Prompt injection in carbon capture post-combustion amine scrubber AI

Four adversarial injection surfaces in post-combustion amine scrubber AI

1. Absorber column differential pressure and liquid level flooding display AI (Honeywell Forge Absorber AI, Aspen GDOT PCC Optimiser, ABB Ability Genix — absorber column flooding detection AI)

The absorber column in a post-combustion capture system operates at low pressure (approximately 1–1.05 bar absolute, as the flue gas is at near-atmospheric pressure) with a large upward gas velocity through the packing or trays (superficial gas velocity approximately 1.5–3.0 m/s in structured packing, depending on the packing type and the L/G ratio). Flooding in a packed absorber column — the process upset in which the upward gas flow carries the descending liquid up the column rather than allowing it to fall countercurrently — occurs when the gas velocity exceeds the flooding velocity at the prevailing liquid rate, initiating a catastrophic liquid carryover from the absorber top into the treated gas duct, the blower or induced-draft fan downstream, and the CO₂ compressor train. Absorber flooding is detected by monitoring the column differential pressure (the pressure drop across the packed bed increases sharply as flooding approaches, rising from the design operating differential pressure of typically 3–8 mbar/m of packing to 20–40 mbar/m at incipient flooding), the column sump level (the liquid inventory in the absorber sump, measured by differential pressure level transmitters, rises as liquid backs up in the column under flooding conditions), and the treated gas moisture content (indicating liquid carryover). AI systems process rendered displays of these measurements — strip chart trend renders of column differential pressure, sump level, and inlet/outlet gas temperature and composition — to classify absorber hydraulic state: normal (within design envelope), approaching flooding (differential pressure above 70% of flooding differential — L/G reduction advisory), incipient flooding (above 90% — immediate L/G reduction), and flooding (above flooding differential — emergency gas bypass and absorber unloading).

An adversarial perturbation on a rendered absorber column differential pressure trend display image that suppresses the rising pressure differential — applying a ±8 DN downward shift to the pixel region encoding the differential pressure trend line above the approaching-flooding threshold — causes the absorber process AI to classify an absorber approaching flooding as normal hydraulic operation, suppressing the L/G reduction advisory and the flooding prevention action. The consequence of an undetected absorber flooding event: amine solvent carryover into the treated gas blower (amine fouling of impeller and housing), the direct contact cooler (if fitted upstream of the CO₂ compressor), and the CO₂ compressor first stage (amine deposition on compressor internals — requiring a shutdown for mechanical cleaning at a remediation cost of $200,000–$1M per compressor cleaning event). At Boundary Dam Unit 3 — where the absorber flooding problem was a primary cause of multiple unplanned outages in the first years of operation — the remediation required redesign of the flue gas blower and the absorber liquid distribution system, delaying the achievement of design capture capacity by more than two years. Adversarial suppression of the absorber flooding display AI removes the only automated early-warning layer before the hydraulic cascade.

2. Lean/rich amine heat exchanger fouling temperature profile display AI (Fluor Econamine FG+ process AI, BASF OASE Blue process AI, Siemens SPPA-T3000 PCC AI — lean/rich HX approach temperature AI)

The lean/rich amine heat exchanger — typically the largest heat exchanger in a post-combustion capture plant, with a heat duty of 40–80 MW on a 500 MWe-capture scale — recovers heat from the hot regenerated lean amine (approximately 100–120°C at the stripper bottom) to preheat the CO₂-loaded rich amine before the stripper. The approach temperature on the lean/rich heat exchanger — the temperature difference between the hot lean amine outlet and the cold rich amine inlet — is the primary indicator of heat exchanger fouling: as heat transfer fouling (from iron corrosion products, heat stable salt precipitates, amine degradation product deposits, and suspended particulates from the gas/liquid contactors) accumulates on the heat transfer surfaces, the approach temperature increases from the clean design value (typically 8–15°C) toward 30–50°C at severe fouling. Lean/rich heat exchanger fouling reduces the rich amine preheat temperature, which increases the steam demand in the stripper reboiler for the same CO₂ capture rate — increasing the thermal energy penalty per tonne of CO₂ captured, the parasitic power demand of the capture plant, and the reboiler steam extraction from the power plant, reducing net electrical output. More critically, severe heat exchanger fouling concentrations can produce localised hot spots in the heat exchanger channels where iron catalysts from corrosion products contact concentrated amine at elevated temperature, initiating accelerated thermal degradation of the amine solvent and heat stable salt (HSS) formation. AI process monitoring systems classify lean/rich HX performance by processing rendered approach temperature trend display images — strip chart renders of the lean outlet and rich inlet temperature channels — to identify fouling progression before it reaches the point of irreversible solvent degradation or thermal runaway in the reboiler.

An adversarial perturbation on a rendered lean/rich HX temperature profile display image that suppresses the rising approach temperature trend — applying a ±8 DN downward shift to the pixel region encoding the approach temperature trend line above the fouling alarm threshold — causes the PCC process AI to classify a progressively fouling heat exchanger as operating within specification, suppressing the cleaning schedule initiation and the solvent quality investigation that elevated approach temperature requires. The consequence is solvent thermal degradation: heat stable salts (acetate, formate, oxalate, succinate — formed by amine oxidative and thermal degradation) accumulate in the solvent to concentrations that: (1) permanently consume amine alkalinity, reducing CO₂ absorption capacity; (2) cause corrosion of carbon steel absorber shells, lean/rich HX shells, reboiler shells, and associated pipework (iron HSS corrosion mechanism); and (3) deposit as solid precipitate in the reclaimer and packing surfaces. The Petra Nova PCC plant (NRG/JX Nippon, Parish Unit 8, Texas, operational 2017–2020) experienced solvent degradation issues contributing to higher-than-projected operating costs that were among the factors in the decision to suspend operations. Adversarial suppression of the lean/rich HX temperature AI delays the detection of the fouling that precedes irreversible solvent degradation.

3. Heat stable salt concentration analyser display AI (Emerson Rosemount 5081 conductivity AI, Metrohm process IC AI, Applied Analytics online FT-IR AI — PCC amine solvent heat stable salt monitoring AI)

Heat stable salts (HSS) in amine-based CO₂ capture solvents — salts formed by the reaction of amine (a base) with strong acids (sulphate from SO₂ in flue gas, chloride from HCl, formate and acetate from amine thermal degradation, thiosulphate from SO₂/O₂ co-capture) that are thermally stable at stripper temperatures and therefore cannot be regenerated by heating — accumulate in the solvent over time as a byproduct of normal operation. Industry practice targets HSS concentrations below 10–15 wt% of total alkalinity (representing the fraction of amine permanently deactivated by strong acid formation), with process alarms at 8–10 wt% and reclaimer activation (vacuum distillation or ion exchange to remove HSS) triggered above the alarm threshold. At HSS concentrations above 15–20 wt%, the solvent pH falls below approximately 8.5 — the threshold for carbon steel corrosion by the acidified amine solution — and corrosion rates in the carbon steel absorber, lean/rich HX shells, and stripper increase sharply, generating iron corrosion products that catalyse further amine oxidative degradation in a self-accelerating cycle. Online HSS monitoring systems — conductivity analysers, in-line ion chromatographs, or FT-IR process analysers — generate rendered displays of HSS concentration trends (percentage of total alkalinity, or millimoles per litre of individual anion species) that AI systems classify to determine whether HSS is within target, approaching the reclaimer trigger threshold, or at the corrosion threshold requiring emergency reclaimer activation and solvent quality investigation.

An adversarial perturbation on a rendered HSS analyser display image that suppresses the HSS concentration trend — applying a ±10 DN downward shift to the pixel region encoding the concentration bar or trend line above the reclaimer alarm threshold — causes the PCC process AI to classify a solvent approaching or at the corrosion threshold as within-specification, suppressing the reclaimer activation and solvent quality management actions that elevated HSS concentration requires. The consequence: solvent pH falls below the carbon steel corrosion threshold; iron corrosion products accumulate in the solvent; the corrosion accelerates oxidative amine degradation (iron-catalysed oxidation of MEA consumes solvent and generates ammonia, formate, and acetate); the fouling deposits from iron corrosion products accumulate in the packed absorber and lean/rich HX tubes; and ultimately the solvent inventory requires full replacement at a cost of $2–$20M for the solvent charge alone, plus shutdown time for system cleaning. ISO 27919-1:2018 Annex A describes solvent quality monitoring requirements for PCC performance evaluation but does not specify adversarial robustness requirements for AI systems classifying rendered solvent analyser displays.

4. CO₂ product stream purity analyser display AI (Emerson X-STREAM CO2 AI, ABB EL3020 analyser AI, AMETEK process analyser AI — CO2 product quality monitoring AI for EOR and geological sequestration)

The CO₂ product stream from a post-combustion capture facility — the dehydrated, compressed CO₂ stream delivered to the CO₂ compression and transport system after leaving the stripper overhead — must meet purity specifications for its intended end use: geological sequestration (typically CO₂ > 95 mol%, H₂O < 500 ppmv, SO₂ < 200 ppmv, H₂S < 200 ppmv, O₂ < 4 mol%, depending on the injection well and aquifer specification); enhanced oil recovery (EOR, typically CO₂ > 95 mol%, H₂S and SO₂ limits set by pipeline specification); or industrial use (food-grade, beverage-grade, or chemical-grade specifications with higher purity and lower impurity limits). Real-time CO₂ product purity is monitored by online analysers — non-dispersive infrared (NDIR) analysers for CO₂ concentration, electrochemical or UV fluorescence analysers for SO₂ and H₂S, paramagnetic analysers for O₂, and gas chromatographs for the complete stream composition — generating rendered displays of multi-component gas stream composition that AI systems classify to determine whether product meets specification, whether an off-spec event is developing (requiring diversion of the CO₂ product to atmospheric vent rather than the sequestration pipeline), or whether an emergency shutdown of the compressor train is required. For CO₂ EOR applications, off-spec CO₂ containing elevated H₂S above pipeline specification creates a regulatory obligation under the Gas Pipeline Safety Regulations and PHMSA 49 CFR Part 195 (pipeline safety for hazardous liquid pipelines).

An adversarial perturbation on a rendered CO₂ product purity analyser display image that suppresses an impurity concentration above specification — applying a ±8 DN downward shift to the pixel region encoding the SO₂, O₂, or H₂S concentration bar above the specification limit — causes the CO₂ product quality AI to classify an off-spec product stream as on-spec, suppressing the product diversion to vent and the compressor isolation that an off-spec classification requires. The consequence: off-spec CO₂ is injected into the geological sequestration reservoir or EOR pipeline. For geological sequestration, elevated SO₂ co-injection forms sulphurous acid (H„SO₃) and sulphuric acid (H₂SO₄) in the formation water near the injection well, which can dissolve formation minerals and alter well injectivity — a problem documented at multiple CO₂ injection demonstration sites (Weyburn-Midale CO₂ monitoring project, Otway Basin Pilot Project). EPA 40 CFR Part 60 Subpart TTTT requires monitoring of CO₂ stream composition for GHG reporting at electric generating units with CCS — adversarial suppression of the purity analyser AI creates a GHG reporting obligation discrepancy and a potential Clean Air Act Section 114 enforcement exposure. Free tier — 10 scans/day, no card required.

Integration: post-combustion amine scrubber AI scanning with Glyphward pre-scan gate

The Glyphward scan gate for post-combustion amine scrubber AI belongs at every rendered-image ingestion boundary in the PCC process monitoring pipeline — before absorber flooding level AI processes rendered differential pressure and sump level trend displays, before lean/rich HX temperature profile AI processes rendered approach temperature displays, before HSS concentration AI processes rendered online analyser outputs, and before CO₂ product purity AI processes rendered multi-component analyser displays. Threshold 40 reflects the process upset and regulatory exceedance consequence: absorber flooding destroys compressor internals ($1M+ remediation); HSS accumulation causes irreversible solvent degradation ($20M+ solvent replacement); CO₂ off-spec delivery creates EPA 40 CFR Part 60 Subpart TTTT reporting violations. The false positive cost is zero — a flagged scan triggers a manual instrument cross-check against DCS-held process data, which takes minutes and has no process consequence.

import asyncio, base64, hashlib, json
from datetime import datetime, timezone
from enum import Enum
from pathlib import Path

import httpx

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

# Post-combustion amine scrubber AI contexts: threshold 40
# ISO 27919-1:2018 (PCC performance evaluation methods);
# EPA 40 CFR Part 60 Subpart TTTT (GHG reporting for CCS at EGUs);
# NETL DOE PCC performance baseline reports.
AMINE_SCRUBBER_THRESHOLD = 40


class AmineScrubberAIContext(Enum):
    ABSORBER_FLOODING    = "absorber_flooding"    # Column DP / sump level AI
    LEAN_RICH_HX         = "lean_rich_hx"         # Approach temperature profile AI
    HEAT_STABLE_SALT     = "heat_stable_salt"     # HSS concentration analyser AI
    CO2_PURITY           = "co2_purity"           # CO2 product stream purity AI


class AdversarialAmineScrubberImageError(Exception):
    """Raised when Glyphward detects adversarial content in a post-combustion
    amine scrubber AI rendered process monitoring image above threshold 40.

    Consequence if not raised:
    - ABSORBER_FLOODING: flooding suppressed → amine carryover into CO2
      compressor train → impeller fouling → compressor trip; $1M+ cleaning.
    - LEAN_RICH_HX: fouling suppressed → reboiler temperature exceedance
      → amine thermal degradation → irreversible HSS accumulation → solvent
      replacement; Boundary Dam Unit 3 operational precedent.
    - HEAT_STABLE_SALT: HSS suppressed → solvent pH below carbon steel
      corrosion threshold → iron corrosion product accumulation → accelerated
      oxidative degradation cycle → $20M+ solvent charge replacement.
    - CO2_PURITY: off-spec CO2 stream suppressed → off-spec product injected
      → geological sequestration reservoir impurity; EPA 40 CFR Part 60
      Subpart TTTT reporting discrepancy; PHMSA pipeline spec violation.
    Fail-safe: halt AI classification; cross-check against DCS-held
    process historian data before resuming AI-driven optimisation.
    """

    def __init__(self, scan_id: str, score: int,
                 context: AmineScrubberAIContext,
                 plant_id: str, unit_id: str,
                 flagged_region: dict | None = None) -> None:
        self.scan_id = scan_id
        self.score = score
        self.context = context
        self.plant_id = plant_id
        self.unit_id = unit_id
        self.flagged_region = flagged_region
        super().__init__(
            f"Adversarial amine scrubber image: "
            f"context={context.value} score={score} "
            f"plant={plant_id} unit={unit_id} scan_id={scan_id}"
        )


async def scan_amine_scrubber_image(
    image_bytes: bytes,
    context: AmineScrubberAIContext,
    plant_id: str,
    unit_id: str,
    client: httpx.AsyncClient,
) -> dict:
    """Scan a post-combustion amine scrubber AI rendered process monitoring
    image for adversarial content.

    Fail-safe contract: AdversarialAmineScrubberImageError or httpx error →
    halt AI process classification for the affected monitoring context;
    cross-check against DCS process historian data before resuming
    AI-driven absorber optimisation or product diversion decisions.
    """
    image_hash = hashlib.sha256(image_bytes).hexdigest()
    payload = {
        "image": base64.b64encode(image_bytes).decode(),
        "source": f"amine_scrubber:{context.value}:{plant_id}:{unit_id}",
        "metadata": {
            "plant_id": plant_id,
            "unit_id": unit_id,
            "context": context.value,
            "image_sha256": image_hash,
        },
    }
    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["score"] > AMINE_SCRUBBER_THRESHOLD:
        raise AdversarialAmineScrubberImageError(
            scan_id=result["scan_id"],
            score=result["score"],
            context=context,
            plant_id=plant_id,
            unit_id=unit_id,
            flagged_region=result.get("flagged_region"),
        )
    return result

Deploy scan_amine_scrubber_image at each PCC process monitoring AI rendered-image ingestion boundary: before absorber flooding AI (threshold 40), before lean/rich HX temperature profile AI (threshold 40), before HSS concentration analyser AI (threshold 40), and before CO₂ product purity AI (threshold 40). On AdversarialAmineScrubberImageError for ABSORBER_FLOODING context: immediately verify absorber column differential pressure and sump level against DCS process historian tag data and cross-check with local field gauge readings before resuming AI-driven L/G optimisation. See also: CO2 geological sequestration injection well AI prompt injection (downstream adversarial surface) and chemical plant process safety AI prompt injection (related process safety AI context). Get early access

Related questions