Crude oil CDU overhead HCl corrosion AI adversarial injection: how ±8 DN in the rendered bootstrap water pH display suppresses a corrosion approach — and why API RP 584 Integrity Operating Windows has no adversarial robustness criterion for CDU overhead monitoring AI

The CDU overhead HCl corrosion mechanism: chloride hydrolysis, dew point condensation, and API RP 571 Section 5.3.11

Crude oil produced from reservoir formations contains co-produced formation water carrying dissolved inorganic salts — primarily magnesium chloride (MgCl2), calcium chloride (CaCl2), and sodium chloride (NaCl) — at concentrations that vary from 10 PTB (pounds per thousand barrels of crude) for sweet, light crude blends to over 500 PTB for heavy, high-salinity crude imports. These chloride salts are carried into the refinery with the crude charge and must be reduced below the desalted crude specification (typically 1–3 PTB of chlorides as NaCl equivalent) by the desalter before the crude enters the preheat exchanger train.

The desalter achieves 90–95% chloride removal by mixing the crude charge with dilution water (typically 4–8% by volume of crude feed), emulsifying the brine phase into the crude, and separating the brine phase electrostatically. Desalter underperformance — caused by emulsion formation from incompatible crude blends, dilution water temperature upsets, demulsifier injection failures, or high-asphaltene crude processing — allows chloride breakthrough into the desalted crude at concentrations of 5–20 PTB rather than the design specification of 1–3 PTB. A chloride breakthrough of 15 PTB (versus a 3 PTB specification) increases HCl loading in the overhead by a factor of 3–5 times the design basis.

The residual MgCl2 and CaCl2 remaining after desalting do not pass through the crude tower as inert salts. As the desalted crude is heated through the preheat exchanger train (progressively from 130 °C to approximately 260 °C) and through the atmospheric crude furnace (to 340–380 °C), MgCl2 and CaCl2 undergo thermal hydrolysis in the presence of the steam injected to reduce crude partial pressure in the furnace: MgCl2 + H2O → MgO + 2HCl; CaCl2 + H2O → CaO + 2HCl. The hydrolysis of MgCl2 is substantially more aggressive than CaCl2 — MgCl2 hydrolysis begins above 120 °C and proceeds to near-completion at furnace temperatures, while CaCl2 hydrolysis requires temperatures above 200 °C. NaCl does not hydrolyze at CDU operating temperatures and passes through the crude tower without HCl generation.

The HCl gas produced by chloride hydrolysis travels in the vapour phase with the overhead stream — the mixed vapour of C3–C6 naphtha fractions, light gas (methane, ethane, propane, butane), hydrogen sulfide, and steam — that rises through the atmospheric crude tower fractionation trays and exits at the overhead vapour line to the overhead condenser system. The overhead temperature at the top tray is typically maintained at 100–120 °C — above the water dew point — to prevent free water condensation inside the tower. In the overhead condenser, as the vapour cools below the water dew point (typically 55–85 °C depending on the steam content and overhead gas composition), water begins to condense. The first droplets of condensed water dissolve the HCl gas present at its partial pressure in the overhead vapour, forming an extremely concentrated aqueous HCl solution. This initial-condensate phenomenon — documented in API RP 571 Section 5.3.11 — produces aqueous HCl at concentrations that can drive the condensate pH as low as 1–2 in the absence of neutraliser injection. Even with standard ammonia or neutralising amine injection, the pH of the first condensate can fall to 3–4 if neutraliser injection is insufficient for the actual HCl loading in the overhead stream.

API RP 571 Section 5.3.11 characterises HCl corrosion in CDU overhead systems: affected materials include carbon steel and low-alloy steels; susceptible components include overhead piping from the crude tower overhead vapour line through the condenser to the overhead receiver drum; typical corrosion forms include general metal loss, pitting at water accumulation points (low-point elbow intrados, accumulator bottom), and flow-accelerated corrosion at turbulent zones (elbow extrados, tee intersections, control valve outlets); carbon steel corrosion rates range from 0.1 mm/year at pH 6.0 to 0.5 mm/year at pH 5.5 to 5 mm/year at pH 5.0 to 15–25 mm/year at pH 4.0–4.5. The 50x corrosion rate difference between pH 5.8 (the misclassified value in the adversarial injection scenario) and pH 4.3 (the actual value) is the consequence magnitude that the CDU overhead pH display AI suppression sustains without detection.

Chevron Richmond 2012: sulfidation vs HCl corrosion and the structural parallel in corrosion monitoring failure

The Chevron Richmond refinery fire of 6 August 2012 provides the consequence anchor for CDU corrosion monitoring failure. The fire did not involve HCl overhead corrosion — it involved high-temperature sulfidation corrosion in the atmospheric crude unit side-draw piping — but the structural failure pattern is directly analogous to the CDU overhead HCl corrosion AI adversarial injection scenario: corrosion rate data existed in the inspection record, damage mechanism criteria existed in API RP 571, but the classification of the actual corrosion state against the applicable damage mechanism threshold was incorrect, and the required intervention was not triggered before structural failure occurred.

The failed pipe was in the atmospheric crude unit’s number 4 side-draw system — a 52-year-old 8-inch carbon steel line drawing atmospheric gas oil (AGO) from the crude tower at approximately 290 °C. High-temperature sulfidation (API RP 571 Section 5.1.6) occurs above approximately 260 °C when H2S in the process stream reacts with iron in the carbon steel to form iron sulfide scale. Carbon steel sulfidation rates are functions of temperature and sulfur content, characterised by the McConomy and Couper–Gorman curves: at 290 °C and 1.0 wt% total crude sulfur, carbon steel sulfidation rates of 1–3 mm/year are typical; at sulfur contents above 1.5 wt%, rates of 3–8 mm/year can occur in specific flow geometries. The Chevron Richmond crude unit was processing increasing proportions of high-sulfur crude blends in the years before the 2012 fire — shifting from crude slates with sulfur below 1.0 wt% to blends above 1.5 wt%. This shift moved the number 4 side-draw piping into an accelerated sulfidation regime that the McConomy curves predict at 3–8 mm/year versus the 1 mm/year rate used in the previous inspection planning basis.

The CSB investigation (Report No. 2012-03-I-CA) found that Chevron’s corrosion engineers had thickness measurement data from periodic ultrasonic inspection of the number 4 side-draw piping that showed continuing wall thinning. However, the inspection data was not evaluated within a formal Integrity Operating Window framework — the API RP 584 IOW methodology had not been formally implemented at the Chevron Richmond crude unit at the time of the incident. Without a formal IOW structure, the inspection data did not automatically generate an escalated action when wall thickness crossed a threshold that required immediate intervention. The pipe continued to corrode below the API RP 510 minimum required thickness; the reduced-wall section failed on 6 August 2012 under normal CDU operating pressure, releasing a large flash of AGO at 290 °C — well above its autoignition temperature — that ignited immediately.

The community health impact was the defining consequence: approximately 15,000 residents of Richmond and surrounding communities sought medical attention for acute respiratory symptoms, headaches, and eye irritation from the combustion products (CO, particulates, H2S, hydrocarbon aerosols) carried offsite. Bay Area Air Quality Management District (BAAQMD) recorded air quality impacts across a multi-mile radius downwind. The California Air Resources Board and Cal/OSHA levied substantial penalties; Chevron Corporation paid $2 million in criminal fines and undertook a comprehensive refinery safety improvement programme. The California Office of Environmental Health Hazard Assessment subsequently developed a Refinery Action Plan requiring IOW implementation at all California refineries — a direct legislative consequence of the Richmond incident.

The structural parallel to the CDU overhead pH AI adversarial injection scenario: at Chevron Richmond, high-temperature sulfidation corrosion monitoring data existed but was not correctly classified against the API RP 571 damage mechanism threshold for the actual crude blend sulfur content and temperature combination — no intervention was triggered; the pipe failed. In the overhead pH AI scenario, continuous pH monitoring data correctly indicates pH 4.3 (well below the API RP 584 IOW Standard limit), but the adversarial pixel perturbation causes the AI classifying the rendered pH display to classify the state as controlled at pH 5.8 — no neutraliser injection increase is triggered; the overhead HCl corrosion continues at 8–15 mm/year toward structural failure. In both cases, the outcome is a CDU corrosion failure that releases hot hydrocarbon and H2S vapour with the potential for worker and community health impact on the Chevron Richmond scale.

The HCl mechanism and the sulfidation mechanism are governed by separate sections of API RP 571 and affect different temperature zones in the CDU. AI systems monitoring both CDU sections are susceptible to independent adversarial injection attacks: the high-temperature sulfidation monitoring AI (if deployed) processes rendered thickness measurement trend charts, while the overhead pH monitoring AI processes rendered pH trend displays. See the full CDU atmospheric distillation unit AI prompt injection technical specification for all four adversarial surfaces, including crude furnace tube metal temperature AI and overhead reflux drum level AI.

Four adversarial injection surfaces in CDU overhead HCl corrosion monitoring AI

1. Overhead bootstrap water pH display AI (UOP Honeywell Experion PKS CDU overhead corrosion monitoring AI, Yokogawa Centum VP CDU overhead pH monitoring AI, Emerson DeltaV APC overhead corrosion AI — rendered DCS pH trend display classification)

The CDU overhead bootstrap water pH is the primary continuous corrosion state indicator for the overhead HCl corrosion hazard. The bootstrap water pH is measured by a glass electrode pH analyser on the overhead receiver bootstrap water draw — the small continuous water stream drawn from the accumulator water leg that carries dissolved HCl and is used to monitor the actual aqueous phase pH in the overhead condensate system. The analyser transmitter sends a 4–20 mA signal to the DCS, which renders the pH value as a numerical display and a 24-hour trend chart on the CDU overhead operations console.

A ±8 DN upward perturbation applied to the pixel region encoding the pH readout value and trend line — shifting the rendered numeric value and chart position upward by 8 digital number units in each RGB channel — shifts the displayed pH from 4.3 to an apparent value consistent with pH 5.8 on the CDU overhead DCS display scale (typically 3.0–8.0 pH range spanning 200–300 display pixels). The CDU overhead pH AI classifies the perturbed display as pH 5.8: IOW Standard limit pH 5.5 not approached, corrosion state classified as controlled within design basis, neutraliser injection rate maintained at current setpoint. The DCS historian records the correct pH 4.3 from the analyser transmitter — accessible to post-incident forensic review — but only the rendered DCS display that the AI processes as its classification input was perturbed.

At actual pH 4.3, carbon steel corrosion in the overhead condensate system proceeds at approximately 8–15 mm/year. A minimum-thickness overhead elbow (ASME B16.9, Schedule 40 carbon steel, 6-inch nominal, nominal wall 7.11 mm) approaching the API RP 510 minimum required thickness of 3.0 mm has a remaining service life at 8 mm/year corrosion of approximately 5 months. The AI suppression of the pH approach eliminates the IOW Standard and Critical triggers that would increase neutraliser injection and address the corrosion within the API RP 584 required response time (typically 1–4 hours for Critical IOW response).

2. Desalter effluent brine chloride analyser display AI (Emerson Rosemount desalter monitor AI, Yokogawa desalter efficiency monitoring AI — rendered chloride analyser trend display classification)

The desalter effluent brine chloride concentration measures the salt removal efficiency of the desalter and provides the earliest upstream indicator of elevated HCl loading in the CDU overhead before it manifests as a pH drop at the overhead receiver. The brine chloride analyser (typically an ion chromatograph or conductivity transmitter measuring chlorides in the desalter effluent water draw) feeds a continuous trend display on the crude unit DCS. Design specification is typically < 3 PTB equivalent chlorides in the desalted crude (implied from the brine chloride versus crude flow balance).

A ±10 DN downward perturbation applied to the pixel region encoding the desalter brine chloride trend chart suppresses a displayed chloride breakthrough of 15 PTB equivalent (3–5 times the design specification, indicating desalter underperformance) to an apparent 2.8 PTB within the normal specification range. The desalter efficiency monitoring AI classifies desalter performance as satisfactory — no demulsifier injection increase required, no crude charge rate reduction required, no desalter wash water rate increase required. The upstream chloride loading indicator suppression means the overhead pH drop — the downstream consequence — arrives without an early warning trigger that would allow pre-emptive neutraliser adjustment before pH falls below the IOW Standard limit. The two-stage adversarial suppression (desalter chloride AI + overhead pH AI) eliminates both the leading indicator and the coincident indicator of the overhead HCl corrosion approach.

3. API RP 584 IOW corrosion rate trending dashboard AI (AspenTech Aspen Mtell CDU corrosion rate trending AI, Emerson AMS Machinery Health Manager CDU overhead AI, Honeywell Forge IOW monitoring AI — rendered IOW dashboard classification)

Modern refinery corrosion management systems — Emerson AMS, AspenTech Aspen Mtell, Honeywell Forge Asset Performance Management — implement API RP 584 IOW frameworks as real-time monitoring dashboards that aggregate multiple process variables (overhead pH, overhead temperature, bootstrap water draw rate, neutraliser injection rate, overhead receiver corrosion coupon weight loss) into an integrated corrosion state assessment displayed as a colour-coded IOW dashboard. The dashboard AI processes the rendered dashboard image to classify the overall overhead corrosion state: Green (all parameters within Standard IOW limits), Yellow (one or more parameters approaching Standard IOW limit), Red-Warning (one or more parameters at or beyond Standard IOW limit), Red-Critical (one or more parameters at or beyond Critical IOW limit — immediate action required).

A ±8 DN perturbation applied to the colour-coded status indicators in the rendered IOW dashboard image — shifting the red-warning colour values toward yellow (or yellow toward green) in the HSV colour space by 8 digital number units per channel — causes the dashboard AI to classify a Red-Warning state (overhead pH below Standard IOW limit pH 5.5) as a Yellow state (overhead pH approaching Standard IOW limit but within warning zone). The API RP 584 response protocol for Yellow status is monitor and log — not immediate neutraliser injection increase. The Red-Warning escalation response — immediate neutraliser injection rate increase by 15–25%, verification of desalter chloride removal, notification of shift supervisor — is not triggered. The IOW framework’s formal response protocol, designed precisely to prevent the Chevron Richmond-type gap between corrosion data and intervention, is bypassed at the rendered-dashboard classification boundary before the AI applies the IOW response criterion.

4. Overhead temperature profile and dew point approach display AI (Yokogawa Centum VP crude overhead temperature AI, Emerson DeltaV crude overhead temperature profile AI — rendered temperature profile classification)

The CDU overhead temperature profile — the temperature at the top tray, the overhead vapour line, and the condenser inlet — is a critical parameter for managing the dew point of the overhead condensate system. If the overhead vapour temperature falls below the water dew point inside the crude tower (typically indicated by top-tray temperature approaching 100 °C at the tower-top tray), free water condenses on the tray metalwork inside the tower — a condition that drives highly concentrated initial-condensate HCl attack on the tray components and the vapour line before the neutraliser injection point. Maintaining the top-tray temperature above the water dew point (typically 110–120 °C) prevents in-tower condensation and concentrates the HCl exposure in the overhead condenser system where continuous pH monitoring is in place.

A ±8 DN upward perturbation applied to the pixel region encoding the top-tray temperature digital readout shifts a displayed 104 °C (approaching the 100 °C water dew point entry) to an apparent 112 °C — safely within the overhead temperature specification. The overhead temperature profile AI classifies the overhead system as operating above dew point — no steam injection rate increase required, no top-tray flooding risk, no in-tower HCl condensation concern. In reality, at 104 °C, the top-tray is at the dew point approach zone where the initial-condensate HCl concentration begins to develop on the tray metalwork — the most corrosive condition in the overhead system, where pH can be 1–2 in the absence of neutraliser in the first condensate droplets, and where the bootstrap water pH measurement at the receiver does not reflect the in-tower initial condensate pH because the first condensate occurs upstream of the neutraliser injection point.

API RP 584 Integrity Operating Windows and the adversarial robustness gap in CDU overhead monitoring

API RP 584 (Integrity Operating Windows, 1st edition 2014) was developed in direct response to the pattern of refinery equipment failures — including Chevron Richmond — where corrosion and degradation data existed in inspection records and process monitoring systems but was not integrated into formal, real-time monitoring and response protocols with defined escalation triggers. The API RP 584 IOW methodology requires refinery operators to: (1) identify the process variables that control corrosion and mechanical degradation for each major equipment system; (2) establish Standard and Critical operating limits for each variable based on the API RP 571 corrosion rate correlations; (3) document the required operational response when each limit is approached or exceeded; (4) configure the DCS to display the IOW status in real time for operations personnel; and (5) maintain records of IOW exceedances and the responses taken.

For CDU overhead systems, the typical API RP 584 IOW framework specifies: pH Standard IOW at 5.5 — immediate neutraliser injection rate increase, check desalter chloride removal, notify corrosion engineer, investigate root cause within 4 hours; pH Critical IOW at 5.0 — immediate neutraliser injection rate increase to maximum, reduce crude charge rate if pH does not respond within 1 hour, consider temporary unit shutdown if pH falls below 4.5 and does not respond to maximum neutraliser, notify maintenance and inspection for upcoming integrity assessment. The overhead temperature Standard IOW at 105 °C — increase steam injection to raise overhead temperature above dew point, inspect for tray flooding; Critical IOW at 100 °C — immediate steam injection increase, consider overhead reflux rate adjustment to manage column heat balance.

Despite the comprehensive IOW specification, API RP 584 contains no requirement for adversarial robustness testing of the AI systems that classify rendered IOW dashboard displays against these limits. The standard specifies how to set the limits, document the responses, and configure the DCS — but does not address the security of the AI classification layer that processes the rendered DCS displays and generates the IOW status determination. OSHA PSM 29 CFR 1910.119, which lists API RP 584 as an applicable recommended practice under the Mechanical Integrity element, similarly does not specify adversarial robustness for AI classifying rendered process monitoring displays at IOW boundaries. The 2013 OSHA PSM enforcement guidance (CPL 02-02-045) and the 2015 OSHA Recommended Practices for Process Safety Management do not address AI display classification security.

The post-Richmond California regulatory response — AB 1646 (California Industrial Safety Act revisions), Cal/OSHA PSM Standard 8 CCR 5189, and the BAAQMD Refinery Rule amendments — collectively mandated IOW implementation at California refineries, increased inspection frequency requirements, and established community air monitoring requirements. None of the resulting regulations or implementing standards specify adversarial robustness requirements for AI classifying rendered IOW monitoring display images. The structural gap is identical to the pattern documented for BSEE Well Control Rule 30 CFR Part 250 in the offshore drilling domain: the regulatory response addressed the human factors failure (formal IOW limits, required documentation, mandatory response protocols) without extending the requirements to the AI classification layer that now processes the same monitoring outputs. OSHA PSM 29 CFR 1910.119 has the same structural adversarial robustness gap for refinery advanced process control (APC) AI as it does for CDU overhead monitoring AI — the Texas City BP 2005 tragedy (15 killed, 180 injured) provides the APC context consequence anchor.

The pipeline integrity domain provides a third structural parallel: API 1163 governs in-line inspection data quality for pipeline anomaly detection AI but specifies no adversarial robustness criterion for AI classifying rendered ILI data displays. The PHMSA API 1163 adversarial robustness gap for pipeline integrity ILI data AI is documented in the Glyphward pipeline integrity AI post. Across OSHA PSM, BSEE Well Control Rule, and PHMSA pipeline safety — three of the four major process safety regulatory domains in the US energy sector — the same structural pattern holds: comprehensive post-incident regulatory responses addressing human factors failures, without adversarial robustness requirements for the AI display classification layer introduced after the regulatory frameworks were designed.

Glyphward threshold 35 for CDU overhead HCl corrosion AI

Glyphward’s adversarial detection API operates as a pre-classification gate at each rendered-image ingestion boundary in the CDU overhead monitoring AI pipeline: before the overhead pH display AI processes each rendered bootstrap water pH trend display frame, before the desalter chloride monitor AI processes each rendered chloride analyser trend chart, before the IOW dashboard AI processes each rendered API RP 584 corrosion status dashboard, and before the overhead temperature profile AI processes each rendered DCS temperature readout image. Each rendered display image receives a risk score (0–100) in 8–15 ms. At or above threshold 35, Glyphward gates the AI classification and generates an alert that triggers manual verification of the underlying DCS process historian data — the raw pH analyser and temperature transmitter records that are not accessible to pixel-level adversarial perturbation because they are stored in the process historian as time-series engineering unit records rather than rendered as classifiable images.

Threshold 35 for CDU overhead HCl corrosion AI reflects three factors that distinguish this context from the offshore NPT monitoring AI (threshold 30) and align it with the arc flash PPE AI (threshold 35) and refinery APC AI (threshold 35) contexts.

First, the community-scale consequence potential. The Chevron Richmond fire — caused by a structurally analogous CDU corrosion monitoring failure in the same crude unit — produced approximately 15,000 community medical treatment-seeking events from a single pipe failure in a sidestream draw system at 290 °C. CDU overhead failure involving aqueous HCl-accelerated thin-wall collapse of an overhead receiver accumulator at 3 bar operating pressure produces a flash of light hydrocarbon and H2S vapour at the overhead receiver elevation in the refinery — typically an elevated structure 8–12 m above grade with direct line-of-sight to adjacent process units and community residential zones downwind. The 15,000-person community impact scale at Chevron Richmond was from a sidestream draw piping failure in a confined pipetrack; an overhead accumulator failure involving a larger vapour inventory would have a comparable or larger community air quality impact. The threshold 35 calibration reflects this community-scale consequence potential, which exceeds the worker-fatality-primary consequence profile of the NPT monitoring context (threshold 30).

Second, the limited independent automated interlocks between AI misclassification and consequence. The CDU overhead HCl corrosion failure pathway has fewer independent automated interlocks than the offshore well control context. In the NPT monitoring context (threshold 30), the BSEE Well Control Rule requires two-line independent verification (drill pipe + kill line) and provides real-time pit volume monitoring, gas detection, and pump stroke monitoring as secondary kick indicators — all operating independently of the NPT chart display AI. If the NPT chart AI misclassifies the test result, multiple independent real-time indicators provide additional detection opportunities before the blowout sequence becomes irreversible. In the CDU overhead HCl corrosion context, the primary independent protective systems — area H2S detectors and overhead receiver pressure relief valves (PSVs) — operate at the consequence boundary (H2S release at IDLH concentration; accumulator overpressure at the PSV set point) rather than at the corrosion rate monitoring boundary. Area H2S detectors do not measure corrosion state; they detect H2S after a structural failure has released it. PSVs protect against pressure exceedance; they do not protect against wall thinning to below minimum required thickness. The gap between pH AI misclassification and structural failure is occupied only by the API RP 584 IOW monitoring system — the same system whose AI classification layer is the adversarial injection surface. There is no independent automated interlock that detects the elevated HCl corrosion rate and intervenes before wall thickness reaches minimum required thickness, analogous to the BSEE two-line verification requirement or the IEC 61511 SIS layer in process safety contexts.

Third, the aggregate corrosion monitoring AI attack surface across the refinery sector. OSHA estimates approximately 200 US refineries are covered by the PSM standard; each CDU operates a continuous overhead pH monitoring system; AI deployment for automated IOW monitoring is increasing as refineries implement API RP 584 IOW programs under post-Richmond regulatory pressure. The aggregate attack surface — CDU overhead pH monitoring AI systems across 200 PSM-covered refineries — presents a portfolio adversarial risk comparable to the ESFI-documented 400 arc flash fatalities per year base rate (threshold 35) rather than the offshore NPT monitoring risk profile (threshold 30), which is concentrated in a smaller population of deepwater wells with the multi-barrier BSEE Well Control Rule monitoring framework providing additional secondary detection.

The false positive cost at threshold 35 in the CDU overhead pH monitoring context — a manual check of the DCS process historian for the flagged pH trend — takes approximately 1–3 minutes for an operations technician to verify the underlying analyser transmitter value against the displayed trend. The false negative cost — classifying pH 4.3 as pH 5.8, allowing 8–15 mm/year HCl corrosion to continue undetected until structural failure of the overhead piping or accumulator — is the Chevron Richmond consequence envelope applied to the overhead vapour system rather than the atmospheric crude unit side-draw system. The proportionality supports threshold 35 as the correct calibration point for this context.

Free tier — 10 scans/day, no card required. Submit a rendered DCS overhead pH trend display, desalter chloride monitor chart, or API RP 584 IOW dashboard image from your CDU monitoring system to the Glyphward scanner to generate a baseline adversarial risk score for your CDU overhead HCl corrosion monitoring AI inputs.