressure sensor (Bosch BMP581) recorded the reference gas and pressure signals, respectively, for phase alignment. Each measurement cycle consisted of 180 s of H2_2 exposure followed by 180 s of recovery in N2_2. The measurement program began with five conditioning cycles at 5 vol %, followed by an ascending concentration series with five cycles per concentration. The uncoated support showed no measurable response. The median PT100 temperatures for the programs designated room temperature (or 25 ∘C25\ ^\circ\mathrm{C}), 40 ∘C40\ ^\circ\mathrm{C}, and 55 ∘C55\ ^\circ\mathrm{C} were 23.7, 40.0, and 55.0 °C, respectively. The thick-coating device was compared at 24.0 and 30.0 °C.

Response definitions

Responses were referenced to the local N2_2 baseline, R0R_0, defined as the median resistance during the interval from 40 to 10 s before H2_2 onset. The late response is 100(Rlate−R0)/R0100(R_{\mathrm{late}} - R_0)/R_0, where RlateR_{\mathrm{late}} is the mean resistance from 140 to 160 s after onset. The response time, t90t_{90}, is the first time at which the signed resistance change from R0R_0 reaches 90 % of the change from R0R_0 to RlateR_{\mathrm{late}}. The recovery time is defined analogously, using the resistance at switch-off as the starting value and the mean resistance from 140 to 160 s after the switch to N2_2 as the endpoint. Gas-switch times were identified from the first sustained change in the reference sensor signal. For recordings without telemetry and for the aged room-temperature and 40 °C recordings with lagging reference-sensor transitions, gas-switch times were instead identified from the largest absolute derivative of the smoothed resistance trace near the programmed switch (Section SS 8). Experimental uncertainties are cycle standard deviations calculated from five cycles per concentration in each run, or ten cycles when two runs are pooled, as specified in the captions. Uncertainties in fitted parameters are reported separately as conditional 95 % cycle-bootstrap intervals.

Global fit

The two-channel model was fitted jointly to all 300 post-treatment cycles of the treated device over the full 0.05 to 5 vol % H2H_2 range at nominal temperatures of 25, 40, and 55 °C. Hydrogen loading follows Equation (2), using r=0.6r = 0.6, ηH=0.063\eta_{\mathrm{H}} = 0.063, usurf=10u_{\mathrm{surf}} = 10 kJ nm mol−1^{-1} H, and the measured size distribution of the treated device represented on a grid of 15 diameters. The two channels follow Equations (3) and (4), with first-order kinetics. The temperature dependences of As(T)A_{\mathrm{s}}(T) and τs(T)\tau_{\mathrm{s}}(T) were parameterized using bounded Arrhenius-type relations with a reference temperature of 40 °C. The same cycle-local baseline normalization was applied to measured and simulated cycles. The simulated states were propagated continuously throughout the measurement program, including the conditioning cycles. The objective function used a soft-L1L_1 loss on group means, with repeatability scaling and balanced weights for temperature, concentration, and gas phase, and was minimized by constrained nonlinear least squares. Confidence intervals were obtained from 200 cycle-bootstrap refits. The complete model specification, parameter bounds, residual diagnostics, and holdout tests are provided in Sections SS 5 to SS 7.

Acknowledgments

This work was funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) – SFB 1615 – 503850735 (SMART Reactors for Future Process Engineering), project A03. P. Huber also acknowledges funding from the DFG under Germany’s Excellence Strategy – EXC 3120/1 – 533771286 (BlueMat, Water-Driven Materials). The authors thank the Centre for Molecular Water Science (CMWS), Hamburg, Germany, for scientific exchange and support.

Conflicts of Interest

The materials described in this research are the subject of European patent application No. 26210112.4. Niclas Alkazaz, Robert Zierold, and Patrick Huber are the inventors. Hamburg University of Technology (TUHH) will own any patent granted from this application.

Supporting information

Electron microscopy and feature-size analysis; bulk thermodynamic reference markers; extended room-temperature recordings; device comparisons and telemetry; full model specification; complete comparisons of model predictions and measurements at each temperature; residual diagnostics; cycle-bootstrap intervals and temperature holdout tests; and measurement setup and data processing (PDF).

Data and code availability

Measurement data and analysis code that support the findings of this study are available from the corresponding author upon reasonable request.

Electron microscopy

EDS maps of the second as-deposited device were acquired at 15 kV, with spectrum live times of 1730 s for the top view and 3991 s for the cross section (Figure SF Figure 1). The paired electron images and elemental maps show the same fields, whereas the calibrated companion SEM images show separate fields. Map intensity represents qualitative X-ray counts rather than a corrected Pd concentration.

The feature-size analysis in Figure SF Figure 2 included 83 selected features of the 1100-cycle coating and 118 manually outlined features of the treated device after thermal treatment. In the post-treatment image, the 200 nm scale bar spanned 151 pixels. For each feature, a circular diameter equivalent to its projected area was calculated from the outline midline. The midline area was estimated from the inner and outer areas to reduce sensitivity to marker width. The features of the treated device have a mean diameter of 29.8±5.729.8 \pm5.7 nm, a median of 29.0 nm, an interquartile range of 25.5 to 33.2 nm, and a range of 19.3 to 48.2 nm. These images characterize the sampled regions; they do not provide a quantitative coverage profile over the full pore depth.

Calibrated SEM

(a) Calibrated SEM

EDS SEM field

(b) EDS SEM field

Pd Lα counts

(c) Pd Lα counts

Map-sum spectrum

(d) Map-sum spectrum

Calibrated SEM

(e) Calibrated SEM

EDS SEM field

(f) EDS SEM field

Pd Lα counts

(g) Pd Lα counts

Map-sum spectrum

(h) Map-sum spectrum

Figure SF 1. EDS characterization of an as-deposited 580-cycle Pd ALD device in top view (a–d) and cross section (e–h). (a,e) Calibrated companion SEM images with 5 and 10 μm\mu\mathrm{m} scale bars, respectively. (b,f) Electron images of the EDS fields. (c,g) Pd Lα\alpha count maps. (d,h) Map-sum spectra on logarithmic axes, expressed in counts per second per 10 eV channel, with Si and Pd line positions indicated. The companion SEM images show fields different from the EDS fields. Data were acquired at 15 kV, with live times of 1730 s and 3991 s for the top view and cross section, respectively. Map brightness represents qualitative X-ray counts.

Figure SF 2. Feature-size analysis of the Pd networks. (a,c) The 1100-cycle coating before thermal treatment: 83 features with a mean diameter of 42.4 ±\pm 10.3 nm and a median of 40.9 nm. (b,d) The treated device after treatment in air at 200 ∘C^\circ\mathrm{C} for 10 min: 118 manually outlined features with a mean projected-area-equivalent diameter of 29.8 ±\pm 5.7 nm and a median of 29.0 nm. Histograms have 5 nm bins. Solid red and dashed blue lines indicate the mean and median, respectively. The distribution in (d) provides the morphological input for the two-channel model.

Bulk thermodynamic reference markers

The bulk Pd–H plateau markers in Figure 2 and Figure 3 of the main text are calculated from

ln⁡(pplp∘)=ΔH−TΔSRT,(5)\ln\left(\frac{p_{\mathrm{pl}}}{p^\circ}\right)=\frac{\Delta H-T\Delta S}{RT}, \tag*{(5)}

using p∘=1p^\circ= 1 bar, the bulk absorption enthalpy ΔH=−38.4\Delta H = -38.4 kJ mol−1^{-1} H2H_2[11, 39] and the entropy ΔS=−94.5\Delta S = -94.5 J mol−1^{-1} K−1^{-1} H2H_2, which is also used in the model. At a total pressure of 1 bar, the plateau concentrations are 1.62, 3.40, and 6.66 vol % at 25, 40, and 55 °C, respectively. The value at 25 °C is a reference at the nominal temperature; the measured room-temperature median was 23.7 °C. These markers provide bulk thermodynamic context and represent neither fitted transition pressures nor measured phase boundaries of the network. Particle size, hysteresis, and mechanical constraint shift absorption and desorption relative to bulk Pd.[28, 29, 31]

Initial calibration, repeatability, and kinetics

The reference device provides the positive-response reference shown in Figure 2 of the main text. Its mean late responses at 0.2 and 3 vol % H2H_2 were 1.70 and 15.14 %, respectively. The calibration curve flattens above approximately 3 vol %. Linear fits give R2=0.985R^2 = 0.985 over 0.05 to 1 vol %, R2=0.913R^2 = 0.913 over 1 to 5 vol %, and R2=0.963R^2 = 0.963 over the full range. Similar concentration-dependent calibration regimes occur in closely spaced Pd nanoparticle arrays.[18] Baseline drift was 0.27 % over 0.1 to 1 vol % and 0.61 % over 1 to 5 vol %. A concentration-dependent drift component becomes apparent near 2 vol %. Low-concentration responses recur at the start of a second, partially recorded ascending sequence while the N2N_2 baseline remains elevated (Figure SF Figure 3). The baseline offset therefore persists into subsequent low-concentration cycling.

The second as-deposited device shows the same positive response trend after an initial run affected by drift (Figure SF Figure 4). Under the common late-response definition, the two devices give 0.33 ± 0.07 and 0.16 ± 0.08 % at 0.05 vol %, and 17.85 ± 0.01 and 16.18 ± 0.02 % at 5 vol %. The resistance of the second as-deposited device is higher: 200 to 350 Ω, compared with 39 to 46 Ω for the reference device. This difference may reflect reduced Pd deposition caused by temperature gradients during ALD, when heating was applied from below without an upper cover. The reference device responds more slowly at low concentrations than at high concentrations. An additional local maximum in response time occurs near 2 vol %. Response and recovery are fastest at 5 vol %.

Figure SF 3. Extended room-temperature raw resistance recording of the reference device (580 ALD cycles, as-deposited). Figure 2 of the main text shows the smoothed initial calibration, in which the short spikes are removed; RlateR_{\mathrm{late}}, all response values, and t90t_{90} times were calculated from unsmoothed data. The first complete ascending program spans 0.05 to 5 vol % H2\mathrm{H}_{2} in N2\mathrm{N}_{2} and is followed by the start of a second program. At each concentration, five 180 s exposures alternate with 180 s N2\mathrm{N}_{2} recovery periods. Gray bands mark H2\mathrm{H}_{2} exposure, and concentration labels are in vol %. Asterisks mark the five initial conditioning exposures at 5 vol % in each program. Low-concentration responses recur in the second program while the N2\mathrm{N}_{2} baseline remains elevated.

Figure SF 4. Room-temperature comparison of the reference device and second as-deposited device, each prepared with 580 ALD cycles in separate runs using the same recipe. The upper panel shows the raw resistance of the second as-deposited device during two ascending concentration programs; the first run exhibits downward drift. The lower panel shows the late responses of the reference device (blue, first run) and second as-deposited device (coral, second run). Each cycle comprises 180 s of H2_2 exposure and 180 s of N2_2 recovery. Points show means, and bars show cycle standard deviations (n=5n = 5). The response is 100(R−R0)/R0100(R - R_0)/R_0 averaged from 140 to 160 s after an onset determined from the resistance trace; R0R_0 is the median resistance from 40 to 10 s before onset.

Temperature measurements and controls

Room-temperature comparison of two treated devices

The treated device and replication device each received a single treatment for 10 min at 200 °C in air and were measured near room temperature using the same program. The treated device provided one complete run, and the replication device provided two complete runs. The recordings for both devices had a median PT100 temperature of 23.70 °C. This comparison was evaluated separately from the model-fitting data set.

Gas switches were aligned to the first sustained change in the reference-sensor signal. The baseline window extended from 40 to 10 s before onset, and the late-response window extended from 140 to 160 s after onset; both were evaluated on a 1 s grid. Invalid startup telemetry was excluded before clock alignment.

At 5 vol%, the late response was 19.20 ± 0.04% for the treated device and 18.56 ± 0.09 and 17.78 ± 0.04% for the two runs of the replication device. At 0.05 vol%, the corresponding responses were −0.45 ± 0.12, −0.59 ± 0.08, and −0.19 ± 0.03%, respectively. At 0.1 vol%, the replication device gave −0.12 ± 0.06 and +0.30 ± 0.02%, placing this concentration at its null point. The low- and high-concentration polarity pattern therefore recurs across the two treated devices, although the response magnitude varies between devices and successive runs. Local N₂ baselines were approximately 39 Ω for the treated device and 41 to 43 Ω for the replication device (Figure SF 5).

Figure SF 5. Room-temperature responses of the treated device (blue) and replication device (two runs, coral and teal), both prepared with 580 ALD cycles and treated independently. The recordings for both devices had a median PT100 temperature of 23.70 °C. (a,b) Mean late responses over the full and low-concentration ranges, respectively. (c) Local N₂ baselines. (d–f) Mean normalized transients at 0.05, 1, and 5 vol% H₂. Bars and bands show cycle standard deviations (n=5n = 5 per run). Exposure and recovery are aligned separately to the reference-gas transitions. Recovery begins at a nominal time of 180 s. Pale bands mark exposure.

Thick-coating device at room temperature and 30 °C

The thick-coating device, prepared with 1100 ALD cycles, was measured at median PT100 temperatures of 24.0 and 30.0 °C (Figure SF 6). At 30 °C, the exposure transients overshoot at 0.05 and 0.1 vol %, and the recovery transient undershoots after hydrogen removal at the lowest concentration. The room-temperature transients are slower and exhibit less overshoot. Heating reduces the exposure response and shortens both the response and recovery times. The pronounced room-temperature kinetic maximum near 2 vol % is substantially reduced at 30 °C. This trend is consistent with a contribution from phase conversion to the slower room-temperature response,6 together with faster hydrogen exchange and a shift in the channel balance.

Figure SF 6. Temperature-dependent response of the thick-coating device near room temperature (blue, median PT100 temperature 24.04 °C) and at nominal 30 °C (red, median 29.99 °C). Each concentration includes five cycles, each comprising 180 s of H₂ exposure and 180 s of N₂ recovery. (a) Mean transients at 0.05, 0.1, 0.5, 1, 2, and 5 vol %, normalized to the late exposure response. Bands show cycle standard deviations, and pale shading marks exposure. (b) Mean late response. (c,d) Mean response and recovery t90t_{90}, respectively. Bars in (b–d) show cycle standard deviations (n=5n = 5). Recovery is displayed from a nominal time of 180 s. The phase-transition and percolation-path annotations are schematic.

Post-treatment recording at 55 °C

Figure SF 7 shows the reference-gas, temperature, and resistance traces recorded simultaneously during the post-treatment program at 55 °C. The local pre-exposure N₂ baseline rises from 43.52 to 46.24 Ω\Omega over the ascending concentration program. From 100 to 168 s after onset, the cycle-averaged exposure response changes by less than 0.08 percentage points at every concentration. Over the corresponding recovery window, the cycle-averaged response continues to rise by 0.10 to 0.16 percentage points. Exposure and recovery thus evolve differently. The recovery behavior is consistent with continued relaxation of the device state; its possible contribution to the model mismatch at 55 °C is discussed in Section SS 6.

Figure SF 7. Concentration program for the treated device at nominal 55 °C after treatment in air at 200 °C for 10 min. (a) Reference H₂ signal. (b) PT100 temperature. (c) Raw resistance. All panels share an elapsed-time axis. Programmed concentrations range from 0.05 to 5 vol %, with five 180 s exposures alternating with 180 s N₂ recovery periods at each concentration. Gray bands mark exposure. The resistance decreases during every exposure.

Model specification and global fitting

Two electrical response channels

The channel sum, reference resistance, and meanings of the scattering and connectivity contributions are defined in (1) and the accompanying text of the main manuscript. The individual channel curves are model assignments, and their sum is compared with the measured response after the cycle-local normalization specified below.

Hydrogen loading

For a particle of diameter DD, the equilibrium hydrogen loading xDx_D, expressed as H/Pd, follows a Lacher-type isotherm, [42]

12ln⁡(pH2p0)=ln⁡(xDr−xD)−EeffxDRT+E0(T,D)RT.(6)\frac{1}{2}\ln\left(\frac{p_{\mathrm{H}_2}}{p_0}\right)=\ln\left(\frac{x_D}{r-x_D}\right)-\frac{E_{\mathrm{eff}}x_D}{RT}+\frac{E_0(T,D)}{RT}. \tag*{(6)}

pH2=Ptotc/100p_{\mathrm{H}_2}=P_{\mathrm{tot}}c/100, where Ptot=1 barP_{\mathrm{tot}}=1\ \mathrm{bar} and cc is the programmed concentration in vol%. The reference pressure is p0=1 barp_0=1\ \mathrm{bar}, rr is the maximum site occupancy, and EeffE_{\mathrm{eff}} is the effective interaction between absorbed hydrogen atoms. Temperatures are expressed in kelvin, and energies are given per mole of H. The insertion energy comprises a size-independent term and a surface term,

E0(T)=E0,ref−(T−Tref)ΔSf2,E0(T,D)=E0(T)+usurfD,(7)E_0(T)=E_{0,\mathrm{ref}}-\frac{(T-T_{\mathrm{ref}})\Delta S_{\mathrm{f}}}{2}, \qquad E_0(T,D)=E_0(T)+\frac{u_{\mathrm{surf}}}{D}, \tag*{(7)}

where Tref=313.15 KT_{\mathrm{ref}}=313.15\ \mathrm{K} and ΔSf=−94.5 J K−1 mol−1 H2\Delta S_{\mathrm{f}}=-94.5\ \mathrm{J\,K^{-1}\,mol^{-1}\ H_2}. The factor of two converts the entropy to a per-mole-of-H basis. The ensemble loading is the number-weighted mean

x=∑iwixiDi∑iwi(8)x=\frac{\sum_i w_i x_i D_i}{\sum_i w_i} \tag*{(8)}

over 15 equally spaced diameters spanning D‾±2.5σD\overline{D}\pm2.5\sigma_D, subject to a lower limit of 12 nm, with Gaussian weights derived from the size distribution of the treated device. The electrical target laws are applied to this average loading. The imposed temperature dependence follows Pd–H thermodynamics. Particle size, surface contributions, and elastic coupling shift nanoparticle equilibria relative to bulk Pd. [28] Chemomechanical calculations also relate mechanical constraint to the miscibility gap. [31]

Scattering channel

Sseq(x,T)=As(T)[ρ(x)ρ(0)−1].(9)S_{\mathrm{s}}^{\mathrm{eq}}(x,T)=A_{\mathrm{s}}(T)\left[\frac{\rho(x)}{\rho(0)}-1\right]. \tag*{(9)}

ρ(x)/ρ(0)\rho(x)/\rho(0) is the relative resistivity of PdHx\mathrm{PdH}_x, interpolated from tabulated data. [12] In the dilute α\alpha phase, resistivity increases approximately linearly with H/Pd\mathrm{H}/\mathrm{Pd}. [23] As(T)A_{\mathrm{s}}(T) scales the intrinsic PdHx\mathrm{PdH}_x response to that of the network. Thin-film measurements show that mechanical constraint affects solubility much more strongly than it affects resistivity at a given hydrogen content. [48] This observation supports treating hydrogen loading, x(pH2,T)x(p_{\mathrm{H}_2},T), separately from electrical transduction, ρ(x)\rho(x).

Connectivity channel

Speq(x)=−Pmax⁡[1−exp⁡(−κgapx)],κgap=10ηHD‾d∗.(10)S_{\mathrm{p}}^{\mathrm{eq}}(x)=-P_{\max}\left[1-\exp(-\kappa_{\mathrm{gap}}x)\right], \qquad\kappa_{\mathrm{gap}}=\frac{10\eta_{\mathrm{H}}\overline{D}}{d^*}. \tag*{(10)}

Pmax⁡P_{\max} is the maximum fractional decrease in resistance. ηH\eta_{\mathrm{H}} is the linear lattice-expansion coefficient per H/Pd\mathrm{H}/\mathrm{Pd}, D‾\overline{D} is the mean diameter in nanometers, and d∗d^* is an effective junction-distance scale in ångström. The factor of 10 converts nanometers to ångström. The exponential law represents a finite population of critical junctions that progressively close as hydrogen loading increases. [15, 18] The measured resistance constrains the combined sensitivity κgap\kappa_{\mathrm{gap}}, from which d∗d^* is calculated using the fixed values of ηH\eta_{\mathrm{H}} and D‾\overline{D}.

Kinetics

Each channel approaches its equilibrium target through first-order relaxation,

dSjdt=Sjeq−Sjτj(T),j∈{s,p}.(11)\frac{dS_j}{dt}=\frac{S_j^{\mathrm{eq}}-S_j}{\tau_j(T)}, \qquad j\in\{\mathrm{s},\mathrm{p}\}. \tag*{(11)}

Both equilibrium targets are zero during N2\mathrm{N}_2 recovery. Each channel has its own time constant, which is the same for absorption and desorption within that channel. These effective time constants can incorporate surface exchange, hydrogen transport, lattice relaxation, and rearrangement of the electrical network. In shape-controlled Pd nanocrystals, hydrogen uptake and release exhibit different sensitivities to vertices and facets, [43] illustrating the microscopic processes represented by this effective description.

Temperature parameterization

AsA_s is parameterized by a bounded inverse-temperature relation,

as(T)=as,ref+EAsR(1T−1Tref),As(T)=Amin⁡+Amax⁡−Amin⁡1+exp⁡[−as(T)],(12)a_s(T) = a_{s,\mathrm{ref}} + \frac{E_{A_s}}{R}\left(\frac{1}{T} - \frac{1}{T_{\mathrm{ref}}}\right), \qquad A_s(T) = A_{\min} + \frac{A_{\max} - A_{\min}}{1 + \exp[-a_s(T)]}, \tag*{(12)}

where Amin⁡=0.02A_{\min} = 0.02, Amax⁡=3A_{\max} = 3, and as,ref=ln⁡[(As(Tref)−Amin⁡)/(Amax⁡−As(Tref))]a_{s,\mathrm{ref}} = \ln[(A_s(T_{\mathrm{ref}}) - A_{\min})/(A_{\max} - A_s(T_{\mathrm{ref}}))]. The scattering time constant is parameterized as

ln⁡[τs(T)τs(Tref)]=Eτ,sR(1T−1Tref).(13)\ln\left[\frac{\tau_s(T)}{\tau_s(T_{\mathrm{ref}})}\right] = \frac{E_{\tau,s}}{R}\left(\frac{1}{T} - \frac{1}{T_{\mathrm{ref}}}\right). \tag*{(13)}

EAsE_{A_s} and Eτ,sE_{\tau,s} are the slopes of these imposed relations. The logistic bound means that EAsE_{A_s} differs from the Arrhenius energy of AsA_s itself. The direct Arrhenius slope between the fitted values at 25 and 55 °C is Rln⁡(2.558/0.192)/(1/298.15 K−1/328.15 K)=70 kJ mol−1R\ln(2.558/0.192)/(1/298.15\ \mathrm{K} - 1/328.15\ \mathrm{K}) = 70\ \mathrm{kJ\ mol^{-1}}.

Global fitting procedure

The joint fit used the post-treatment programs of the treated device at nominal temperatures of 25, 40, and 55 °C. One complete run of 75 cycles was available at each of 25 and 55 °C, and two complete runs of the same length, totaling 150 cycles, were available at 40 °C. Each run comprised five cycles at 0.05 and 0.1 vol %, at concentrations from 0.2 to 1.0 vol % in steps of 0.1 vol %, and at concentrations from 2 to 5 vol % in steps of 1 vol %. The five initial conditioning cycles at 5 vol % initialized the model states but were excluded from the objective. Gas transitions were identified from the first sustained response of the reference sensor and represented as instantaneous concentration steps. The chamber time constant was fixed to zero. Each measured cycle was normalized to its local pre-exposure baseline R0R_0. For simulated cycle cc, bj,cb_{j,c} denotes the median channel state over the pre-exposure window from 40 to 10 s before onset. The quantities compared with the measurements are

S~j,c(t)=Sj(t)−bj,c1+bs,c+bp,c,S~c(t)=S~s,c(t)+S~p,c(t).(14)\widetilde{S}_{j,c}(t) = \frac{S_j(t) - b_{j,c}}{1 + b_{s,c} + b_{p,c}}, \qquad\widetilde{S}_c(t) = \widetilde{S}_{s,c}(t) + \widetilde{S}_{p,c}(t). \tag*{(14)}

Measured responses are expressed as 100(R−R0)/R0100(R-R_0)/R_0, and simulated responses as 100S~c100\widetilde{S}_c. The simulated states evolve continuously across gas switches and cycles, retaining the effects of incomplete recovery. Comparing responses relative to each cycle’s local baseline prevents long-term baseline drift from dominating the fit.

The objective was a soft-L1L_1 loss minimized by constrained nonlinear least squares. Within each temperature, run, concentration, and gas-phase group, the five measured cycles were averaged on a common time grid and compared with the mean model prediction. Each difference was scaled by the larger of a 0.05 percentage-point noise floor and the group’s median robust cycle-to-cycle standard deviation. A factor proportional to the inverse square root of the number of runs and time points balanced the groups. Consequently, each temperature, concentration, and gas phase had comparable influence despite the unequal numbers of runs. Penalty terms enforced the parameter constraints.

The size-distribution parameters were fixed at D‾=29.78 nm\overline{D} = 29.78\ \mathrm{nm} and σD=5.69 nm\sigma_D = 5.69\ \mathrm{nm}. The remaining fixed quantities were ηH=0.063\eta_{\mathrm{H}} = 0.063, r=0.6r = 0.6, and usurf=10 kJ nm mol−1 Hu_{\mathrm{surf}} = 10\ \mathrm{kJ\ nm\ mol^{-1}\ H}. The parameter bounds were −4-4 to 8 kJ mol−1 H8\ \mathrm{kJ\ mol^{-1}\ H} for E0,refE_{0,\mathrm{ref}}, 1010 to 30 kJ mol−1 H30\ \mathrm{kJ\ mol^{-1}\ H} for EeffE_{\mathrm{eff}}, 0.020.02 to 33 for AsA_s, 00 to 0.100.10 for Pmax⁡P_{\max}, 11 to 300 (H/Pd)−1300\ (\mathrm{H/Pd})^{-1} for κgap\kappa_{\mathrm{gap}}, and 0.30.3 to 600 s600\ \mathrm{s} for the time constants. The bounds for both temperature-slope parameters, EAsE_{A_s} and Eτ,sE_{\tau,s}, were −200-200 to 200 kJ mol−1200\ \mathrm{kJ\ mol^{-1}}. Temperature dependence enters through E0(T)E_0(T), As(T)A_s(T), and τs(T)\tau_s(T). The parameters EeffE_{\mathrm{eff}}, Pmax⁡P_{\max}, κgap\kappa_{\mathrm{gap}}, and τp\tau_p are shared across temperatures. Nine parameters were fitted.

Fitted parameters

Table ST 1 lists the point estimates. The size-independent insertion energy is E0,ref=1.651 kJ mol−1 HE_{0,\mathrm{ref}} = 1.651\ \mathrm{kJ\ mol^{-1}\ H} at 40 °C, giving E0(T)=0.942E_0(T) = 0.942, 1.651, and 2.360 kJ mol−1 H\mathrm{kJ\ mol^{-1}\ H} at 25, 40, and 55 °C, respectively. At the mean diameter, the surface term usurf/Du_{\mathrm{surf}}/D adds 0.336 kJ mol−1 H\mathrm{kJ\ mol^{-1}\ H}. The scattering amplitude decreases successively from 2.558 to 1.055 and 0.192, an overall decrease by a factor of 13.3. The scattering time constant decreases successively from 25.78 to 11.94 and 5.94 s, an overall decrease by a factor of 4.3, with Eτ,s=39.82 kJ mol−1E_{\tau,s} = 39.82\ \mathrm{kJ\ mol^{-1}}.

This energy lies within the range of apparent activation energies reported for hydrogen absorption and desorption in Pd nanoparticles. [45, 46] Heated Pd nanowires likewise exhibit smaller response magnitudes and faster kinetics. [40] The shared connectivity parameters are Pmax⁡=0.03704P_{\max}=0.03704, κgap=165.50\kappa_{\mathrm{gap}}=165.50 (H/Pd)−1^{-1}, and τp=17.77\tau_p=17.77 s. With the fixed expansion coefficient and mean diameter, κgap\kappa_{\mathrm{gap}} corresponds to d∗=0.1134d^*=0.1134 Å.

TT (°C)E0(T)E_0(T) (kJ mol−1^{-1} H)As(T)A_s(T)τs(T)\tau_s(T) (s)τp\tau_p (s)Null point (vol % H2_2)
250.9422.55825.7817.770.120
401.6511.05511.9417.771.361
552.3600.1925.9417.7725.90

Table ST 1. Point estimates from the joint fit. E0(T)E_0(T) is the size-independent insertion energy; the isotherm includes the additional term usurf/Du_{\mathrm{surf}}/D for each particle diameter. The null point is the modeled equilibrium concentration at which the total response changes sign. The null point at 55 °C is extrapolated beyond the tested concentration range.

Parameter sensitivity and identifiability

Replacing the size distribution of the treated device with the 1100-cycle distribution changes the total response trajectories by at most 3.2×10−53.2 \times10^{-5} percentage points over 25 800 modeled points, while shifting the fitted insertion energy and the size-derived value of d∗d^*. The total response is therefore well constrained, whereas the assignment of morphology-dependent parameters is not. The interaction parameter EeffE_{\mathrm{eff}} reaches its lower bound of 10 kJ mol−1^{-1} H. In the Lacher model, the critical temperature is Tc=rEeff/(4R)T_c=rE_{\mathrm{eff}}/(4R), giving 180 K at this bound and 541 K for the bulk value Eeff≈30E_{\mathrm{eff}}\approx30 kJ mol−1^{-1} H. The fitted branch is consequently single-valued at all three temperatures. The model includes no explicit α\alpha–β\beta phase fraction, transformation kinetics, or hysteresis. Any room-temperature contribution from phase conversion would be reflected in the effective parameters and structured residuals. The local Jacobian condition number is 1973. The strongest correlations are r=0.967r=0.967 between E0,refE_{0,\mathrm{ref}} and EAsE_{A_s}, and r=−0.956r=-0.956 between the scattering reference time constant and its temperature slope.

Full model comparisons and residual diagnostics

Figures SF 8 to SF 10 show the complete first run at each temperature, together with the modeled equilibrium response and late-response residuals. Experimental points are averages of the raw resistance samples from 140 to 160 s. The modeled late-response curves are averages of predictions stored at 4 s intervals over the same window. Full-run plots concatenate nominal 360 s cycles, whereas the state calculations retain the recorded gas-switch chronology.

Figure SF 8. Joint two-channel model at nominal 25 ∘C25\,^\circ\mathrm{C} (median PT100 temperature 23.70 ∘C23.70\,^\circ\mathrm{C}) for the treated device. (a) First run displayed on a concatenated timeline, with five cycles per concentration, each comprising 180180 s of H2\mathrm{H}_2 exposure and 180180 s of N2\mathrm{N}_2 recovery. Gray bands mark exposure. The measured response (pale blue), model sum (blue), scattering contribution (red), and connectivity contribution (green) are each referenced to the cycle-local baseline. (b) Equilibrium channel responses relative to zero loading (lines) and measured late responses (points). (c) Mean late residuals, defined as measurement minus model, from 140140 to 160160 s on the common 44 s grid. Bars in (b,c) show cycle standard deviations (n=5n = 5).

Figure SF 9. Joint two-channel model at 40 ∘C40\,^\circ\mathrm{C} for the treated device. (a) First of two runs displayed on a concatenated timeline. Gray bands mark H2H_2 exposure. The measured response (pale blue), model sum (blue), scattering contribution (red), and connectivity contribution (green) are referenced to the cycle-local baseline. (b) Equilibrium channel responses relative to zero loading (lines) and measured late responses (points). (c) Mean late residuals, defined as measurement minus model. Points and bars in (b,c) show means and cycle standard deviations across both runs (n=10n = 10).

Figure SF 10. Joint two-channel model at 55 °C for the treated device. (a) First run displayed on a concatenated timeline. Gray bands mark H2H_2 exposure. The measured response (pale blue), model sum (blue), scattering contribution (red), and connectivity contribution (green) are referenced to the cycle-local baseline. (b) Equilibrium channel responses relative to zero loading (lines) and measured late responses (points). (c) Mean late residuals, defined as measurement minus model. Bars in (b,c) show cycle standard deviations (n=5n = 5).

Residual statistics

The temperature-balanced mean group RMSE is 0.474 percentage points. The overall pointwise RMSE and mean absolute error are 0.569 and 0.314 percentage points, respectively. The pointwise RMSEs at 25 ∘C25\,^\circ\mathrm{C}, 40 ∘C40\,^\circ\mathrm{C}, and 55 ∘C55\,^\circ\mathrm{C} are 0.828, 0.321, and 0.636 percentage points, respectively. At 25 ∘C25\,^\circ\mathrm{C} and 5 vol %, the exposure RMSE is 2.28 percentage points, with a model-minus-measurement bias of 1.75 percentage points. Figure SF 11 shows these deviations by concentration and gas phase.

At 55 ∘C55\,^\circ\mathrm{C}, the measured late response reaches a minimum of −4.22-4.22 % at 0.3 vol %, whereas the fitted minimum is −3.18-3.18 % near 0.8 vol %. At 0.5 vol %, the measured response is −4.04-4.04 %, and the fitted scattering and connectivity contributions are +0.31+0.31 and −3.45-3.45 %, respectively, giving a total of −3.14-3.14 %. The shared connectivity limit of 100Pmax⁡=3.704100P_{\max}=3.704 % restricts the depth of the modeled minimum, and cycle-local normalization does not remove this restriction. The evolution of the baseline and recovery at 55 ∘C55\,^\circ\mathrm{C} (Figure SF 7) suggests that the junction weighting may change during the run.

Figure SF 11. Concentration-resolved diagnostics of the joint fit to the treated device at nominal 25, 40, and 55 °C (top to bottom). Left panels show pointwise RMSE; right panels show signed bias, defined as model minus measurement. Red circles represent H₂ exposure, and blue squares represent N₂ recovery. Values are expressed in percentage points of 100(R−R0)/R0100(R-R_0)/R_0. Each statistic includes five cycles per concentration and run. The statistics at 40 °C are averaged across both runs.

Cycle-bootstrap uncertainty and temperature holdout tests

Complete cycles, each comprising paired exposure and recovery phases, were resampled with replacement within each temperature, run, and concentration group. Resampling preserved five cycles per group and the temporal structure within each cycle. The gas program, morphology, temperature assignments, and model structure were held fixed. Each constrained soft-L1L_1 refit began near the global optimum, used the original bounds, and allowed up to 500 function evaluations. All 200 refits converged. Table ST Table 2 reports intervals from the 2.5th to the 97.5th percentile for the nine fitted and six derived quantities. These intervals quantify cycle-to-cycle variation under within-group exchangeability. The interval for EeffE_{\mathrm{eff}} is affected by its active lower bound.

QuantityUnitsPoint estimatePercentile interval
E0,refE_{0,\mathrm{ref}}kJ mol−1^{-1} H1.651[1.639, 1.716]
EeffE_{\mathrm{eff}} (lower bound)kJ mol−1^{-1} H10.000[10.000, 10.118]
As(40 ∘C)A_{\mathrm{s}}(40\,^\circ\mathrm{C})–1.055[1.045, 1.082]
EAsE_{A_{\mathrm{s}}}kJ mol−1^{-1}123.111[121.431, 130.383]
τs(40 ∘C)\tau_{\mathrm{s}}(40\,^\circ\mathrm{C})s11.942[11.604, 12.076]
Eτ,sE_{\tau,\mathrm{s}}kJ mol−1^{-1}39.817[39.269, 40.965]
100Pmax⁡100P_{\max}%3.704[3.666, 3.719]
κgap\kappa_{\mathrm{gap}}(H/Pd)−1^{-1}165.497[159.908, 178.421]
τp\tau_{\mathrm{p}}s17.774[17.355, 17.966]
As(25 ∘C)A_{\mathrm{s}}(25\,^\circ\mathrm{C})–2.558[2.541, 2.621]
As(55 ∘C)A_{\mathrm{s}}(55\,^\circ\mathrm{C})–0.192[0.178, 0.195]
τs(25 ∘C)\tau_{\mathrm{s}}(25\,^\circ\mathrm{C})s25.776[25.302, 25.980]
τs(55 ∘C)\tau_{\mathrm{s}}(55\,^\circ\mathrm{C})s5.936[5.669, 6.056]
Null point 25 ∘^\circCvol % H2_20.120[0.115, 0.123]
Null point 40 ∘^\circCvol % H2_21.361[1.350, 1.364]

Table ST 2. Conditional 95 % cycle-bootstrap percentile intervals from 200 refits, shown alongside the point estimates from the all-temperature fit.

As a sensitivity check, one temperature program was omitted at a time. The model was refitted to the other two programs using six optimization starts, and the omitted program was predicted using its recorded gas-switch times and the same baseline convention. All three fits converged. Table ST 3 reports the pointwise RMSE. The holdout test at 40 ∘^\circC involves interpolation between the training temperatures; the other two tests involve extrapolation. The refit that omitted 40 ∘^\circC reached the 200 kJ mol−1^{-1} upper bound of EAsE_{A_{\mathrm{s}}}. The refit that omitted 25 ∘^\circC reached the 300 (H/Pd)−1^{-1} upper bound of κgap\kappa_{\mathrm{gap}}. The fits to pairs of temperatures did not predict the response at the omitted temperature. The holdout errors therefore quantify the limits of extrapolation rather than a failure of the joint description.

Held-out TT (°C)Training TT (°C)All-temperature RMSEHeld-out RMSE
2540, 550.8284.257
4025, 550.3212.662
5525, 400.6360.834

Table ST 3. Prediction errors from temperature holdout tests, expressed in percentage points. The all-temperature fit provides the descriptive benchmark.

Measurement setup and data processing

Gas mixing and electrical measurements

Figure SF 12 shows the measurement chamber. Two mass-flow controllers (MC series, Alicat Scientific) mix a premixed stream of 5 vol% H2_2 in N2_2 with pure N2_2 at a total flow rate of 1 standard liter per minute.

The chamber has a volume of 7054 mm37054\ \mathrm{mm}^3 and operates near 1 bar. Using the nominal flow rate at chamber conditions gives a residence time V/QV/Q of 0.42 s and an ideal well-mixed exchange time t90=ln⁡(10) V/Qt_{90} = \ln(10)\,V/Q of 0.98 s. The actual volumetric flow rate depends on the controller reference conditions and the chamber temperature and pressure.

Four tungsten-carbide pins contact the sample. A Keithley 6220 supplies alternating currents of +1 and −1 mA, and a Keithley 2182 nanovoltmeter measures the voltage. A Raspberry Pi Pico 2 H records the resistance, pressure, temperature, and reference H2\mathrm{H}_2 signal. The sensing element of an Infineon Hydrogen 2 Click module (TCIXTMA1) provides the reference signal, and a Bosch BMP581 monitors the chamber pressure. A four-wire Class A PT100 in contact with a sample corner measures the temperature through a MAX31865 converter. This measurement provides feedback to a Kapton heater powered at 5 V through a MOSFET. A host computer logs the data through USB. The reference and pressure signals are used only for monitoring and phase alignment; no amplitude correction is applied to the resistance.

Figure SF 12. Four-point resistance measurement during controlled hydrogen exposure. Two mass-flow controllers (MFC) mix 5 vol % H2\mathrm{H}_2 in N2\mathrm{N}_2 with pure N2\mathrm{N}_2 at a total flow rate of 1 standard liter per minute through a 7.05 mL chamber. Four tungsten-carbide (WC) pins contact the sample for voltage measurement with an alternating ±1\pm1 mA current. A PT100 measures the temperature at a sample corner, and a thermal-conductivity hydrogen sensor (TCI) provides a reference gas signal. A computer records all signals.

Gas-phase alignment

The reference device and the second as-deposited device were measured without synchronized reference-gas or temperature telemetry. For these recordings, the programmed switches were aligned to the largest absolute local derivative of an 11-point-smoothed resistance trace within a bounded window around each expected switch. The detected H2\mathrm{H}_2 exposure intervals lasted 160 to 205 s. The first run of the second as-deposited device was excluded because strong downward drift overlapped the ascending concentration program.

In the paired aging comparison, the reference-sensor transitions lagged the resistance changes in the aged room-temperature and 40 ∘C40\,^\circ\mathrm{C} recordings. These recordings were aligned to the largest absolute derivative of a seven-point-smoothed resistance trace near each programmed switch, giving exposure intervals of 167 to 188 s. All other recordings used the reference-sensor timing. Each complete run began with five conditioning cycles at 5 vol %, followed by five cycles at each of 15 concentrations. The aged series included one complete run at each of room temperature and 40 ∘C40\,^\circ\mathrm{C}, and two runs at 55 ∘C55\,^\circ\mathrm{C}. The treated series included one run at each of nominal 25 and 55 ∘C55\,^\circ\mathrm{C}, and two runs at 40 ∘C40\,^\circ\mathrm{C}. The 40 ∘C40\,^\circ\mathrm{C} recording of the thick-coating device was excluded because a resistance discontinuity occurred from approximately 3 vol % onward.

Response definitions

R0R_0 is the median resistance from 40 to 10 s before H2H_2 onset. The late response is the mean of 100(R−R0)/R0100(R-R_0)/R_0 from 140 to 160 s after onset. This finite-time definition applies to both positive and negative transients. The response time t90t_{90} is the first crossing of 90 % of the signed change from R0R_0 to the late level. The recovery time is the first crossing of 90 % of the signed change from the resistance at switch-off to the mean resistance from 140 to 160 s after the switch to N2N_2. Each search is restricted to the corresponding gas phase. Because these definitions use the first crossing, the reported times can include early crossings during overshoot. Programmed concentrations and gas-phase times are used without an independent correction for gas delivery.

References

  1. [1]Møller, K. T.; Jensen, T. R.; Akiba, E.; Li, H.-W. Hydrogen - A Sustainable Energy Carrier. Prog. Nat. Sci.:Mater. Int. 2017, 27, 34–40.DOI
  2. [2]Qanbar, M. W.; Hong, Z. A Review of Hydrogen Leak Detection Regulations and Technologies. Energies 2024, 17, 4059.DOI
  3. [3]Hüber t, T.; Boon-Brett, L.; Black, G.; Banach, U. Hydrogen Sensors – A Review. Sens. Actuators, B 2011, 157, 329–352.
  4. [4]Coward, H. F.; Jones, G. W. Limits of Flammability of Gases and Vapors; Bulletin 503, 1952.
  5. [5]Buttner, W.; Burgess, R.; Post, M.; Rivkin, C. Summary and Findings from the NREL/DOE Hydrogen Sensor Workshop (June 8, 2011); Technical Report NREL/TP-5600-55645, 2012.DOI
  6. [6]Hassenstein, P.; Krueger, J.-D. H.; Paschalidis, L.; Voß, D.; Skiborowski, M.; Albert, J. Glycerol Hydrogenolysis Goes SMART – Separating Reaction Steps as a Key for Tailoring Selectivity and Reaction Control. Ind. Eng. Chem. Res. 2026, 65, 11298–11309.DOI
  7. [7]Santra, S. Koopman–PCE-Based Confidence-Bounded Fault Diagnosis and Recovery for Hydrogen Demand Tracking in a Membrane Reactor. Comput. Chem. Eng. 2026, 213, 109728.DOI
  8. [8]Koo, W.-T.; Cho, H.-J.; Kim, D.-H.; Kim, Y. H.; Shin, H.; Penner, R. M.; Kim, I-D. Chemiresistive Hydrogen Sensors: Fundamentals, Recent Advances, and Challenges. ACS Nano 2020, 14, 14284–14322.DOI
  9. [9]Darmadi, I.; Nugroho, F. A. A.; Langhammer, C. High-Performance Nanostructured Palladium-Based Hydrogen Sensors – Current Limitations and Strategies for Their Mitigation. ACS Sensors 2020, 5, 3306–3327.DOI
  10. [10]Luo, S.-X. L.; Yuan, W.; Xue, M.; Feng, H.; Bezdek, M. J.; Palacios, T.; Swager, T. M. Chemiresistive Hydrogen Sensing with Size-Limited Palladium Nanoparticles in Iptycene-Containing Poly(arylene ether)s. ACS Nano 2023, 17, 2679–2688.DOI
  11. [11]Flanagan, T. B.; Oates, W. A. The Palladium-Hydrogen System. Annu. Rev. Mater. Sci. 1991, 21, 269–304.DOI
  12. [12]Otterson, D. A.; Smith, R. J. Absorption of Hydrogen by Palladium and Electrical Resistivity up to Hydrogen–Palladium Atom Ratios of 0.97; NASA Technical Note NASA TN D-5441, 1969.
  13. [13]Favier, F.; Walter, E. C.; Zach, M. P.; Benter, T.; Penner, R. M. Hydrogen Sensors and Switches from Electrodeposited Palladium Mesowire Arrays. Science 2001, 293, 2227–2231.pubmed.ncbi.nlm.nih.gov/11567132
  14. [14]Xu, T.; Zach, M. P.; Xiao, Z. L.; Rosenmann, D.; Welp, U.; Kwok, W. K.; Crabtree, G. W. Self-Assembled Monolayer-Enhanced Hydrogen Sensing with Ultrathin Palladium Films. Appl. Phys. Lett. 2005, 86, 203104.DOI
  15. [15]Kiefer, T.; Villanueva, L. G.; Fargier, F.; Favier, F.; Brugger, J. The Transition in Hydrogen Sensing Behavior in Noncontinuous Palladium Films. Appl. Phys. Lett. 2010, 97, 121911.DOI
  16. [16]Lee, J.; Shim, W.; Lee, E.; Noh, J.-S.; Lee, W. Highly Mobile Palladium Thin Films on an Elastomeric Substrate: Nanogap-Based Hydrogen Gas Sensors. Angew. Chem., Int. Ed. 2011, 50, 5301–5305.
  17. [17]Ramanathan, M.; Skudlarek, G.; Wang, H. H.; Darling, S. B. Crossover Behavior in the Hydrogen Sensing Mechanism for Palladium Ultrathin Films. Nanotechnology 2010, 21, 125501.DOI
  18. [18]Xie, B.; Zhang, S.; Liu, F.; Peng, X.; Song, F.; Wang, G.; Han, M. Response Behavior of a Palladium Nanoparticle Array Based Hydrogen Sensor in Hydrogen–Nitrogen Mixture. Sens. Actuators, A 2012, 181, 20–24.DOI
  19. [19]Das, S. S.; Kopnov, G.; Gerber, A. Positive vs Negative Resistance Response to Hydrogenation in Palladium and Its Alloys. AIP Adv. 2020, 10, 065129.DOI
  20. [20]Gupta, D.; Dutta, D.; Kumar, M.; Barman, P. B.; Sarkar, C. K.; Basu, S.; Hazra, S. K. A Low Temperature Hydrogen Sensor Based on Palladium Nanoparticles. Sens. Actuators, B 2014, 196, 215–222.DOI
  21. [21]Kim, D.-H.; Kim, S.-J.; Shin, H.; Koo, W.-T.; Jang, J.-S.; Kang, J.-Y.; Jeong, Y. J.; Kim, I.-D. High-Resolution, Fast, and Shape-Conformable Hydrogen Sensor Platform: Polymer Nanofiber Yarn Coupled with Nanograined Pd@Pt. ACS Nano 2019, 13, 6071–6082.DOI
  22. [22]Simmons, J. G. Generalized Formula for the Electric Tunnel Effect between Similar Electrodes Separated by a Thin Insulating Film. J. Appl. Phys. 1963, 34, 1793–1803.DOI
  23. [23]Tóth, J.; Péter, L.; Bakonyi, I.; Tompa, K. Peculiarities of the Electrolytic Hydrogenation of Pd as Revealed by Resistivity Measurements. J. Alloys Compd. 2005, 387, 172–178.
  24. [24]Yamauchi, M.; Ikeda, R.; Kitagawa, H.; Takata, M. Nanosize Effects on Hydrogen Storage in Palladium. J. Phys. Chem. C 2008, 112, 3294–3299.DOI
  25. [25]Pundt, A.; Kirchheim, R. Hydrogen in Metals: Microstructural Aspects. Annu. Rev. Mater. Res. 2006, 36, 555–608.DOI
  26. [26]Kirchheim, R. Hydrogen Solubility and Diffusivity in Defective and Amorphous Metals. Prog. Mater. Sci. 1988, 32, 261–325.DOI
  27. [27]Lemier, C.; Weissmüller, J. Grain boundary segregation, stress and stretch: Effects on hydrogen absorption in nanocrystalline palladium. Acta Mater. 2007, 55, 1241–1254.DOI
  28. [28]Griessen, R.; Strohfeldt, N.; Giessen, H. Thermodynamics of the Hybrid Interaction of Hydrogen with Palladium Nanoparticles. Nat. Mater. 2016, 15, 311–317.DOI
  29. [29]Syrenova, S.; Wadell, C.; Nugroho, F. A. A.; Gschneidtner, T. A.; Diaz Fernandez, Y. A.; Nalin, G.; Świtlik, D.; Westerlund, F.; Antosiewicz, T. J.; Zhdanov, V. P.; Moth-Poulsen, K.; Langhammer, C. Hydride Formation Thermodynamics and Hysteresis in Individual Pd Nanocrystals with Different Size and Shape. Nat. Mater. 2015, 14, 1236–1244.DOI
  30. [30]Baldi, A.; Narayan, T. C.; Koh, A. L.; Dionne, J. A. In Situ Detection of Hydrogen-Induced Phase Transitions in Individual Palladium Nanocrystals. Nat. Mater. 2014, 13, 1143–1148.DOI
  31. [31]Dyck, A.; Böhlke, T.; Pundt, A.; Wagner, S. Phase Transformation in the Palladium Hydrogen System: Effects of Boundary Conditions on Phase Stabilities. Scr. Mater. 2024, 247, 116117.DOI
  32. [32]Elam, J. W.; Zinovev, A.; Han, C. Y.; Wang, H. H.; Welp, U.; Hryn, J. N.; Pellin, M. J. Atomic Layer Deposition of Palladium Films on Al₂O₃ Surfaces. Thin Solid Films 2006, 515, 1664–1673.DOI
  33. [33]Weber, M. J.; Mackus, A. J. M.; Verheijen, M. A.; Longo, V.; Bol, A. A.; Kessels, W. M. M. Atomic Layer Deposition of High-Purity Palladium Films from Pd(hfac)₂ and H₂ and O₂ Plasmas. J. Phys. Chem. C 2014, 118, 8702–8711.DOI
  34. [34]Lausecker, C.; Muñoz-Rojas, D.; Weber, M. Atomic Layer Deposition (ALD) of Palladium: From Processes to Applications. Crit. Rev. Solid State Mater. Sci. 2024, 49, 908–930.DOI
  35. [35]Zou, Y.; Li, J.; Cheng, C.; Wang, Z.; Ong, A. J.; Goei, R.; Li, X.; Li, S.; Tok, A. I. Y. Atomic Layer Deposition of Palladium Thin Film from Palladium (II) Hexafluoroacetylacetonate and Ozone Reactant. Thin Solid Films 2021, 738, 138955.DOI
  36. [36]Lehmann, V.; Föll, H. Formation Mechanism and Properties of Electrochemically Etched Trenches in n-Type Silicon. J. Electrochem. Soc. 1990, 137, 653–659.DOI
  37. [37]Scheen, G.; Bassu, M.; Douchamps, A.; Zhang, C.; Debliquy, M.; Francis, L. A. Palladium Nanoparticle Deposition via Precipitation: A New Method to Functionalize Macroporous Silicon. Sci. Technol. Adv. Mater. 2014, 15, 065002.pubmed.ncbi.nlm.nih.gov/27877732
  38. [38]Brinker, M.; Huber, P. A Mott–Schottky Analysis of Mesoporous Silicon in Aqueous Electrolyte Solution by Electrochemical Impedance Spectroscopy. Electrochim. Acta 2024, 483, 144038.
  39. [39]Flanagan, T. B.; Luo, W.; Clewley, J. D. Calorimetric Enthalpies of Absorption and Desorption of Protium and Deuterium by Palladium. J. Less-Common Met. 1991, 172–174, 42–55.DOI
  40. [40]Yang, F.; Taggart, D. K.; Penner, R. M. Joule Heating a Palladium Nanowire Sensor for Accelerated Response and Recovery to Hydrogen Gas. Small 2010, 6, 1422–1429.DOI
  41. [41]Langhammer, C.; Zhdanov, V. P.; Zorić, I.; Kasemo, B. Size-Dependent Kinetics of Hydriding and Dehydriding of Pd Nanoparticles. Phys. Rev. Lett. 2010, 104, 135502.pubmed.ncbi.nlm.nih.gov/20481892
  42. [42]Lacher, J. R. A Theoretical Formula for the Solubility of Hydrogen in Palladium. Proc. R. Soc. London, Ser. A 1937, 161, 525–545.DOI
  43. [43]Johnson, N. J. J.; Lam, B.; MacLeod, B. P.; Sherbo, R. S.; Moreno-Gonzalez, M.; Fork, D. K.; Berlinguette, C. P. Facets and Vertices Regulate Hydrogen Uptake and Release in Palladium Nanocrystals. Nat. Mater. 2019, 18, 454–458.DOI
  44. [45]Stolås, A.; Darmadi, I.; Nugroho, F. A. A.; Moth-Poulsen, K.; Langhammer, C. Impact of Surfactants and Stabilizers on Palladium Nanoparticle–Hydrogen Interaction Kinetics: Implications for Hydrogen Sensors. ACS Appl. Nano Mater. 2020, 3, 2647–2653.DOI
  45. [46]Alekseeva, S.; Strach, M.; Nilsson, S.; Fritzsche, J.; Zhdanov, V. P.; Langhammer, C. Grain-Growth Mediated Hydrogen Sorption Kinetics and Compensation Effect in Single Pd Nanoparticles. Nat. Commun. 2021, 12, 5427.DOI
  46. [47]Kiefer, T.; Favier, F.; Vazquez-Mena, O.; Villanueva, G.; Brugger, J. A Single Nanotrench in a Palladium Microwire for Hydrogen Detection. Nanotechnology 2008, 19, 125502.DOI
  47. [48]Wagner, S.; Pundt, A. Electrical Resistivity and Hydrogen Solubility of PdHₓ Thin Films. Acta Mater. 2010, 58, 1387–1394.DOI
  48. [49]Nugroho, F. A. A.; Darmadi, I.; Cusinato, L.; Susarrey-Arce, A.; Schreuders, H.; Bannenberg, L. J.; da Silva Fanta, A. B.; Kadkhodazadeh, S.; Wagner, J. B.; Antosiewicz, T. J.; Hellman, A.; Zhdanov, V. P.; Dam, B.; Langhammer, C. Metal–Polymer Hybrid Nanomaterials for Plasmonic Ultrafast Hydrogen Detection. Nat. Mater. 2019, 18, 489–495.DOI
  49. [50]Yang, F.; Kung, S.-C.; Cheng, M.; Hemminger, J. C.; Penner, R. M. Smaller Is Faster and More Sensitive: The Effect of Wire Size on the Detection of Hydrogen by Single Palladium Nanowires. ACS Nano 2010, 4, 5233–5244.DOI
  50. [51]Offermans, P.; Tong, H. D.; van Rijn, C. J. M.; Merken, P.; Brongersma, S. H.; Crego-Calama, M. Ultralow-Power Hydrogen Sensing with Single Palladium Nanowires. Appl. Phys. Lett. 2009, 94, 223110.DOI
  51. [52]Jo, M.-S.; Kim, K.-H.; Lee, J.-S.; Kim, S.-H.; Yoo, J.-Y.; Choi, K.-W.; Kim, B.-J.; Kwon, D.-S.; Yoo, I.; Yang, J.-S.; Chung, M.-K.; Park, S.-Y.; Seo, M.-H.; Yoon, J.-B. Ultrafast (~0.6 s), Robust, and Highly Linear Hydrogen Detection up to 10% Using Fully Suspended Pure Pd Nanowire. ACS Nano 2023, 17, 23649–23658.DOI
  52. [53]Kim, K.-H.; Jo, M.-S.; Kim, S.-H.; Kim, B.; Kang, J.; Yoon, J.-B.; Seo, M.-H. Long-Term Reliable Wireless H₂ Gas Sensor via Repeatable Thermal Refreshing of Palladium Nanowire. Nat. Commun. 2024, 15, 8761.DOI
  53. [54]Zemlyanov, D.; Aszalos-Kiss, B.; Kleimenov, E.; Teschner, D.; Zafeiratos, S.; Hävecker, M.; Knop-Gericke, A.; Schlögl, R.; Gabasch, H.; Unterberger, W.; Hayek, K.; Klötzer, B. In Situ XPS Study of Pd(111) Oxidation. Part 1: 2D Oxide Formation in 10⁻³ mbar O₂. Surf. Sci. 2006, 600, 983–994.
  54. [55]Delmelle, R.; Amin-Ahmadi, B.; Sinnaeve, M.; Idrissi, H.; Pardoen, T.; Schryvers, D.; Proost, J. Effect of Structural Defects on the Hydriding Kinetics of Nanocrystalline Pd Thin Films. Int. J. Hydrogen Energy 2015, 40, 7335–7347.DOI
  55. [56]Gor, G. Y.; Huber, P.; Bernstein, N. Adsorption-Induced Deformation of Nanoporous Materials—A Review. Appl. Phys. Rev. 2017, 4, 011303.DOI
  56. [57]Detsi, E.; Chen, Z. G.; Vellinga, W. P.; Onck, P. R.; De Hosson, J. T. M. Reversible Strain by Physisorption in Nanoporous Gold. Appl. Phys. Lett. 2011, 99, 083104.DOI
  57. [58]Wahl, P.; Traußnig, T.; Landgraf, S.; Jin, H.-J.; Weissmüller, J.; Würschum, R. Adsorption-Driven Tuning of the Electrical Resistance of Nanoporous Gold. J. Appl. Phys. 2010, 108, 073706.DOI
  58. [59]Hakamada, M.; Kato, N.; Miyazawa, N.; Mabuchi, M. Water-Adsorption Effect on Electrical Resistivity of Nanoporous Gold. Scr. Mater. 2016, 123, 30–33.DOI
  59. [60]Olichwer, N.; Meyer, A.; Yesilmen, M.; Vossmeyer, T. Gold Nanoparticle Superlattices: Correlating Chemiresistive Responses with Analyte Sorption and Swelling. J. Mater. Chem. C 2016, 4, 8214–8225.DOI
  60. [61]Cheng, C.; Zou, Y.; Li, J.; Ong, A. J.; Goei, R.; Huang, J.; Li, S.; Tok, A. I. Y. Adsorption and Reaction Mechanisms of Direct Palladium Synthesis by ALD Using Pd(hfac)₂ and Ozone on Si (100) Surface. Processes 2021, 9, 2246.DOI

Paper details

vlnc:W_ny9cy9fm
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents