In-Situ Hydrogen Sensing: The Palladium-Nickel Thin-Film Advantage

The gas that moves first

Every dissolved gas analysis (DGA) programme is built around a hierarchy of urgency, and hydrogen sits at the top of it. Hydrogen is produced at lower temperatures than any other fault gas, which makes it the earliest chemical evidence that something has begun inside a transformer. Partial discharge generates it. Arcing generates it. Thermal decomposition of oil generates it. It appears — in different proportions — in essentially every fault type, which is exactly why the ratio methods lean on it: the classic partial-discharge signature is a CH₄/H₂ ratio below 0.1, and a C₂H₂/H₂ ratio above 3 is read as a tap-changer leak path into the main tank rather than a fault in the tank itself.

The thresholds reflect that primacy. IEC 60599 puts the typical hydrogen concentration at around 150 µL/L, and IEEE C57.104-2019 grades hydrogen across four conditions from below 100 µL/L to above 1,800 µL/L. In practice it is hydrogen’s rate of change that most often triggers the first escalation — long before acetylene arrives to confirm the fault has become serious.

That makes the hydrogen channel the single most consequential measurement in an online monitor. It is also the one measurement that a photoacoustic instrument cannot make.

Two gases the optical bench cannot see

Photoacoustic spectroscopy (PAS) detects a gas by the light it absorbs: a molecule takes up a photon at one of its vibrational transitions, the energy becomes heat, the heat becomes a pressure wave, and a microphone hears it. No absorption, no signal. And absorption of infrared light requires a changing dipole moment during the vibration.

Hydrogen is a homonuclear diatomic — two identical atoms sharing a symmetric bond. Its vibration produces no dipole change, so its fundamental vibrational transitions are not infrared-active. The same reasoning applies to nitrogen. And it applies, identically, to oxygen. This is not a limitation of one vendor’s optical design; it is a selection rule, and it binds every infrared absorption technique — photoacoustic, non-dispersive infrared (NDIR) and Fourier-transform infrared (FTIR) alike. (We examine the hydrogen case in detail in why photoacoustic DGA cannot measure hydrogen.)

The practical consequence is visible across the whole market: every photoacoustic DGA monitor obtains hydrogen, oxygen and nitrogen from sensors that are not the optical path. An ABB patent describing an FTIR-based analyser (US 10,832,854 B2) states the point plainly, specifying that O₂, H₂ and N₂ “may be performed optically and/or with non-optical sensors” — resistive, capacitive or thermo-conductive — and naming a paramagnetic analyser for oxygen. A nine-gas headline is, structurally, an optical path plus a set of auxiliary channels. The question worth asking is not how many gases are listed. It is how each one is actually measured.

Oxygen is not a fault gas — but it decides how you read the others

Oxygen carries no fault signature of its own. IEEE C57.104-2019 groups it with nitrogen as a gas entering through residual air, air ingress or a nitrogen preservation system — not a decomposition product. Yet it earns its place on the measurement list for three reasons.

• It identifies the preservation regime. IEEE C57.104-2019 uses an O₂/N₂ ratio of 0.2 to separate low-oxygen units (hermetic, membrane-sealed or nitrogen-blanketed) from air-breathing ones — though a 2025 analysis of 225,238 oil samples put the practical split nearer 0.05 and questioned the 0.2 threshold (Draper & Dukarm, CIGRE Science & Engineering No. 36, 2025). This is not housekeeping: the 2019 edition conditions its typical-value tables on age and on that ratio, and the limits for low-oxygen units are higher by a factor of roughly five for low-temperature hydrocarbons. A unit that flips category between samples changes its own screening criteria.
• It flags excessive oxidation. IEC 60599 reads an O₂/N₂ ratio below 0.3 as evidence that oxygen is being consumed faster than it is replenished, pointing to oil oxidation or accelerated paper ageing. Air-saturated oil sits near 0.5.
• It has a wide dynamic range that carries information. A 2025 fleet study of 225,238 oil samples from 11,529 transformers found free-breathing units sitting near 20,000 µL/L, sealed units at 4,000–10,000, and nitrogen-blanketed units below 4,000 (Draper & Dukarm, CIGRE Science & Engineering No. 36, 2025). Where oxygen lands within that spread says something about the unit — and about the preservation regime it belongs to.

Oxygen is therefore a contextual gas, not a threshold gas — and, like hydrogen, it is invisible to the optical path.

Extract, or measure where the gas already lives?

There are two ways to get a dissolved gas number. The first is to take it out: draw an oil sample (or draw oil into a monitor), separate the dissolved gas from the oil by vacuum, stripping or headspace extraction, and analyse the gas phase. This is what laboratories do, what gas chromatographs do, and what PAS monitors do. The second is to leave the gas where it is and put a sensor directly into the oil.

The first approach is general — it hands you every gas at once, in one matrix, and the analyser does not care what it is looking at. The second is selective, and it only makes sense for gases whose sensing mechanism works at the oil interface. Hydrogen and oxygen are precisely those gases. And for hydrogen, the extraction route carries a specific and quantifiable tax.

Why extraction loses hydrogen first

When gas partitions between oil and a void, how much stays dissolved is governed by the Ostwald solubility coefficient. Hydrogen has the lowest coefficient of every diagnostic gas. IEC 60567 Annex A (Table A.1, mineral oil) gives the coefficients below; the retention columns are computed from them for a 95 mL oil / 5 mL void system:

Read the first row against the last. Hydrogen loses roughly four times the fraction that ethane loses. And the loss is not merely a loss — it is a distortion, because it is gas-selective. A 5 mL void in a 100 mL sample container shifts the CH₄/H₂ ratio from 4.00 to 6.94, a change of +73 %, and pulls a true 70 µL/L hydrogen reading down to 36.0 µL/L. Since CH₄/H₂ is one of the ratio-method partial-discharge criteria, a sampling artefact can move a case across a diagnostic boundary.

Two further details matter. First, the coefficient is not a fixed constant: IEC 60567 notes that solubility coefficients depend on both temperature and oil density, and warns that the Annex A values are means taken from oils in use in 2006 which newer oils may not match — so a recovery figure is only meaningful at a stated temperature and oil type. Second, the mechanisms differ: vacuum and Toepler extraction recover hydrogen extremely well (the standard notes these methods remove about 97 % of the more soluble gases and higher percentages of the less soluble ones, so hydrogen is the best-recovered gas in that route). The hydrogen problem is not vacuum extraction. It is sample-container headspace, static headspace extraction, and field sampling — which is why the sampling standards set a container-integrity criterion of hydrogen loss below 2.5 % per week. That criterion sat in IEC 60567 clauses 4.2–4.3 of the 2005 and 2011 editions; oil sampling has since moved to IEC 60475:2022.

A sensor that never leaves the oil has no container, no headspace, no transport and no extraction step. For hydrogen specifically, that removes the largest term in the error chain rather than shrinking a small one.

Inside the palladium-nickel film

Palladium has an unusual relationship with hydrogen, and it is the basis of the sensor. When hydrogen reaches a palladium surface:

• Dissociative adsorption — the H₂ molecule splits into atomic hydrogen at the metal surface, with a sticking probability close to unity.
• Absorption — individual hydrogen atoms enter the face-centred-cubic lattice and occupy the octahedral interstitial sites, the largest available voids.
• Palladium hydride formation — the absorbed hydrogen forms PdH_x and acts as an electron-scattering centre, raising the film’s electrical resistivity.
• Reversibility — when the hydrogen concentration falls, the process runs backwards and the film returns towards its baseline resistance.

Two numbers convey the scale of the effect. Hydrogen occupies the octahedral interstitial sites of the palladium lattice up to a limiting composition near PdH_0.7 — roughly 70 % site occupancy (Manchester, San-Martin & Pitre, Journal of Phase Equilibria 15(1):62–83, 1994). And in the dilute phase the dissolved hydrogen concentration follows Sieverts’ law, C_H ∝ √p(H₂) — so the resistance response tracks roughly the square root of the hydrogen partial pressure, with reported exponents around 0.4–0.5, rather than rising linearly. That is worth stating correctly, because a great deal of sensor marketing implies a straight line.

Selectivity is the other half of the story. Hydrogen is small enough, and dissociates readily enough, to enter the lattice; methane, ethylene, acetylene, carbon monoxide and carbon dioxide are not. The film is selective by physical mechanism rather than by filtering — which is a different and more durable proposition than a sensor that responds to whatever changes its surface chemistry.

The problem with pure palladium — and what nickel fixes

Pure palladium has a defect that becomes fatal in a sensor that measures over years. As dissolved hydrogen increases, the metal does not simply absorb it smoothly. It undergoes a phase transition: from the dilute α phase (below roughly 0.017 H/Pd at room temperature) into the hydrogen-rich β phase (above roughly 0.58 H/Pd), with a two-phase region in between. The β phase has a lattice constant near 4.02 Å against 3.89 Å for the α phase — a linear expansion of about 3.3–3.5 %, or roughly 10–11 % by volume (phase boundaries and lattice constants: Manchester, San-Martin & Pitre 1994; the volumetric figure is derived from the lattice constants).

Three failures follow. The pressure at which the transition occurs is not the same on the way up as on the way down, which produces hysteresis: the reading depends on the history of the hydrogen pressure, not only its present value. Across the two-phase plateau a large change in absorbed hydrogen occurs for almost no change in pressure, so the sensor loses concentration resolution in exactly that band. And the repeated swelling and contracting generates stress that cracks and eventually delaminates the film.

Alloying palladium with nickel addresses the root cause rather than the symptom. Nickel has a smaller lattice constant than palladium (3.52 Å against 3.89 Å), so adding it contracts the lattice, reduces the interstitial volume available, and shifts hydride formation to higher pressures. The Sandia National Laboratories work on Pd/Ni films is the reference here. Hughes and Schubert reported in 1992 that nickel addition suppresses the α-to-β transition of pure palladium across alloys of 8–20 atomic percent nickel (Journal of Applied Physics 71(1):542–544); the 1995 follow-up with Buss found that alloys above 8 at% Ni showed no phase change up to 630 Torr of hydrogen at ambient temperature (Journal of the Electrochemical Society 142(1):249–254). And Noh, Flanagan and co-workers found that at 15 % nickel, hysteresis disappeared completely after eight absorption–desorption cycles (Scripta Metallurgica et Materialia 25(9):2177–2180, 1991).

The trade-off is real and worth naming: adding nickel reduces hydrogen sensitivity, and very high nickel fractions introduce additional hydride phases of their own. Practical sensor alloys sit in a band — Sandia tested 8–20 at% Ni; other work has found an optimum near 8 at% Ni.

Why in-situ sensing changes the error budget

The argument for measuring hydrogen in place is not that it makes a better sensor. It is that it removes steps. Sampling, transport, container headspace and oil–gas separation are the dominant terms in the hydrogen error chain, and an immersed sensor performs none of them.

The gain in measurement frequency is the other half, and it is quantifiable. CIGRE’s TF15 analysis framed the trade-off directly (TB 409, Table 31): to detect gassing within 15 %, online monitoring needs roughly six measurements per day at 2 % reproducibility where laboratory analysis needs one per day at 10 %. A 100 ppm/year gassing rate rising from a 10 ppm baseline is detectable in about four days online against 24 days by laboratory. For a gas whose value lies in its rate of change, that difference is the whole point.

What in-situ sensing does not do is automatically deliver better accuracy, and it is worth being straight about that. CIGRE’s own multi-vendor programme (TB 409, 2010) measured participating laboratories at roughly ±12 % while individual online monitors ranged from about ±9 % to over ±80 %, with several exceeding the ±15 % target. Long-term field drift data for in-situ hydrogen sensors remains thin — the most recent published stability figures are laboratory-scale, and manufacturers’ ten-year calibration-free claims are, at present, not independently audited. In-situ sensing also does not remove the need for calibration: the sensor still reports against gas-in-oil standards and still requires a conversion from partial pressure to ppm in oil.

What can be said without qualification is structural: it eliminates the steps that lose hydrogen, and it delivers enough measurements to see a trend form while it is still a trend.

Engineering the in-situ hydrogen channel

The DGA-500 is built on this reasoning. Its sensing element is a palladium-nickel alloy thin film operating at ambient temperature with no oil–gas separation stage — the probe sits in the oil, mounted at the transformer’s bottom valve or a bushing’s oil sampling port, and reads dissolved hydrogen directly. Manufacturer specifications put the range at 5–5,000 ppm with 1 ppm resolution, accuracy of ±15 % of reading or ±5 ppm (whichever is greater), and a T63 response under 30 minutes above 100 ppm. The unit is IP67 and communicates over Modbus RTU / RS-485, with relay outputs for hydrogen level, rate of change and oil temperature, and a stated ten-year design life with no consumable.

Two design details matter for the long-term argument. Because the alloy film is selective by lattice physics rather than surface chemistry, the channel does not need oxygen to function — distinguishing it from metal-oxide and catalytic principles, which are built on oxygen chemisorption and behave poorly in the oxygen-depleted environment of transformer oil (mechanisms reviewed in Hübert et al., Sensors and Actuators B 157(2):329–352, 2011). And because the sensor sits in oil rather than in a gas stream, field verification is a direct comparison: draw a sample near the probe, run it per ASTM D3612-02(2026), and recalibrate if the deviation exceeds 15 %.

The wider architecture follows the same logic. The DGA-900 pairs a laser photoacoustic optical path for the hydrocarbon gases and carbon oxides with a dedicated hydrogen channel, because the optical path cannot see hydrogen and no honest instrument pretends otherwise. Where hydrogen and moisture are the entire requirement — distribution transformers, bushing monitoring, fleet-wide screening — the DGA-500 stands alone, and the DGA-300 offers the same sensing technology as an OEM probe for third-party monitors and digital-twin platforms. Read the palladium-nickel thin-film deep-dive for the sensor architecture, or contact PAS DGA to discuss what fits your fleet.

Sources

• IEC 60567:2023 (Ed. 5.0, 2023-12-08), Oil-filled electrical equipment — Sampling of free gases and analysis of free and dissolved gases in mineral oils and other insulating liquids — Guidance — Annex A, Table A.1 (Ostwald solubility coefficients, mineral oil); clause 4.2–4.3 container integrity in the 2005/2011 editions.
• IEC 60475:2022 — sampling of insulating liquids; carries the syringe-integrity criterion (clause 4.2, Annex B) that oil sampling moved to when IEC 60567:2023 narrowed its scope.
• IEEE C57.104-2019, IEEE Guide for the Interpretation of Gases Generated in Mineral Oil-Immersed Transformers — O₂/N₂ preservation discriminator and hydrogen condition ratings.
• IEC 60599:2022 — typical concentrations and fault-type interpretation.
• ASTM D3612-02(2026) — Standard Test Method for Analysis of Gases Dissolved in Electrical Insulating Oil by Gas Chromatography.
• Draper, Z. H. & Dukarm, J. J., “Determination of oil preservation types from atmospheric gases to contextualize DGA in power transformers,” CIGRE Science & Engineering No. 36, February 2025 — 225,238 samples from 11,529 transformers.
• CIGRE Technical Brochure 409, Report on Gas Monitors for Oil-Filled Electrical Equipment, WG D1.01/TF15, 2010 — laboratory vs online accuracy; Table 31 detection-time comparison.
• US Patent 10,832,854 B2 (ABB Schweiz AG), Dissolved gas analysis devices, systems, and methods — non-optical measurement of O₂, H₂ and N₂.
• Manchester, F. D., San-Martin, A. & Pitre, J. M., “The H-Pd (Hydrogen-Palladium) System,” Journal of Phase Equilibria 15(1):62–83, 1994 — α/β phase boundaries and lattice constants.
• Hughes, R. C. & Schubert, W. K., “Thin films of Pd/Ni alloys for detection of high hydrogen concentrations,” Journal of Applied Physics 71(1):542–544, 1992.
• Hughes, R. C., Schubert, W. K. & Buss, R. J., “Solid-State Hydrogen Sensors Using Palladium-Nickel Alloys: Effect of Alloy Composition on Sensor Response,” Journal of the Electrochemical Society 142(1):249–254, 1995.
• Noh, H., Flanagan, T. B., Gavra, Z., Johnson, J. R. & Reilly, J. J., “The disappearance of hysteresis for the hydride phase transition in palladium-nickel alloys,” Scripta Metallurgica et Materialia 25(9):2177–2180, 1991.
• Hübert, T., Boon-Brett, L., Black, G. & Banach, U., “Hydrogen sensors – A review,” Sensors and Actuators B: Chemical 157(2):329–352, 2011.
• Range, resolution, accuracy, T63, IP rating, communications and design life for the DGA-500 are manufacturer data.