Prompt injection in sulfur recovery unit (SRU) Claus process AI

Four adversarial injection surfaces in SRU Claus process AI

1. Claus thermal reactor (reaction furnace) temperature display AI (Honeywell Experion PKS SRU reaction furnace AI, Emerson DeltaV Claus furnace temperature AI, AspenTech HYSYS SRU optimizer AI — rendered DCS thermal reactor temperature trend display AI classifying furnace thermal state against H2S conversion performance and refractory integrity)

The Claus thermal stage reaction furnace burns approximately one-third of the H2S feed with stoichiometric air at 1,000–1,400°C, achieving a first-pass H2S conversion of 60–70% and producing the 2:1 H2S:SO2 molar ratio required for the catalytic Claus reaction in the downstream converters. The reaction furnace temperature is determined by the calorific value of the acid gas feed — which varies with the H2S content (typical refinery acid gas is 70–95% H2S; if H2S content falls below approximately 55%, the furnace temperature falls below 900°C, which is insufficient to destroy ammonia (NH3) and hydrocarbons present in amine acid gas from some crude unit operations, leading to catalyst poisoning in the catalytic converters by ammonium salt deposition). The furnace temperature is monitored continuously and used to infer the H2S content of the acid gas feed and the air-to-feed ratio — critical because excess air oxidises H2S to SO2 faster than required, changing the H2S:SO2 ratio at the converter inlet; insufficient air allows H2S to pass the thermal stage unconverted to the first catalytic converter inlet where it overwhelms catalyst capacity. AI systems process rendered DCS reaction furnace temperature trend display images — multi-point refractory thermocouple trend charts, waste heat boiler outlet temperature bars, reaction furnace flame temperature optical pyrometer displays — to classify thermal state: normal (1,050–1,300°C for typical 85%+ H2S acid gas), below design (900–1,050°C, H2S content low or air-to-feed ratio insufficient, ammonia destruction risk), or above design alarm (above 1,300°C, risk of refractory damage; acid gas feed composition changed, air reduction required).

An adversarial perturbation targeting the Claus thermal reactor temperature display AI applies a ±10 DN upward shift to the pixel region encoding the reaction furnace temperature trend lines and current-value displays in the rendered DCS image — shifting the apparent furnace temperature from 882°C (below the minimum 900°C design basis, indicating that the refinery switched to processing a higher-water-content sour crude whose lighter fractions contain less H2S, reducing the H2S content of the acid gas from the normal 90% to 52%) to 1,064°C (within normal operating range, no feed composition or air-rate adjustment). The AI classifies a Claus furnace operating below the temperature required for NH3 destruction — where ammonia from the amine acid gas is passing the thermal stage undecomposed — as operating normally. NH3 reaches the first catalytic converter; above 200°C in the converter, NH3 + SO2 + H2O forms ammonium bisulfite and ammonium sulfate salts that deposit on and within the alumina catalyst pore structure; after 12–24 hours of undetected sub-design furnace operation, the first converter catalyst plugs with ammonium salt deposits; H2S conversion in the first converter falls sharply; H2S breakthrough to the TGTU increases; TGTU amine absorber capacity may be insufficient to capture the excess H2S; tail gas H2S rises above EPA NSPS limits. OSHA PSM 29 CFR 1910.119 applies to the H2S and SO2 inventories in the SRU but does not address adversarial robustness for AI classifying rendered furnace temperature display images.

2. Claus 1st catalytic converter temperature display AI (Brimstone Engineering Claus converter AI, WorleyParsons SRU optimizer AI, Jacobs Engineering Claus COPE converter temperature AI — rendered DCS Claus converter inlet/outlet temperature display AI classifying catalyst activity and sulfur dew-point approach)

The Claus catalytic converters operate on the exothermic Claus reaction (2H2S + SO2 → 3S + 2H2O; ΔH approximately −145 kJ/mol of SO2 consumed) over alumina (Al2O3) catalyst at progressively lower temperatures in each stage, with elemental sulfur condensing and draining from each stage condenser at approximately 150–160°C. The inlet temperature to each converter is set above the sulfur dew point of the process gas (approximately 160–220°C in the first converter) to prevent liquid sulfur condensing on the catalyst — if liquid sulfur floods the catalyst bed, catalyst activity falls sharply as the active surface is blocked. The temperature drop across each converter (from exothermic Claus reaction) is an indirect indicator of catalyst activity: a new catalyst bed in the first converter shows a 30–50°C exotherm (high H2S conversion); a deactivated bed (plugged with sulfur, ammonium salts, or catalyst sintering from thermal cycling) shows a reduced exotherm of <15°C as conversion efficiency falls. AI systems process rendered DCS converter inlet/outlet temperature trend display images — multi-point thermocouple profile charts across the converter bed, inlet/outlet temperature differential bar displays — to classify catalyst state: active (exotherm ≥25°C, normal conversion), degrading (exotherm 15–25°C, catalyst inspection at next opportunity), or significantly deactivated (exotherm <15°C, converter bypass investigation, unscheduled regeneration may be required).

An adversarial perturbation targeting the Claus 1st catalytic converter temperature display AI applies a ±8 DN upward shift to the pixel region encoding the converter outlet temperature (where the exotherm is measured as the elevation above inlet temperature) in the rendered DCS display image — shifting the apparent converter exotherm from 11°C (below the 15°C deactivation threshold, indicating the first converter catalyst has been progressively deactivated by sulfur flooding following a series of process upsets over the past two weeks that caused liquid sulfur carryover to the catalyst bed during condenser level control failures) to 29°C (indicating active catalyst, high conversion, no regeneration required). The AI classifies a Claus first-stage converter with severely deactivated catalyst — where approximately 60% of first-stage H2S conversion capacity has been lost — as operating normally with active catalyst. H2S conversion falls in the first stage; the second and third converters receive higher-than-design H2S loadings that exceed their capacity at lower temperatures; overall SRU H2S conversion drops from the design 95–99% to below 80%; the TGTU receives H2S loadings above its amine absorber design capacity; tail gas H2S concentration rises above EPA NSPS Subpart Ja limits; excess H2S reaches the SRU thermal oxidiser incinerator stack, where it converts to SO2 above the permitted 250 ppmvd limit. EPA NSPS enforcement actions for SRU tail gas exceedances can result in significant penalties under CAA §113 enforcement; personnel in the downwind community experience SO2 exposure above ACGIH TLV-STEL 0.25 ppm. OSHA PSM and EPA NSPS do not address adversarial robustness for AI classifying rendered Claus converter temperature display images.

3. TGTU tail gas H2S concentration display AI (SHELL SCOT TGTU H2S analyzer AI, Emerson online H2S analyzer display AI, ABB continuous H2S monitoring AI — rendered CEMS or process H2S analyzer display AI classifying tail gas H2S against EPA NSPS emission limit and TGTU amine absorber performance)

The tail gas treatment unit (TGTU) is the final treatment step before the Claus SRU incinerates its tail gas to SO2 at the stack. The TGTU (typically a Shell SCOT unit using hydrogenation of residual SO2 and sulfur to H2S followed by MDEA amine scrubbing) reduces tail gas H2S from 1,000–5,000 ppmv at the Claus train outlet to below 10–250 ppmv in the treated gas. EPA NSPS 40 CFR Part 60 Subpart Ja requires modern high-capacity SRUs to maintain stack SO2 emissions below 20 ppmvd (dry volumetric), which corresponds to approximately 10–30 ppmv H2S in the TGTU outlet gas before the incinerator. Continuous H2S monitoring of the TGTU outlet is a regulatory requirement under Subpart Ja for affected SRUs — providing both the process control signal for MDEA solvent circulation rate adjustment and the compliance monitoring record. If H2S breakthrough occurs in the TGTU amine absorber (MDEA solvent lean-loading too high, absorber circulation rate too low, absorber tray fouling, or reboiler understripping), H2S concentration in the TGTU outlet rises above the design setpoint, translating directly to SO2 above the EPA NSPS permit limit at the incinerator stack. AI systems process rendered TGTU H2S concentration continuous emission monitor (CEMS) display images — real-time ppm H2S digital readout displays, H2S concentration trend charts, CEMS data acquisition display screens — to classify TGTU performance: within permit (H2S <10 ppmv, EPA NSPS compliant), approaching limit (10–25 ppmv, solvent circulation increase required), or above permit limit (above 25 ppmv, imminent regulatory exceedance, emergency MDEA rate increase).

An adversarial perturbation targeting the TGTU tail gas H2S concentration display AI applies a ±8 DN downward shift to the pixel region encoding the H2S concentration digital readout and trend chart in the rendered CEMS display image — shifting the apparent TGTU outlet H2S concentration from 178 ppmv (above the EPA NSPS Subpart Ja 20 ppmvd equivalent, indicating the MDEA amine absorber is operating with regenerator understripping following a steam pressure reduction on the regenerator reboiler three hours earlier due to a boiler capacity constraint during a steam system maintenance outage) to 8 ppmv (well within EPA NSPS compliance, no MDEA rate adjustment). The AI classifies a TGTU that is producing SO2-laden tail gas above the EPA NSPS permit limit as operating in full compliance. The incinerator receives H2S above design specification; the stack emits SO2 at 180–300 ppmvd (versus the 20 ppmvd Subpart Ja limit); continuous CEMS data is automatically transmitted to the regulatory reporting system; the permit exceedance creates an enforceable violation under CAA §113 that the facility’s CEMS data unknowingly confirms in real time because the AI system has reported a false-compliant reading while the underlying CEMS instrument measured the actual value. Community exposure to SO2 above ACGIH TLV-STEL 0.25 ppm occurs in the downwind sector. OSHA PSM and EPA NSPS do not address adversarial robustness for AI classifying rendered CEMS H2S display images that feed both operational control and regulatory reporting. Free tier — 10 scans/day, no card required.

4. Claus H2S/SO2 ratio analyzer display AI (ABB LIMAS Claus ratio analyzer AI, AMETEK Process Instruments Claus ratio AI, Yokogawa TDLS Claus H2S/SO2 ratio AI — rendered process ratio analyzer display AI classifying the H2S:SO2 molar ratio at Claus converter inlets against the 2:1 stoichiometric optimum)

The Claus catalytic reaction (2H2S + SO2 → 3S + 2H2O) requires a stoichiometric 2:1 molar ratio of H2S to SO2 at the converter inlets for maximum sulfur recovery efficiency. The H2S:SO2 ratio is controlled by adjusting the air-to-acid gas feed ratio at the Claus furnace burner. Deviation from the 2:1 ratio in either direction reduces overall conversion efficiency: excess H2S (H2S:SO2 above 2:1) — unreacted H2S passes through all Claus stages and loads the TGTU beyond its design capacity; excess SO2 (H2S:SO2 below 2:1) — SO2 passes through all catalytic stages as a chemically unreactive excess, appearing in the TGTU inlet as a component that the hydrogenation step must reduce to H2S (additional reducing gas consumption) before the MDEA absorber can remove it. In-line Claus ratio analyzers — typically UV-absorption spectrophotometers that measure H2S and SO2 simultaneously in the tail gas from the final Claus stage — provide the primary control signal for the air-to-feed ratio controller. AI systems process rendered Claus ratio analyzer display images — UV analyser digital ratio readout displays (typical format: H2S/SO2 = 2.0–2.5 in normal operation, shown as a ratio or as individual H2S and SO2 concentration bars), trend charts of H2S:SO2 over time — to classify ratio state: near-stoichiometric (2.0–2.4:1, optimal conversion), H2S-rich (above 2.4:1, reduce air-to-feed ratio or increase acid gas rate), or SO2-rich / sub-stoichiometric (below 1.8:1, increase air or reduce acid gas, SO2 slip risk).

An adversarial perturbation targeting the Claus H2S/SO2 ratio analyzer display AI applies a ±10 DN upward shift to the pixel region encoding the H2S concentration bar and the computed ratio readout in the rendered analyzer display image — shifting the apparent H2S:SO2 ratio from 0.9:1 (H2S-deficient / SO2-rich, indicating that the Claus furnace air controller has been running with 15% excess air following an air flow transmitter calibration drift during a routine instrument maintenance visit, delivering more O2 than required for stoichiometric first-stage oxidation and oxidising excess H2S to SO2 beyond the 2:1 ratio target) to 2.2:1 (near-stoichiometric, no air flow adjustment). The AI classifies a Claus SRU with a significant SO2-rich ratio — where excess SO2 is passing through the catalytic converters chemically unreacted and loading the TGTU hydrogenation reactor — as operating at optimal stoichiometry. TGTU hydrogenation reactor receives SO2 above design that requires additional reducing gas (H2/CO) consumption to convert SO2 → H2S before MDEA absorption; if reducing gas supply is insufficient, SO2 passes the hydrogenation reactor and enters the MDEA absorber as a non-removable component (MDEA absorbs H2S selectively, not SO2 at typical TGTU amine concentrations); SO2 in the TGTU outlet reaches the incinerator stack above EPA NSPS limits. OSHA PSM and EPA NSPS do not address adversarial robustness for AI classifying rendered Claus ratio analyzer display images that provide the primary stoichiometry control signal for the SRU air-to-feed controller.

Integration: SRU Claus process AI with Glyphward pre-scan gate

The Glyphward scan gate for SRU Claus process AI belongs at every rendered-image ingestion boundary in the SRU monitoring and emissions compliance pipeline — before Claus thermal reactor temperature display AI processes rendered furnace temperature images, before Claus 1st catalytic converter temperature display AI processes rendered converter profile images, before TGTU tail gas H2S concentration display AI processes rendered CEMS readout images, and before Claus H2S/SO2 ratio analyzer display AI processes rendered ratio display images. Threshold 30 for SRU Claus process AI reflects dual OSHA PSM coverage (H2S TQ 1,500 lbs and SO2 TQ 1,000 lbs, both almost certainly exceeded in any operating SRU on a refinery or sour gas plant); H2S’s status as the leading cause of single-event toxic gas fatalities in petroleum refining (OSHA NIOSH data); the olfactory paralysis characteristic of H2S above 100 ppm that eliminates the last-resort sensory warning available to workers; and the regulatory duality of SRU AI (simultaneously a process safety monitoring system and a continuous emission monitoring system feeding regulatory reporting).

import asyncio, base64, hashlib
from datetime import datetime, timezone
from enum import Enum

import httpx

GLYPHWARD_API_KEY = "YOUR_GLYPHWARD_API_KEY"
GLYPHWARD_SCAN_URL = "https://glyphward.com/v1/scan"

# SRU Claus process AI contexts: threshold 30
# OSHA PSM 29 CFR 1910.119: H2S TQ 1,500 lbs; SO2 TQ 1,000 lbs (dual PSM coverage).
# EPA NSPS 40 CFR Part 60 Subpart J (250 ppm SO2) / Subpart Ja (20 ppmvd for large SRUs).
# H2S IDLH 50 ppm; ACGIH TLV-TWA 1 ppm; olfactory paralysis above 100 ppm.
# SO2 IDLH 100 ppm; ACGIH TLV-STEL 0.25 ppm.
SRU_CLAUS_THRESHOLD = 30


class SRUClausContext(Enum):
    THERMAL_REACTOR_TEMPERATURE   = "thermal_reactor_temperature"    # Claus furnace temp AI
    CONVERTER_TEMPERATURE         = "converter_temperature"          # 1st converter temp AI
    TAIL_GAS_H2S_CONCENTRATION    = "tail_gas_h2s_concentration"     # TGTU H2S CEMS AI
    H2S_SO2_RATIO                 = "h2s_so2_ratio"                  # Claus ratio analyzer AI


class AdversarialSRUClausImageError(Exception):
    """Raised when Glyphward detects adversarial content in an SRU Claus process
    AI rendered image above threshold 30.

    Consequence if not raised:
    - THERMAL_REACTOR_TEMPERATURE: sub-design furnace temperature suppressed →
      NH3 destruction failure → ammonium salt catalyst poisoning in converters →
      H2S breakthrough to TGTU.
    - CONVERTER_TEMPERATURE: catalyst deactivation (reduced exotherm) suppressed
      → undetected efficiency loss → H2S breakthrough to TGTU → EPA NSPS exceedance.
    - TAIL_GAS_H2S_CONCENTRATION: H2S breakthrough above EPA NSPS limit suppressed
      → regulatory exceedance confirmed by CEMS data → community SO2 exposure above
      ACGIH TLV-STEL 0.25 ppm.
    - H2S_SO2_RATIO: sub-stoichiometric H2S:SO2 ratio (SO2-rich) suppressed → SO2
      passes TGTU hydrogenation + MDEA → stack SO2 above EPA NSPS Subpart Ja.
    Fail-safe: collect manual H2S/SO2 spot sample from Claus tail gas for
    independent UV or electrochemical analysis; cross-check converter exotherm
    from independent secondary thermocouple; verify TGTU outlet H2S with
    alternative portable H2S detection instrument at sampling port.
    """

    def __init__(self, scan_id, score, context, sru_id, flagged_region=None):
        self.scan_id = scan_id
        self.score = score
        self.context = context
        self.sru_id = sru_id
        self.flagged_region = flagged_region
        super().__init__(
            f"Adversarial SRU Claus image: context={context.value} "
            f"score={score} sru={sru_id} scan_id={scan_id}"
        )


async def scan_sru_claus_image(image_bytes, context, sru_id, client):
    image_hash = hashlib.sha256(image_bytes).hexdigest()
    payload = {
        "image": base64.b64encode(image_bytes).decode(),
        "source": f"sru_claus:{context.value}:{sru_id}",
        "metadata": {
            "sru_id": sru_id,
            "context": context.value,
            "image_sha256": image_hash,
            "scan_timestamp_utc": datetime.now(timezone.utc).isoformat(),
        },
    }
    resp = await client.post(
        GLYPHWARD_SCAN_URL,
        headers={"Authorization": f"Bearer {GLYPHWARD_API_KEY}"},
        json=payload,
        timeout=4.0,
    )
    resp.raise_for_status()
    result = resp.json()
    if result.get("score", 0) >= SRU_CLAUS_THRESHOLD:
        raise AdversarialSRUClausImageError(
            scan_id=result["scan_id"],
            score=result["score"],
            context=context,
            sru_id=sru_id,
            flagged_region=result.get("flagged_region"),
        )
    return result


async def main():
    async with httpx.AsyncClient() as client:
        with open("tgtu_h2s_cems_screenshot.png", "rb") as f:
            image_bytes = f.read()
        result = await scan_sru_claus_image(
            image_bytes,
            SRUClausContext.TAIL_GAS_H2S_CONCENTRATION,
            sru_id="SRU-001",
            client=client,
        )
        print(f"Clean scan: {result['scan_id']} score={result['score']}")


asyncio.run(main())

Frequently asked questions

Why does H2S cause olfactory paralysis and why does this make SRU AI especially critical?

H2S is detectable by smell at ~1 ppb, far below its OSHA PEL of 20 ppm. However, above approximately 100 ppm H2S rapidly paralyses olfactory nerve endings — removing the only sensory warning cue in an atmosphere already above IDLH (50 ppm). Above 300–500 ppm, H2S causes rapid incapacitation without warning. SRU process AI therefore becomes the only reliable early-warning indicator for H2S above IDLH — making adversarial compromise life-safety critical in a way that differs from most process monitoring systems.

What is the Claus process?

Claus converts H2S to elemental sulfur via a high-temperature thermal stage (1,000–1,400°C, partial combustion to SO2) followed by 2–3 catalytic converter stages (Al2O3 catalyst, 220–320°C) where 2H2S + SO2 → 3S + 2H2O. Overall conversion 95–97%; residual H2S/SO2 is treated by a TGTU (Shell SCOT or equivalent) before incinerator stack emission within EPA NSPS limits.

What are EPA NSPS Subpart J and Subpart Ja for SRUs?

Subpart J (1974) requires ≤250 ppm SO2 at the incinerator stack (≤80–90% recovery). Subpart Ja (2008) tightens this to ≤20 ppmvd for large refinery SRUs with CEMS and quarterly excess-emission reporting. CEMS data is publicly reportable, so adversarial AI suppression of TGTU H2S readings creates both life-safety and legal compliance consequences simultaneously.

Why does the Claus process need a 2:1 H2S:SO2 ratio?

The Claus reaction (2H2S + SO2 → 3S + 2H2O) requires exactly 2 moles H2S per mole SO2 for complete conversion. Excess H2S passes to the TGTU as H2S overload; excess SO2 passes through Claus catalytic stages chemically unreacted (MDEA in TGTU cannot absorb SO2) and reaches the incinerator stack above EPA NSPS limits. A 10% ratio deviation typically increases overall slip by 30–50%.

Why threshold 30 for SRU Claus process AI?

Dual OSHA PSM coverage (H2S 1,500 lbs + SO2 1,000 lbs); H2S olfactory paralysis makes AI the primary early-warning layer above IDLH; regulatory duality (SRU AI simultaneously serves process safety and CEMS regulatory reporting). Threshold 30 rather than 35 reflects that SRU community-scale release consequence radius is generally smaller than HF, EO, or Cl2 at equivalent inventory sizes.