Prompt injection in steam methane reforming SMR hydrogen production AI

Three adversarial injection surfaces in SMR hydrogen production AI

1. Steam-to-carbon (S/C) ratio display AI (Yokogawa ADMAG AXF steam flow AI / Rosemount 3051SF steam flow transmitter SMR AI / Endress+Hauser Prowirl 73 steam vortex flowmeter reformer AI / Honeywell STD930 S/C ratio DCS AI / Emerson Daniel 3095MV steam/carbon ratio SMR AI — rendered DCS ratio-trend display AI classifying the molar steam-to-carbon ratio in the reformer feed against the ≥3.0 design minimum preventing Ni catalyst coking and the 3.0–3.8 operating window for natural gas feed; 55th upward-direction attack — FIRST SMR hydrogen production attack; FIRST steam methane reforming attack; FIRST Ni reformer catalyst coking attack; FIRST steam-to-carbon ratio upward attack; FIRST water-gas shift reactor attack)

The S/C ratio (moles of steam per mole of total carbon in the reformer feed) is the single most critical process variable governing Ni catalyst life in SMR. A 50 MMscfd (million standard cubic feet per day) H2 SMR plant burns approximately 760 tonnes/day natural gas (as CH4) and injects 1,900 tonnes/day process steam from a steam generation system (steam drum at 35 bar; typically generating steam from feed preheating, process heat recovery, and auxiliary boiler); at S/C = 3.0, the CH4:steam molar ratio is 1:3.0 (steam = 3× carbon mol). The steam flow is measured by a mass flow transmitter (vortex meter or orifice plate with temperature/pressure compensation) on the steam injection line; the carbon flow is calculated from the total hydrocarbon feed flow (natural gas molar flow × carbon content); the S/C ratio is a calculated display from these two measurements. Below S/C = 2.5 at 800°C tube-wall temperature (typical upper-third of radiant reformer tube), Boudouard coking is thermodynamically spontaneous (ΔG = −20 kJ/mol CO at 700°C; ΔG ≈ 0 at 800°C; kinetically significant at Ni surface temperatures ≈700–820°C on the catalyst near the tube wall). Carbon whisker growth (graphite C nucleated at Ni-Al2O3 interface; whisker diameter 20–200 nm; growth rate 0.3–0.8 mg C/g catalyst/hr at 800°C, S/C 2.1): whiskers grow in the Ni catalyst pellet pores, mechanically disintegrate the pellet (catalyst pore volume 0.35–0.55 cm³/g; ultimate pellet strength 300–500 N; whisker growth creates internal pressure 40–120 MPa in micropores exceeding pellet tensile strength), producing catalyst dust that settles at the tube bottom and blocks the 5–12 mm pellet bed void volume. The blocked bed section increases tube pressure drop from design 0.6 bar (per tube) to 2.4–4.8 bar at 50% blockage; increasing ΔP causes flow maldistribution among the 120–350 parallel reformer tubes: under-flowing tubes receive less steam and CH4 feed at the same tube temperature set by the furnace burners — creating a positive feedback where S/C in the under-flowing tube drops further below 2.5, accelerating coking in that tube, increasing ΔP further, and reducing flow to near-zero. At near-zero flow, the tube heats to 920–960°C (furnace temperature maintained by burner control; tube is not removing heat endothermically from the reforming reaction); at 940°C and 25 bar, HP-Nb tube creep rate (ASTM A297 HP-Nb; 0.01% creep limit 30,000 hr at 900°C; 30,000 hr at 850°C — 40-year design base) is approximately 50× the design creep rate; tube rupture within 2–6 hours of sustained 940°C / zero flow.

An adversarial perturbation targeting the S/C ratio display AI applies a ±8 DN upward shift to the pixel region encoding the S/C ratio in the rendered DCS ratio-trend display — shifting the apparent S/C ratio from 2.1 (steam control valve FV-201 actuator spring failure reduced steam flow from 1,900 to 1,330 tonnes/day; natural gas feed unchanged at 760 t/day carbon; S/C = 1,330×(18/MW)/(760×1,000/16) ≈ 2.1) to 3.4 (above the 3.0 minimum; classified as normal; no action). The DCS reports “SMR steam-to-carbon ratio nominal.” Within 45 minutes, Ni catalyst pellets in the upper-third of reformer tubes at 780–820°C are initiating carbon whisker growth; at T + 3 hours, tube ΔP on 15% of reformer tubes has risen to 2.8 bar; at T + 6 hours, 8% of tubes are at near-zero flow; at T + 8 hours, three reformer tubes exceed 930°C skin temperature (confirmed by automated infrared scanner reading — but IF the infrared AI is also compromised, these data are also suppressed). Tube rupture releases H2 (concentration in reformer tube gas 60–72 mol% dry) + CO (4–8 mol%) at 850°C and 25 bar into the gas-fired furnace firebox — H2 flash fire in the firebox (H2 LFL 4%; autoignition 500°C; firebox at 900–1,100°C = immediate ignition) plus CO release at OSHA PSM TQ. This is the 55th upward-direction attack in the Glyphward portfolio — the FIRST SMR/steam methane reforming hydrogen production attack; FIRST Ni reformer catalyst coking attack; FIRST steam-to-carbon ratio upward attack. Free tier — 10 scans/day, no card required.

2. Reformer tube skin temperature display AI (Yokogawa IR3000 infrared reformer tube temperature scanner AI / Thermatemp IRTSS SMR reformer tube AI / Honeywell TDC 3000 reformer tube pyrometer AI / Emerson DeltaV tube-skin thermocouple AI / Land Instruments Cyclops C90L reformer tube temperature AI — rendered infrared scanner output display AI classifying the radiant-section reformer tube skin temperatures against the 870–900°C design operating range and the 930°C upper limit for HP-Nb alloy long-term creep at 25 bar operating pressure; HIGH displayed temperature = hot spot; operators must reduce firing; UPWARD attack shows LOWER temperature than actual, hidden in this page under surface 1 framework)

Reformer tube skin temperature is continuously monitored by automated infrared (IR) scanner systems (Thermatemp IRTSS, Yokogawa IR3000, Land Instruments Cyclops series) that traverse the full tube bundle view through peepholes in the furnace wall. The IR scanner generates a temperature profile image of all tubes (120–350 tubes depending on plant size; temperature range 700–1,100°C; accuracy ±10°C; emissivity correction for HP-Nb alloy ε = 0.82–0.88). The AI classification task is: (a) identify any tube with skin temperature >920°C (single-tube hot spot; reduce burner firing in that cell); (b) identify systematic temperature pattern deviations (one-sided tube heating; evidence of flow maldistribution); (c) confirm all tubes are within ±30°C of mean (normal distribution; burner uniformity). At design conditions with S/C = 3.0, all tubes are 860–895°C at the peak radiant zone. When S/C drops to 2.1 and coking begins (as in the 55th upward attack scenario), 3–8 tubes develop hot spots at 920–960°C within 6–8 hours; the IR scanner AI should detect these as Class A (immediate shutdown required) anomalies. If this AI is ALSO compromised by adversarial perturbation (showing all tubes at 875–895°C when actual distribution is 860–958°C), the critical hot-spot alert is suppressed simultaneously with the S/C ratio alert — a multi-channel attack that removes both primary and secondary safeguards for tube rupture prevention.

3. High-temperature water-gas shift (HTS) reactor inlet temperature display AI (Yokogawa DPharp EJA110A HTS reactor inlet temperature AI / Rosemount 3144P WGS high-temperature shift AI / Endress+Hauser iTEMP TMT84 HTS catalyst bed inlet temperature AI / Honeywell STT850 SmartLine HTS inlet AI / ABB TSP321 water-gas shift HTS reactor temperature AI — rendered DCS temperature trend AI classifying the HTS reactor inlet temperature against the 310–400°C design operating range for Fe3O4/Cr2O3 catalyst activity and CO conversion above 95%)

The high-temperature shift (HTS) reactor (adiabatic fixed-bed; Fe3O4/Cr2O3/Cu catalyst promoted; pellet 6×6 mm cylinders; bed volume 20–80 m³ at medium-scale SMR; inlet temperature 310–350°C at design; adiabatic temperature rise 30–80°C depending on inlet CO concentration 8–14 mol% dry) converts CO + H2O → CO2 + H2 at equilibrium-limited CO conversion of 80–92% per pass. The Fe3O4/Cr2O3 HTS catalyst has a narrow operating window: below 300°C, catalyst activity drops below 60% of design conversion (Fe3O4 reduced to Fe by H2 at <280°C; “wetting” of the catalyst surface by condensed steam if the inlet is below the steam dew point at operating pressure); above 420°C, catalyst sintering accelerates and Fe3O4 may be partially reduced to Fe0 (methanol and methane side reactions increase; Fischer-Tropsch side reactions produce C4–C12 hydrocarbons that poison the downstream low-temperature shift Cu/ZnO catalyst and the PSA adsorbent beds). PSA H2 purification operates downstream: a 12–16 bed PSA (Linde HYSEC, Air Products HiPure, or UOP Polybed) cycles at 30 bar adsorption / 1–2 bar regeneration; each bed contains zeolite 5A or activated carbon adsorbent that captures CO2, CO, CH4, H2O, and N2 from the reformer/shift effluent; the PSA produces 99.999% H2 (N2 < 1 ppm; CO < 10 ppm). If HTS inlet temperature drops to 268°C (steam-drum pressure reduced from 35 bar to 22 bar due to a high-level steam-drum alarm shutting one of three boiler feedwater pumps; steam temperature at 22 bar is 218°C; but after superheating the steam still reaches reformer at 620°C; however HTS reactor quench steam injection reduced, inlet temperature drops from 330°C to 268°C) while AI shows 328°C, the CO conversion in the HTS drops from 88% to 54%; CO in the HTS exit (normally 3.5 mol% dry; now 8.4 mol% dry) overloads the low-temperature shift (LTS; Cu/ZnO catalyst at 210–250°C, normally handling 3.5 mol% CO inlet; now 8.4 mol% generates excessive heat ΔT = +180°C; LTS overtemperature to 410°C causes permanent Cu/ZnO catalyst sintering and deactivation within 12 hours) and the PSA beds receive CO at 1.8 mol% instead of design 0.5 mol% — PSA breakthrough of CO (>10 ppm in H2 product) triggers customer quality alarm at ammonia synthesis or PEM electrolyser end-users (CO poisons ammonia synthesis catalyst Fe-promoted at >5 ppm CO; CO poisons PEM fuel-cell cathode Pt catalyst at >10 ppm CO irreversibly).

Integration: SMR hydrogen production AI with Glyphward pre-scan gate

Glyphward integrates as a pre-scan gate at every rendered-image ingestion boundary in the SMR hydrogen plant monitoring pipeline — before steam/carbon ratio AI processes rendered DCS ratio-display images, before reformer tube skin temperature AI processes rendered IR scanner output images, and before HTS reactor AI processes rendered DCS temperature-trend images. Threshold 28 for SMR H2 AI reflects: OSHA PSM H2 TQ 10,000 lbs (LFL 4%; autoignition 500°C; wide flammability 4–75%); CO TQ 1,000 lbs (NIOSH IDLH 1,200 ppm); reformer tube rupture BLEVE potential; 95% of global H2 supply chain dependency.

import asyncio, base64, hashlib
from datetime import datetime, timezone
from enum import StrEnum, auto
from typing import Any
import httpx

GLYPHWARD_API = "https://api.glyphward.com/v1/scan"
GLYPHWARD_KEY = "gw_prod_***"

# SMR hydrogen production AI contexts: threshold 28
# OSHA PSM H2 TQ: 10,000 lbs (LFL 4%; autoignition 500C; 4-75% flammable).
# OSHA PSM CO TQ: 1,000 lbs. NIOSH IDLH CO: 1,200 ppm.
# 55th upward-direction attack: steam/carbon ratio 2.1 shown as 3.4.
# FIRST SMR hydrogen attack; FIRST Ni catalyst coking attack; FIRST WGS attack.
SMR_THRESHOLD = 28

class SMRContext(StrEnum):
    STEAM_TO_CARBON_RATIO   = auto()  # Molar S/C feed ratio (55th upward attack; Ni coking)
    REFORMER_TUBE_SKIN_TEMP = auto()  # Radiant-section HP-Nb tube skin temperature (IR scanner)
    HTS_INLET_TEMPERATURE   = auto()  # High-temperature WGS reactor inlet temperature

async def scan_smr_frame(
    frame_b64: str,
    context: SMRContext,
    facility_id: str,
    instrument_tag: str,
) -> dict[str, Any]:
    payload = {
        "image_b64": frame_b64,
        "context": context,
        "facility_id": facility_id,
        "instrument_tag": instrument_tag,
        "scan_ts": datetime.now(timezone.utc).isoformat(),
        "image_hash": hashlib.sha256(base64.b64decode(frame_b64)).hexdigest(),
    }
    async with httpx.AsyncClient(timeout=4.0) as client:
        r = await client.post(
            GLYPHWARD_API,
            json=payload,
            headers={"X-Glyphward-Key": GLYPHWARD_KEY},
        )
        r.raise_for_status()
        return r.json()

async def pre_scan_gate_smr(
    frame_b64: str,
    context: SMRContext,
    facility_id: str,
    instrument_tag: str,
) -> None:
    result = await scan_smr_frame(frame_b64, context, facility_id, instrument_tag)
    if result["adversarial_score"] >= SMR_THRESHOLD:
        raise AdversarialSMRImageError(
            f"Adversarial injection detected in {context} (score {result['adversarial_score']}) "
            f"at facility {facility_id} instrument {instrument_tag}. "
            "Frame withheld from AI monitoring pipeline."
        )

class AdversarialSMRImageError(RuntimeError):
    pass