Hydrogen sulfide (H2S) refinery amine treating AI adversarial injection: how a causal four-surface attack chain — reboiler steam valve upward — lean amine loading — absorber breakthrough — area CEMS — eliminates both olfactory fatigue warning and engineered alarm simultaneously, with 2,840 ppm H2S (56.8× IDLH) displayed as 94 ppm normal — OSHA PSM TQ 1,500 lbs, dual safeguard elimination, Glyphward threshold 35

H2S chemistry, refinery amine treating, and the OSHA PSM TQ 1,500 lbs calibration

Hydrogen sulfide is a colourless gas at ambient temperature and pressure, with a boiling point of −60.2°C and vapour pressure at 20°C of approximately 18.7 bar — meaning it is stored and transported as a liquefied compressed gas in pressure vessels rated for at least 20 bar, or dissolved in sour process streams under refinery operating pressures of 3–90 bar. H2S is produced in petroleum refining from the high-temperature, high-pressure desulfurisation reactions (hydrodesulfurisation, HDS) of organic sulfur compounds in crude oil: thiophenes, benzothiophenes, dibenzothiophenes, mercaptans, and sulfides react with hydrogen over cobalt-molybdenum or nickel-molybdenum catalysts at 300–400°C and 30–120 bar to release H2S. The resulting H2S-rich refinery off-gas streams — hydrogen-rich recycle gas from hydrotreaters, off-gas from catalytic crackers, absorber off-gas from naphtha and diesel hydrotreaters — contain H2S at concentrations from 0.1% to several volume percent, which must be removed before these streams can be used as hydrogen recycle, processed in downstream units, or burned as refinery fuel gas.

Amine treating achieves this removal through reversible acid-gas absorption chemistry. H2S (pKa1 = 7.0, a weak acid) reacts with alkanolamine bases in aqueous solution to form water-soluble amine-hydrosulfide salts: for MDEA (pKb1 ≈ 5.4; pKb2 ≈ 8.3), the absorption reaction is R3N + H2S ⇋ R3NH⁺ + HS⁻ — reversible at the elevated temperature of the regenerator, where the equilibrium shifts toward amine + H2S regeneration. The capacity of the amine to absorb H2S is quantified as the H2S loading (mol H2S per mol amine): fresh lean amine from the regenerator at design conditions carries <0.01–0.05 mol/mol H2S (essentially free of H2S), providing maximum absorption driving force; rich amine leaving the absorber base carries 0.4–0.7 mol/mol (partially saturated). The lean amine loading specification — the maximum acceptable H2S loading in the regenerated amine returned to the absorber — is typically 0.05–0.10 mol/mol for MDEA in high-H2S-content service; exceeding this specification reduces the absorption driving force in the absorber, increases treated gas H2S residual (breaks the treated gas specification), and signals inadequate regeneration. The reboiler heat input rate — controlled by the steam supply valve — is the primary independent variable determining lean amine quality: insufficient reboiler heat produces inadequate H2S stripping and high lean amine loading. This is the causal root in the adversarial attack chain described in this post.

OSHA PSM 29 CFR 1910.119 Appendix A lists H2S at a threshold quantity of 1,500 lbs (680 kg). H2S inventories at refinery amine treating units — in the form of H2S dissolved in rich amine, in acid gas piping between absorber and Claus plant, and in sour water storage vessels — routinely exceed this TQ by orders of magnitude, placing virtually every sour crude refinery amine treating operation under full PSM program requirements. The 1,500 lb TQ calibration reflects H2S's occupational lethality profile: ACGIH TLV-C 1 ppm (ceiling concentration, not to be exceeded for any duration), NIOSH REL 1 ppm (10-minute ceiling), OSHA PEL 20 ppm ceiling (29 CFR 1910.1000 Table Z-2), and NIOSH IDLH 50 ppm. The IDLH calibration of 50 ppm is particularly consequential: it represents both the maximum concentration from which a healthy adult can escape within 30 minutes without irreversible health effects and — coincidentally — the lower bound of the olfactory fatigue range, where H2S begins to paralyse the very sensory system that provides natural warning at lower concentrations. EPA RMP 40 CFR Part 68 lists H2S at TQ 10,000 lbs for community consequence analysis, with worst-case toxic endpoint distances calculated using ERPG-2 for H2S (30 ppm per published ERPG data) that can extend hundreds of meters from major sour gas processing facilities.

Olfactory fatigue — how H2S eliminates its own natural warning above the NIOSH IDLH

H2S's olfactory fatigue property is arguably the single most counterintuitive toxicological fact in industrial hygiene: a chemical whose primary natural warning mechanism — its extraordinarily low odour threshold of 0.008–0.13 ppm — operates four to five orders of magnitude below the dangerous exposure range, loses that warning capability entirely at the concentration where protection is most urgently needed. At 0.1–10 ppm, H2S produces a strong, offensive rotten-eggs odour that is distinctive, unmistakeable, and essentially impossible to habituate to in the short term. Workers in amine units routinely use the onset of H2S odour as an informal low-level indicator that a process upset has increased atmospheric H2S from background levels — even though this practice carries the serious risk that olfactory fatigue may have already occurred by the time they notice the smell.

The physiological mechanism of olfactory fatigue in H2S is not fully elucidated but involves two overlapping processes. First, receptor-level saturation: H2S activates the T2R38 bitter taste receptor expressed on human olfactory sensory neurons through a mechanism involving polysulfide intermediates generated by H2S oxidation at the mucus layer. At H2S concentrations above approximately 50 ppm, the rate of receptor activation exceeds the recovery rate, saturating the receptor population and reducing the population-average firing rate of olfactory sensory neurons toward the background noise floor. Second, cytochrome inhibition at the olfactory neuron level: H2S inhibits cytochrome c oxidase (complex IV of the mitochondrial electron transport chain) by binding to the haem iron of the cytochrome a3 component, reducing the metabolic capacity of olfactory sensory neurons and impairing their capacity to maintain the ionic gradients needed for sustained action potential generation. This is the same cellular mechanism responsible for H2S's systemic lethality at higher concentrations, but localised initially in the high-metabolic-demand olfactory sensory neurons before systemic effects manifest. The consequence is that H2S at 50–150 ppm produces a brief period of intense rotten-eggs odour perception — often lasting 30 seconds to 2 minutes — followed by complete olfactory paralysis: the smell stops entirely, creating the dangerous impression that the atmosphere has cleared.

The safety implications have been documented in multiple fatal incident investigations. In confined-space H2S fatalities, a recurring pattern involves a worker who reports “smelling eggs briefly then it went away” before entering a space that was not properly purged — the olfactory fatigue masking the persistence of lethal concentrations. The NIOSH recommendation to never rely on smell to detect H2S has been standard industrial hygiene guidance since the 1970s, and is the regulatory basis for the fixed-point CEMS requirement in PSM-covered H2S facilities. The area CEMS network exists specifically to provide an olfaction-independent alarm signal that compensates for the known and well-characterised failure of H2S's natural odour warning above 50 ppm. This is the first safeguard already eliminated — by H2S physiology, not by any adversary — before the adversarial attack on the AI monitoring system is ever applied. The adversarial attack on the area CEMS AI display then eliminates the second and only remaining engineered safeguard, completing the dual safeguard elimination.

Four adversarial injection surfaces in amine treating unit monitoring AI — the causal chain

1. Reboiler steam supply valve position AI — Honeywell Experion PKS amine regenerator reboiler AI / Emerson DeltaV S-series reboiler control AI / Yokogawa OpreX amine regenerator reboiler AI / Fisher DVC6200 digital valve controller AI (±8 DN upward — 8% open near-closed actuator failure displayed as 82% open adequate steam supply)

The amine regenerator reboiler is a kettle or thermosiphon steam heat exchanger that maintains the regenerator column base temperature at 120–130°C (for MDEA service; MEA service requires 105–120°C). The steam supply control valve — typically a Fisher Controls, Flowserve, or Emerson Final Control valve with a pneumatic diaphragm actuator and positioner — is the primary heat-input manipulation variable for the regenerator temperature controller. The positioner receives a 4–20 mA control signal from the DCS regenerator temperature controller and translates it to instrument air pressure on the actuator diaphragm, positioning the valve between full-closed (0%) and full-open (100%). The actuator design is fail-closed for steam services: loss of instrument air pressure causes the actuator spring to return the valve to the closed position, isolating the steam supply as the safe-failure mode. Honeywell Experion PKS and Emerson DeltaV AI monitoring systems process rendered DCS valve position indicator display images — a bar indicator at 0–100% open position, 200 pixels height (0.5% per pixel) — to classify reboiler steam valve status against normal operating range (typically 40–80% open at design regeneration throughput) and alarm setpoints.

The adversarial injection scenario for surface 1: a pneumatic instrument air supply compressor serving the amine unit control valves has developed a seal leak, reducing instrument air header pressure from the design 80 psig to 18 psig over a 6-hour period during the night shift. The reboiler steam supply valve actuator — spring-to-fail-closed, design instrument air signal 3–15 psig at 4–20 mA — has lost the driving instrument air pressure needed to overcome the return spring at design valve positions; at 18 psig header pressure, the actuator can only reach approximately 8% open against the return spring force before the available pressure equals the spring load. The valve is essentially closed: reboiler steam flow drops to approximately 8% of design, supplying approximately 6% of the required regeneration heat. The DCS valve position indicator display shows a bar indicator on the 0–100% scale at 200 pixels (0.5%/px): actual 8% position = (8/100) × 200 = 16 pixels from the bottom. The ±8 DN upward adversarial perturbation on the pixel values encoding the valve position bar shifts the apparent indicator from 16 pixels to approximately 164 pixels, producing an apparent reading of (164/200) × 100 = 82% open: the midpoint of the normal operating range for this reboiler service. The AI classifies the reboiler steam valve as operating normally: no low-position alarm; no instrument air pressure investigation; no operator manual valve override; no temperature alarm for the regenerator bottoms. Over the following 4 hours, the regenerator bottoms temperature falls from 125°C to 82°C; lean amine H2S loading rises progressively from 0.04 mol/mol (normal) to 0.42 mol/mol (4.2× specification maximum). The physically undetected steam valve failure, hidden by the surface 1 upward adversarial perturbation, initiates the entire downstream causal chain. Glyphward free-tier scan — submit a rendered DCS reboiler steam valve position display image for adversarial robustness baseline scoring before deploying AI monitoring on amine treating unit reboiler controls subject to OSHA PSM TQ 1,500 lbs.

2. Lean amine H2S loading analyser AI — Ametek Model 150 online H2S analyser AI / ABB AWT420 analytical transmitter lean amine AI / Emerson Rosemount 5081 lean amine loading analyser AI / Schlumberger InSitu lean amine composition AI (±10 DN downward — 0.42 mol/mol at 4.2× specification maximum displayed as 0.08 mol/mol within-specification)

Lean amine H2S loading — the mol H2S per mol amine in the regenerated amine stream leaving the regenerator base and being pumped to the absorber — is the primary quality indicator of amine regeneration performance. Online lean amine analysers (electrochemical, gas chromatographic, or spectroscopic) measure H2S loading continuously, with typical measurement ranges of 0–0.5 mol/mol and alarm setpoints at 0.10 mol/mol (design maximum for MDEA high-H2S-content service), 0.15 mol/mol (process alert: regeneration performance has degraded significantly; investigate reboiler duty and amine quality), and 0.25 mol/mol (high alarm: amine quality too poor for absorber service; recirculate to regenerator; notify process engineer). AI monitoring systems from Emerson and ABB process rendered online analyser display images — a bar indicator on 0–0.50 mol/mol range, 200 pixels height (0.0025 mol/mol per pixel) — to classify lean amine quality against the specification thresholds.

The adversarial injection scenario for surface 2: after 4 hours of regenerator undersupply (surface 1 root cause), the lean amine H2S loading has reached 0.42 mol/mol. On the 0–0.50 mol/mol display at 200 pixels (0.0025 mol/mol per pixel): actual 0.42 mol/mol = (0.42/0.50) × 200 = 168 pixels from the bottom. The ±10 DN downward adversarial perturbation shifts the apparent analyser bar from 168 pixels to approximately 32 pixels, producing an apparent reading of (32/200) × 0.50 = 0.08 mol/mol: within the 0.10 mol/mol specification maximum, near the lower end of the normal operating range for lean MDEA. The lean amine loading AI classifies the regenerated amine as within specification: no lean amine quality alarm; no call to reduce the lean amine pump rate to allow more regeneration time; no investigation of the reboiler temperature that is already 43°C below setpoint (but undetected due to surface 1). The near-exhausted lean amine at 0.42 mol/mol H2S loading — carrying 0.42 mol H2S per mol MDEA, compared to the <0.05 mol/mol of adequately regenerated lean amine — is continuously pumped by the lean amine circulation pump to the top of the absorber column at the design rate of approximately 500–1,000 tonnes per hour of amine solution. In the absorber, this near-saturated amine encounters the sour gas stream; the absorption driving force (the difference between actual and equilibrium H2S loading) is nearly zero — the amine is too close to saturation to absorb meaningful additional H2S from the gas. The foaming condition in the absorber — from surfactant contamination of the recirculating amine — simultaneously reduces gas-liquid interfacial area to below 15% of design. The combined effect: absorber H2S removal drops from the design >99.9% to below 3%; treated gas H2S rises from the 4–10 ppm specification to 2,840 ppm.

3. Absorber outlet H2S analyser AI — Emerson Rosemount CT5400 absorber outlet H2S AI / Yokogawa GS200 treated gas H2S AI / ABB ACX spectrometer absorber outlet AI / Servomex 4900 absorber overhead H2S AI (±8 DN downward — 2,840 ppm H2S breakthrough at 56.8× IDLH displayed as 94 ppm within instrument range)

The absorber outlet H2S analyser monitors the treated “sweet” gas stream leaving the absorber overhead for H2S residual content, providing the primary process quality measurement for the amine treating unit. In normal operation, treated gas H2S from an MDEA amine treating unit is below 4–10 ppm H2S — the specification required for downstream hydrogen plant, catalytic reformer, or fuel gas system service. The analyser is typically a continuously sampling extractive process GC, an infrared spectrometer (for high-range measurement during upset), or an electrochemical sensor (for ppm-range measurement during normal operation). Because treated gas H2S can range from trace levels (normal) to percent-level concentrations during severe absorber failure, the analyser range for upset monitoring is typically extended to 5,000 ppm or 1 vol%. AI monitoring systems process rendered treated gas H2S analyser display images — a bar indicator on a 0–4,000 ppm display range, 200 pixels height (20 ppm per pixel) — to classify absorber outlet H2S against the normal operating specification.

The adversarial injection scenario for surface 3: with near-exhausted lean amine (0.42 mol/mol, surface 2) recirculating to the foamed and flooded absorber column, treated gas H2S has risen to 2,840 ppm — 56.8× the NIOSH IDLH of 50 ppm and approximately 284× the typical treated gas H2S specification. The downstream consequence begins immediately: the 2,840 ppm H2S-rich treated gas stream enters the downstream process train, poisoning the catalytic reformer guard bed (above 0.1 ppm) and generating SO2 in the refinery fuel gas system above EPA NESHAP limits. More immediately, H2S-laden treated gas vents within the amine unit plot area through absorber overhead relief paths, process equipment flanges, and instrument connections: the unit atmospheric H2S concentration rises to 84 ppm in the amine unit working area (surface 4). On the 0–4,000 ppm absorber outlet H2S display at 200 pixels (20 ppm per pixel): actual 2,840 ppm = (2,840/4,000) × 200 = 142 pixels from the bottom. The ±8 DN downward adversarial perturbation shifts the apparent analyser bar from 142 pixels to approximately 5 pixels, producing an apparent reading of (5/200) × 4,000 = 100 ppm — close to the 94 ppm reading the adversarial attack produces in this scenario (the apparent value varies with exact pixel gradient and rendering artefacts at the bar edge). The absorber outlet H2S AI classifies the treated gas stream as producing a minor upset reading of approximately 94 ppm — above the 10 ppm specification, but classified as a routine instrument exceedance rather than a 56.8× IDLH absorber flooding event. No absorber flooding alarm; no amine foaming investigation; no treated gas diversion to flare to prevent downstream catalyst poisoning; no amine unit evacuation for atmospheric H2S exposure risk.

4. H2S area CEMS AI — Honeywell Analytics Midas H2S fixed-point detector AI / MSA Ultima XE H2S CEMS AI / Dräger X-am 2500 H2S area monitor AI / Industrial Scientific Radius BZ1 H2S network AI (±8 DN downward — 84 ppm at 1.68× IDLH displayed as 2.8 ppm below action level — dual safeguard elimination with olfactory fatigue)

H2S area CEMS networks at PSM-covered amine treating units consist of fixed-point electrochemical H2S detectors placed at absorber base vent points, rich amine pump seals, reboiler connection points, and along the main pipe rack to detect any H2S atmospheric release from amine unit process equipment. Alarm setpoints typically follow a tiered structure: first alarm at 1 ppm (TLV-C ceiling; SCBA donning required for work in the area), second alarm at 10 ppm (OSHA PEL; area evacuation of non-essential personnel), high alarm at 50 ppm (NIOSH IDLH; immediate full evacuation, emergency response activation). The CEMS network at 50 ppm IDLH alarm represents the engineered safeguard designed specifically to compensate for olfactory fatigue: because H2S smell disappears above 50 ppm, the IDLH-alarm CEMS is the only reliable remaining warning mechanism for workers in H2S atmospheres above the IDLH. AI monitoring systems from Honeywell Analytics, MSA Safety, and Dräger Safety process rendered H2S area CEMS display images — a bar indicator on 0–100 ppm display range, 200 pixels height (0.5 ppm per pixel) — to classify atmospheric H2S against the tiered alarm setpoints.

The adversarial injection scenario for surface 4 — and the dual safeguard elimination: two maintenance technicians performing routine lean-amine pump seal inspection in the amine unit are exposed to the 84 ppm H2S atmosphere created by the surface 3 absorber breakthrough venting event. Both technicians initially noticed a strong rotten-eggs smell approximately 40 minutes earlier — when the absorber outlet H2S was approximately 60 ppm and the amine unit atmospheric concentration was first rising above 20 ppm. They attributed the smell to a minor process upset and continued working. At 84 ppm, the smell has now disappeared entirely: olfactory fatigue has paralysed their olfactory nerve, consistent with the physiological time course of olfactory paralysis at this concentration range (typically 30 seconds to 3 minutes above 100 ppm; somewhat longer at 84 ppm). Both technicians believe the process upset has resolved because the smell stopped. This is the elimination of the first safeguard. The amine unit CEMS registers 84 ppm H2S at the nearest fixed-point detector (located 8 metres from the lean amine pump seal station where the technicians are working). On the 0–100 ppm H2S CEMS display at 200 pixels (0.5 ppm per pixel): actual 84 ppm = (84/100) × 200 = 168 pixels from the bottom. The ±8 DN downward adversarial perturbation shifts the apparent CEMS bar from 168 pixels to approximately 6 pixels, producing an apparent reading of (6/200) × 100 = 3 ppm — within rounding of the 2.8 ppm the adversarial attack produces in this scenario. The H2S area CEMS AI classifies the amine unit atmospheric H2S as 2.8 ppm: below the 10 ppm OSHA PEL; well below the first alarm threshold; consistent with routine background-level H2S from minor process equipment leaks. No IDLH alarm; no area evacuation; no emergency H2S response activation. The two maintenance technicians continue their pump seal work in an atmosphere at 84 ppm H2S — 1.68× IDLH — with no olfactory warning (olfactory fatigue) and no electronic warning (CEMS AI adversarial suppression). This is the dual safeguard elimination: both protections against undetected lethal H2S exposure — the natural physiological one and the engineered electronic one — are simultaneously absent. At 84 ppm, continued exposure without SCBA creates acute H2S toxicity risk within minutes; if the absorber breakthrough worsens (as the process chemistry predicts it will, since the root-cause steam valve failure remains unresolved and the lean amine loading continues degrading), atmospheric H2S in the amine unit can reach 300–700 ppm, at which incapacitation occurs within 3–5 minutes, and potentially 700–1,000 ppm, at which H2S's cytochrome c oxidase inhibition causes respiratory muscle paralysis and unconsciousness within 30–60 seconds — too fast for self-rescue once the event is recognised.

The dual safeguard elimination — why H2S amine treating AI adversarial injection is categorically distinct in the Glyphward portfolio

Every other CEMS adversarial attack documented in the Glyphward industrial AI portfolio — HF alkylation, BF3 handling, phosgene production, MIC storage, HCN storage, Cl2 generation, AsH3 semiconductor fab, PH3 fumigation, B2H6 CVD, ClF3 handling — eliminates the engineered electronic alarm while leaving the natural physiological warning at least partially intact. HF at 18 ppm (BF3 handling scenario) produces severe respiratory irritation, burning eyes, and an intensely acidic sensation at the mucosa — symptoms that are unmistakeable and intensify rapidly with continued exposure, driving self-rescue even if the CEMS shows below-TLV-C. Phosgene at 1.2 ppm IDLH has a mild hay-like odour at that concentration that, while subtle, is perceptible. Cl2 at near-IDLH concentrations produces intense eye and respiratory irritation that is unmistakeable. AsH3 at 1.4× IDLH produces garlic-like odour perceptible above 0.5 ppm. In each case, the adversarial AI attack on the CEMS removes the primary engineered safeguard, but the natural physiological response to the toxic atmosphere — smell, irritation, discomfort — provides a residual self-rescue trigger.

H2S at 84 ppm (1.68× IDLH) provides no such residual. The olfactory fatigue mechanism eliminates smell perception before the CEMS adversarial attack is even needed to suppress the engineered alarm. Workers at 84 ppm H2S with olfactory fatigue have zero sensory indication of their exposure: no smell, no respiratory irritation at this concentration (H2S's respiratory irritation threshold is approximately 150–200 ppm, higher than the area concentration in this scenario), no visual indication (H2S is colourless and does not produce visible plumes at atmospheric concentrations). The dual safeguard elimination creates a scenario in which the total absence of any warning signal — sensory or electronic — coexists with a concentration 1.68× IDLH and a process chemistry trajectory pointing toward rapid worsening, because the root-cause steam valve failure (the sole initiating event) remains completely undetected and unresolved by the surface 1 adversarial attack.

The causal chain structure also means that standard incident investigation would encounter exceptional difficulty establishing the link between the adversarial attack and the outcome. A post-incident investigation finding workers incapacitated in the amine unit would investigate the H2S atmospheric source (absorber breakthrough — visible from process data if the absorber outlet analyser data is recovered) and the amine regeneration failure (lean amine loading degradation — visible from the lean amine analyser historian). But the steam valve position display — the surface 1 upward adversarial attack — would appear in the DCS historian as showing 82% open throughout the incident period; the actual valve position (8% open due to instrument air failure) would only be discoverable by a field inspection of the valve and instrument air header, which might not be prioritised in an acute H2S incident response. The adversarial attack on surface 1 creates a false audit trail in which the root cause — steam valve failure — appears normal in the DCS record, directing investigation toward amine quality issues (foaming, degradation products, contamination) as the apparent initiating cause, rather than the instrument air pressure failure that the adversarial attack concealed. See the Glyphward H2S amine treating SEO reference page for the full four-surface pixel-displacement audit and the causal chain process chemistry analysis.

Frequently asked questions