Trichlorosilane (SiHCl3) Siemens CVD polysilicon process AI adversarial injection: how a CVD reactor temperature upward attack (750°C displayed as 1,100°C) converts crystalline silicon rod to pyrophoric fine silicon powder — EPA RMP TQ 2,500 lbs vs OSHA PSM TQ 5,000 lbs dual regulatory gap, HCl generation from TCS pyrophoric hydrolysis, and the 22nd upward-direction attack in the Glyphward industrial AI portfolio

TCS chemistry, Siemens CVD polysilicon production, and the OSHA PSM / EPA RMP threshold quantity calibration

Trichlorosilane occupies an unusual position in the industrial hazardous chemical landscape: it is simultaneously flammable (flash point −28°C, NFPA Class IB; LEL 7.0% in air) and pyrophoric in moist air — two hazard mechanisms distinct in their chemistry but coexisting at every TCS facility. The flash point of −28°C means TCS is immediately flammable at all practical ambient temperatures: the liquid surface generates sufficient vapor for an ignitable mixture with air at every temperature above −28°C, requiring only an external ignition source to combust. TCS vapor combustion (Si–H and Si–Cl bond oxidation) is a conventional flammable gas fire hazard subject to NFPA 30 ignition-source control requirements. The pyrophoricity is a distinct mechanism: TCS reacts with moisture (SiHCl3 + 2H2O → SiO2 + 3HCl + H2; ΔH approximately −310 kJ/mol), and at high hydrolysis rates — when TCS contacts ambient humidity at any atmospheric interface — the exothermic hydrolysis energy and locally elevated H2 concentration from the reaction product can cause spontaneous ignition without external spark or flame. In dry storage under nitrogen blanket, TCS is a conventional flammable liquid; at any interface with ambient humidity (valve packing connections, sampler access ports, vent piping, drain valve seats) TCS can ignite spontaneously via moisture contact. This dual hazard profile generates two independent sentinel signals for TCS monitoring AI: the flammable vapor pathway (TCS area detector reading TCS concentration directly) and the toxic HCl pathway (area HCl detector reading HCl fumes generated by TCS hydrolysis at 3 mol HCl per mol TCS).

The Siemens CVD process is the dominant polysilicon manufacturing route for both the photovoltaic and semiconductor industries, representing approximately 80% of global polysilicon capacity. In the Siemens process, TCS and UHP H2 (≤0.1 ppm H2O) are fed into bell-jar CVD reactors containing pairs of U-shaped silicon seed rods. The rods are heated to 1,050–1,150°C surface temperature by direct resistive current (initial voltage 2,000–4,000 V DC applied to rod tips; rod resistance decreases as cross-section grows from 6 mm to 120–150 mm over 60–100-hour batch cycle). At the hot rod surface, TCS decomposes with H2 to deposit silicon: SiHCl3 + H2 → Si + 3HCl (ΔH approximately −72 kJ/mol at 1,100°C). The 1,050–1,150°C window is critical for two independent reasons. Above 1,150°C, silicon re-evaporation from localised hot spots at electrode contact zones creates porous, inhomogeneous “popcorn silicon” rod structure that fails resistivity and mechanical handling specifications. Below 900°C, surface kinetics slow below the gas-phase nucleation rate: TCS molecules react with H2 in the reactor atmosphere before reaching the rod surface, producing amorphous fine silicon powder (particle diameter <1 μm; specific surface area >100 m²/g) that settles on the reactor floor, walls, and product rod surfaces. Fine Si powder at this particle size is pyrophoric — a hazard class entirely absent from normal Siemens CVD operation, created only when temperature drops below the minimum deposition threshold. NFPA 654 (Standard for the Prevention of Fire and Dust Explosions from the Manufacturing, Processing, and Handling of Combustible Particulate Solids) applies to fine Si powder accumulations in the CVD reactor building. Major producers operating Siemens CVD facilities include GCL-Poly (China; ~100,000 MT/yr), Daqo New Energy (China; ~80,000 MT/yr), Wacker Chemie (Germany; Burghausen and Nünchritz; ~80,000 MT/yr), REC Silicon (Norway/USA; ~8,000 MT/yr), OCI Company (Malaysia; ~20,000 MT/yr), and Tokuyama Corporation (Japan; ~9,000 MT/yr).

TCS’s acute hazard standards reflect both its HCl hydrolysis product toxicity and its pyrophoric fire risk. On contact with ambient moisture, TCS generates HCl fumes (SiHCl3 + 2H2O → SiO2 + 3HCl + H2; net: 3 mol HCl per mol TCS; approximately 0.81 kg HCl per kg TCS released) visible as white aerosol in ambient air. HCl is the sentinel molecule for TCS monitoring: ACGIH TLV-C ceiling 0.5 ppm (no transient exceedance permitted); OSHA PEL 5 ppm (ceiling; 29 CFR 1910.1000 Table Z-1); NIOSH IDLH 50 ppm. TCS vapor density of 4.67 — more than 4.5× air density — causes TCS vapour to settle in below-grade condensate pits, pump cellars, trench cable ducts, and below-floor ventilation channels, requiring sensor placement below 0.3 m height in all sub-grade spaces at TCS facilities. OSHA PSM 29 CFR 1910.119 Appendix A lists trichlorosilane at threshold quantity 5,000 lbs. EPA RMP 40 CFR Part 68 Appendix A lists trichlorosilane at threshold quantity 2,500 lbs — half the OSHA PSM TQ — making TCS the only chemical in the Glyphward industrial AI portfolio where the EPA community-consequence threshold is lower than the OSHA on-site worker protection threshold. AI monitoring of TCS area HCl detectors, Siemens CVD reactor temperature, TCS/STC molar ratio process GCs, and H2 carrier gas moisture analyzers is deployed at polysilicon production facilities running Honeywell Experion PKS, Emerson DeltaV, Siemens SIMATIC S7, and Yokogawa Centum VP DCS platforms — each rendered instrument display presenting a distinct adversarial injection surface classified at Glyphward threshold 35.

Four adversarial injection surfaces in trichlorosilane Siemens CVD polysilicon process AI

1. TCS area HCl detector AI (Dräger Polytron 8700 HCl area monitor AI / MSA Ultima XE HCl electrochemical detector AI / Honeywell Analytics MIDAS-E HCl sensor AI / Industrial Scientific GX-6000 HCl detector AI / Analytical Technology ATI A14 HCl area sensor AI — monitoring hydrogen chloride vapor generated by TCS hydrolysis on moisture contact in TCS storage areas, hydrochlorination reactor buildings, distillation columns, and Siemens CVD reactor facilities for OSHA PEL 5 ppm ceiling and ACGIH TLV-C 0.5 ppm compliance)

Hydrogen chloride is the primary sentinel gas for TCS release events at polysilicon facilities: when TCS contacts atmospheric moisture at any humidity level, the hydrolysis reaction (SiHCl3 + 2H2O → SiO2 + 3HCl + H2) generates HCl fumes immediately and visibly — white aerosol observable by personnel in the area. At the TCS hydrolysis stoichiometry of 3 mol HCl per mol TCS, a TCS leak of 0.2 kg/hr generates approximately 0.16 kg/hr HCl. Area HCl detectors therefore simultaneously function as indirect TCS release detectors (via hydrolysis indicator) and direct HCl acute toxicity monitors. The ACGIH TLV-C of 0.5 ppm is a ceiling with zero tolerance for transient exceedance — HCl is a severe upper respiratory tract irritant with progressive corrosive injury above 1 ppm, with NIOSH IDLH 50 ppm representing the concentration for immediate 30-minute-exposure life threat. TCS vapor density of 4.67 produces accumulation in below-grade spaces below 0.3 m height, while HCl (vapor density 1.27 relative to air) disperses more uniformly; dual sensor heights address both TCS stratification and HCl distribution in service areas.

The adversarial attack uses ±8 DN downward pixel-value shift on the TCS area HCl detector display image. The actual HCl concentration is 28 ppm — 5.6× OSHA PEL 5 ppm; 56% NIOSH IDLH 50 ppm — from a TCS condensate pump PTFE valve packing failure releasing TCS liquid at 0.2 kg/hr into the distillation building atmosphere. At 45% RH ambient, TCS immediately hydrolyzes generating HCl aerosol fumes and white SiO2 solid smoke, both visible in the building. On a 0–60 ppm HCl display at 200 px height (0.3 ppm/px), the actual 28 ppm produces a bar at approximately 93 px. The ±8 DN perturbed image is classified as approximately 1 px — corresponding to 0.3 ppm, below the ACGIH TLV-C 0.5 ppm alarm. The area HCl AI monitoring system reports no alarm, no TLV-C exceedance, no IDLH approach. White SiO2 fumes are visible, but absent a confirmed HCl reading, operators may attribute the fume to normal process ventilation from adjacent equipment. The TCS condensate pump PTFE packing continues releasing at 0.2 kg/hr — generating simultaneous pyrophoric TCS release (spontaneous ignition risk from moisture contact) and HCl toxic exposure — until a scheduled walk-through detects the visible fume signature or a personnel SCBA response is triggered by personal monitor alarm.

2. Siemens CVD reactor deposition temperature AI (Emerson Rosemount 3144P CVD reactor temperature transmitter AI / Yokogawa EJA110A deposition zone temperature AI / Endress+Hauser iTHERM TM411 rod surface temperature AI / Honeywell STG94L thermocouple transmitter CVD AI / Siemens SITRANS T temperature transmitter AI — monitoring silicon seed rod surface temperature to maintain deposition within 1,050–1,150°C optimal range for crystalline polysilicon rod growth; preventing operation below 900°C where amorphous fine silicon powder is produced — 22nd upward-direction attack; first “product itself becomes pyrophoric” attack in the Glyphward portfolio)

The Siemens CVD deposition reaction (SiHCl3 + H2 → Si + 3HCl; ΔH approximately −72 kJ/mol at 1,100°C) proceeds at the hot rod surface by a heterogeneous surface-activation mechanism: adsorbed SiHCl3 is selectively decomposed at crystallographic surface step sites on the hot silicon lattice and incorporated into the growing crystal. This surface-activation mechanism is kinetically favored over gas-phase homogeneous nucleation at 1,050–1,150°C — crystalline rod deposition dominates. Below 900°C, the surface-activation rate constant drops below the gas-phase nucleation rate constant for the TCS/H2/temperature regime: TCS molecules react with H2 in the bulk reactor atmosphere before reaching and adsorbing at rod surface sites, producing suspended nanometer-scale Si clusters that coagulate and deposit as amorphous fine powder on all reactor surfaces. Fine Si powder at <1 μm particle diameter has specific surface area above 100 m²/g; exothermic surface oxidation (Si + O2 → SiO2; ΔH = −910 kJ/mol) occurs fast enough at this surface area to raise particle temperature above Si auto-ignition (~700°C in air) before heat can dissipate — self-sustaining pyrophoric ignition in air without external source. NFPA 654 classifies accumulations of combustible fine Si powder in CVD reactor buildings as regulated combustible dust hazards.

The adversarial attack uses ±10 DN upward pixel-value shift on the Siemens CVD reactor temperature transmitter display image. Actual rod temperature is 750°C — 150°C below the 900°C minimum crystalline deposition threshold — from resistance heating element aging in reactor #3 of a 12-reactor facility: element resistance increased 28% after 7,200-hour service from annealing of the tungsten-rhenium alloy at sustained high current density, reducing power delivery from 120 kW to 85 kW at design voltage at hour 68 of the 72-hour batch cycle. On a 600–1,200°C display at 200 px height (3°C/px), the actual 750°C produces a bar at approximately 50 px. The ±10 DN upward perturbed image is classified as approximately 167 px — corresponding to 1,101°C, apparently within the optimal 1,050–1,150°C deposition range. Fine Si powder production at 750°C: approximately 0.8 kg/hr from gas-phase nucleation. After 4 hours of undetected off-spec operation at 750°C, 3.2 kg of pyrophoric fine Si powder has accumulated on the CVD reactor floor, walls, and outer rod surfaces. This is the 22nd upward-direction attack in the Glyphward industrial AI portfolio and the first attack in the portfolio where the dangerous process condition causes the manufactured product itself to become pyrophoric: at 750°C, TCS CVD produces pyrophoric fine Si powder rather than crystalline Si rod, transforming the product stream from a safely-handled semiconductor material into a spontaneously-igniting combustible dust that awaits ambient air contact when maintenance crews open the reactor bell jar at batch end.

3. TCS/STC molar ratio process GC AI (Shimadzu GC-2030 FID TCS/STC ratio analyzer AI / Thermo Fisher TRACE 1310 GC chlorosilane process analyzer AI / ABB PGC1000 process gas chromatograph TCS purity AI / Yokogawa GC1000 Mark II TCS/SiCl4 ratio AI / Emerson Daniel Danalyzer chlorosilane process GC AI — monitoring TCS to silicon tetrachloride (STC; SiCl4) molar ratio in the Siemens CVD reactor feed to maintain TCS purity ≥98.5 mol%; excess STC is chemically inert at CVD temperatures and at high concentrations produces elevated HCl partial pressure that attacks graphite electrode seals and Hastelloy-C reactor internals)

TCS (BP 31.8°C) and STC (BP 57.7°C) are co-produced in the hydrochlorination reactor at a TCS:STC molar ratio of approximately 4:1 to 9:1 depending on temperature, HCl:Si stoichiometry, and catalyst loading. The TCS/STC distillation column (relative volatility 1.8; 30–50 theoretical stages at 5 bar absolute) separates the mixture to produce TCS distillate at >99.5 mol% purity at design tray efficiency (80–85%). If column tray fouling degrades efficiency — from polymer scale generated by trace olefin impurities in the hydrochlorination feed that polymerise on hot column trays — STC slip-through to the TCS distillate increases. At STC above 1.5 mol% in the CVD reactor feed: deposition rate decreases (SiCl4 + H2 → Si + 4HCl is thermodynamically unfavorable below 1,200°C at typical H2/STC ratios, so excess STC is chemically inert at 1,100°C and acts only as a dilutent reducing TCS partial pressure); HCl partial pressure in the reactor increases; graphite electrode seal chloride stress corrosion accelerates from the 5,000-hour design lifetime to 1,200–1,800 hours. The adversarial attack on the TCS/STC GC display (±8 DN downward: actual 8.2 mol% STC shown as 0.8 mol%) conceals column tray efficiency collapse from 83% to 31% after 4,200 hours of fouling, and the resulting accelerated graphite electrode seal degradation that will manifest as HCl bypass into the reactor building atmosphere at the next undetected seal failure. On a 0–15 mol% STC display at 200 px height (0.075 mol%/px), the actual 8.2 mol% STC bar at approximately 109 px is shifted by the ±8 DN perturbation to appear at approximately 11 px — corresponding to 0.8 mol% STC, within specification. The AI monitoring system reports TCS purity compliant; no column inspection scheduled; distillation continues at 4.6× the STC specification — advancing electrode seal failure on the order of weeks.

4. H2 carrier gas moisture analyzer AI (Servomex 2500 moisture-in-H2 trace analyzer AI / Vaisala DMT330 dew point moisture transmitter H2 AI / Meeco Aquamax moisture analyzer H2 AI / GE Panametrics Series 7 moisture analyzer H2 carrier AI / Michell Instruments XTP601 moisture transmitter H2 AI — monitoring water vapor content in hydrogen carrier gas fed to Siemens CVD reactors to maintain H2O below 0.1 ppm(v); excess moisture causes in-reactor TCS hydrolysis, SiO2 scale deposition, and product polysilicon quality degradation from oxygen incorporation)

The Siemens CVD process requires UHP H2 (≥99.999%) with H2O content below 0.1 ppm(v) — enforced by a molecular sieve dryer train at the CVD building H2 inlet. At 1,100°C CVD temperature, any H2O in the H2 carrier gas reacts essentially instantaneously with TCS via in-reactor hydrolysis (SiHCl3 + 2H2O → SiO2 + 3HCl; thermodynamic equilibrium reached in milliseconds at CVD temperature). Three adverse consequences develop simultaneously: (1) SiO2 solid deposits on all hot reactor surfaces at approximately 0.25 mm/hr at 165 ppm H2O, including on the growing silicon rod, graphite electrode mounting surfaces, and H2 inlet nozzles; (2) additional HCl from hydrolysis raises reactor HCl partial pressure 20–40% above the baseline from TCS deposition alone, accelerating graphite electrode seal and Hastelloy-C distributor chloride corrosion at 3–5× normal rate; and (3) oxygen incorporation from in-situ SiO2 in the growing silicon lattice introduces oxygen donor-type dopants that shift product N-type polysilicon resistivity from the 1,000 ohm·cm specification (equivalent to ~4.5×10²² phosphorus atoms/cm³) to 10–100 ohm·cm (oxygen donor density 2×10¹⁴–2×10¹⁵ atoms/cm³), rendering product off-specification for CZ crystal-puller feedstock and premium TOPCon solar cell applications.

The adversarial attack uses ±8 DN downward pixel-value shift on the H2 moisture analyzer display image. The actual H2O content is 165 ppm(v) — 1,650× the 0.1 ppm specification — from molecular sieve dryer saturation after 4,800 continuous operating hours without regeneration (regeneration cycle deferred due to production schedule priority). The analyzer data system shows an “overrange” flag at the 1 ppm display maximum; the ±8 DN downward perturbed image is classified as approximately 12 px — corresponding to 0.06 ppm, nominally within specification. The AI monitoring system reports “H2 moisture within spec — dryer performance nominal.” In the CVD reactor at 1,100°C and 165 ppm H2O, SiO2 deposits at 0.25 mm/hr on all reactor surfaces simultaneously — advancing graphite electrode seal failure, scaling H2 inlet nozzles, and producing off-specification polysilicon in every batch run on the saturated dryer, undetected by any alarmed monitoring surface until the batch harvest fails product resistivity specification.

The 22nd upward-direction attack: why the CVD temperature upward manipulation is structurally different from all previous downward temperature attacks in the portfolio

Across the Glyphward industrial AI portfolio, adversarial attacks on process temperature displays have consistently been downward — suppressing a temperature that is too high to appear within the safe operating range. Phosgene synthesis reactor temperature: ±10 DN downward, 162°C (catalyst deactivation and Cl2 slip onset) suppressed to appear 68°C. MIC storage tank temperature: ±10 DN downward, 28°C (water reaction exotherm) suppressed to appear 2°C. VCM suspension autoclave temperature: ±10 DN downward, 62°C (exothermic runaway approach) suppressed to appear 46°C. F2 electrolytic cell bath temperature: ±10 DN downward, 142°C (diaphragm fracture and HF evolution onset) suppressed to appear 82°C. In each of these cases: the process is running hotter than its design maximum; excessive temperature is the hazardous condition; and the adversarial manipulation makes the displayed reading appear cooler — within the safe operating range.

The Siemens CVD reactor temperature attack inverts this logic entirely. At 750°C — 300°C below the 1,050°C lower bound of the optimal deposition range and 150°C below the 900°C minimum crystalline deposition threshold — the process is running cooler than its design minimum, and the adversarial pixel shift raises the displayed temperature by 350°C to appear within the optimal 1,050–1,150°C range. The hazardous condition is process intensity deficit — insufficient CVD temperature producing pyrophoric fine powder rather than crystalline rod — not process intensity excess. This is the first process intensity upward attack in the portfolio: a perturbation that conceals a parameter value too low (not too high) relative to its safety-significant operating minimum. The 22nd upward-direction attack joins: (1) cooling water flow upward attacks (1st through 18th in the portfolio) that show deficient protective cooling flow as adequate; and (2) N2 inertisation deficiency-suppression upward attacks (5th, 6th, 7th N2 class entries) that show deficient inert blanket pressure as adequate. The CVD temperature attack is the first upward attack in a third mechanistic class: process-intensity minimum threshold attacks.

The structural significance for adversarial robustness evaluation is practical. Adversarial robustness test suites for AI monitoring systems in safety-critical applications — as defined under ISO/IEC 42001 AI management system requirements, the NIST AI Risk Management Framework Govern/Map/Measure/Manage functions, and the emerging EU AI Act Annex IV Article 9 technical documentation requirements for high-risk AI — predominantly define test adversarial cases as attacks that cause the AI to fail to detect a dangerous condition. For process temperature AI, the canonical test scenario is: actual temperature is dangerously high; adversarial perturbation (downward) causes AI to classify it as normal; alarm missed. An adversarial robustness evaluation framework testing only downward temperature perturbations on AI systems monitoring CVD reactors would not expose the temperature upward attack class — in which the dangerous condition is a temperature that is too low, and the adversarial manipulation is upward. The upward direction is non-intuitive for process temperature: engineers conventionally worry about runaway and overtemperature, not undertemperature. The CVD case is distinctive because it has a safety-significant lower bound as well as an upper bound, and the adversarial attack exploits the lower bound.

The pyrophoric product transformation compounds the attack severity. In all previous temperature attacks in the Glyphward portfolio, the hazardous material being released or approaching release existed before the adversarial manipulation began: phosgene was already in the synthesis reactor, MIC was already in the storage tank, VCM was already in the autoclave, HF was already at the Moissan cell. In the CVD temperature attack, the hazardous material — pyrophoric fine Si powder — does not exist at the start of the attack. It is created progressively by the adversarially concealed process condition: 0.8 kg/hr of pyrophoric fine powder depositing on the reactor floor and walls with every hour of off-spec operation at 750°C. After 4 hours, 3.2 kg accumulates — material that awaits ambient air contact at the batch-end rod harvest maintenance procedure. The ±10 DN upward adversarial attack on the temperature display thus concurrently conceals the accumulation of a secondary pyrophoric hazard that would not be anticipated by maintenance crews relying on AI-displayed temperature history showing a normal 1,100°C batch completion.

EPA RMP TQ 2,500 lbs vs OSHA PSM TQ 5,000 lbs: the dual regulatory gap and the adversarial robustness exemption for intermediate-inventory TCS facilities

The EPA Risk Management Program (40 CFR Part 68) and OSHA Process Safety Management (29 CFR 1910.119) are independent regulatory regimes, each calibrated by independent risk assessment processes and with independent threshold quantity values for the same chemicals. For nearly all chemicals listed in both programs, the EPA RMP TQ is greater than or equal to the OSHA PSM TQ. Trichlorosilane is the only chemical in the Glyphward industrial AI portfolio where the EPA RMP TQ (2,500 lbs) is lower than the OSHA PSM TQ (5,000 lbs) — an asymmetry that creates a distinct regulatory compliance landscape for intermediate-inventory TCS facilities and a corresponding gap in the regulatory mechanisms that would otherwise prompt adversarial robustness evaluation.

The reason for the EPA/OSHA TQ asymmetry lies in the different consequence models. OSHA PSM calibrates TQ for on-site worker protection: the OSHA threshold reflects the inventory at which a TCS release scenario (TCS vapour flash-fire; HCl cloud from hydrolysis) threatens on-site workers near the IDLH (50 ppm HCl). A 5,000-lb (2,268 kg) TCS inventory at a facility creates a credible worst-case on-site HCl release scenario consistent with OSHA’s occupational exposure hazard calibration. EPA RMP calibrates TQ for off-site community consequence: the EPA threshold reflects the inventory at which a TCS release produces an HCl cloud dispersion reaching public receptors beyond the facility fence at concentrations above the IDLH. Because TCS hydrolyzes at 3 mol HCl per mol TCS, a 2,500-lb (1,134 kg) TCS release generates approximately 935 kg HCl as the cloud source term. Under Pasquill-Gifford Class F atmospheric stability (strongly stable, light winds at 1.5 m/s) — the EPA worst-case dispersion scenario — this HCl source term can produce IDLH-exceedance (50 ppm) concentrations at distances extending beyond the facility fence to community receptors. The EPA found that community-consequence IDLH exceedance from TCS releases occurs at a lower inventory threshold (2,500 lbs) than the on-site occupational IDLH threshold (5,000 lbs).

The practical consequence for process safety management at TCS facilities: intermediate-inventory facilities (2,500–5,000 lbs TCS on-site) must implement EPA RMP Program 2 or Program 3 requirements — worst-case release scenario, off-site consequence analysis, community emergency response notification, five-year accident history — but are entirely exempt from OSHA PSM requirements, including the process hazard analysis (PHA) methods most likely to surface AI monitoring adversarial attack scenarios. OSHA PSM’s PHA obligation (HAZOP, What-If, checklist) is the regulatory mechanism that would produce worksheet entries such as: “What if the CVD reactor temperature AI displays 1,100°C when the actual rod temperature is 750°C? What if the area HCl detector AI shows 0.3 ppm when the actual HCl concentration is 28 ppm?” These are natural What-If hazard nodes in a PSM PHA of a Siemens CVD polysilicon facility. EPA RMP’s hazard review has narrower scope: it is primarily oriented toward characterising worst-case release quantity and dispersion for community emergency planning, not toward comprehensive on-site equipment failure analysis or AI monitoring system vulnerability assessment. For the intermediate TCS inventory band (2,500–5,000 lbs), the absence of OSHA PSM PHA requirement means that AI monitoring adversarial attack scenarios are substantially less likely to be identified by the regulatory-mandated analysis process — even while the EPA RMP requirement acknowledges the off-site community hazard significance of that inventory. Neither OSHA PSM nor EPA RMP, at any TCS inventory level, specifies adversarial robustness requirements for AI systems classifying rendered TCS process monitoring displays. Glyphward threshold 35 applies at all inventory levels above 2,500 lbs, reflecting the EPA RMP TQ as the lower bound of regulated community-consequence risk.

Integration: trichlorosilane Siemens CVD polysilicon process AI with Glyphward pre-scan gate

Glyphward integrates as a pre-scan gate between the DCS display capture layer and the AI inference pipeline for each TCS process monitoring context. If the adversarial score meets or exceeds threshold 35 — calibrated on the EPA RMP TQ 2,500 lbs and OSHA PSM TQ 5,000 lbs, the dual flammable-and-pyrophoric hazard profile of TCS, the 22nd upward-direction attack on CVD reactor temperature, and the first “product becomes pyrophoric” attack in the portfolio — the scan raises AdversarialTCSImageError and the monitoring AI does not process the frame.

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"

# Trichlorosilane Siemens CVD polysilicon monitoring contexts: threshold 35
# OSHA PSM 29 CFR 1910.119 Appendix A TCS TQ 5,000 lbs
# EPA RMP 40 CFR Part 68 Appendix A TCS TQ 2,500 lbs (LOWER than PSM TQ)
# ACGIH TLV-C 0.5 ppm HCl (from SiHCl3 hydrolysis: SiHCl3 + 2H2O -> SiO2 + 3HCl)
# Flash point -28 deg C NFPA Class IB; pyrophoric in moist air (H2 from hydrolysis)
# CVD temperature UPWARD attack (22nd): 750 deg C shown as 1,100 deg C
# -> pyrophoric fine Si powder (<1 um, >100 m2/g) accumulating at 0.8 kg/hr
# -> 3.2 kg after 4 hours; NFPA 654 combustible dust at batch-end rod harvest
TCS_THRESHOLD = 35


class TCSProcessContext(Enum):
    AREA_HCL_DETECTOR = "area_hcl_detector"
    CVD_REACTOR_TEMPERATURE = "cvd_reactor_temperature"
    TCS_STC_MOLAR_RATIO = "tcs_stc_molar_ratio"
    H2_MOISTURE_CONTENT = "h2_moisture_content"


class AdversarialTCSImageError(Exception):
    """Raised when any TCS CVD monitoring image scores >= 35.
    AREA_HCL_DETECTOR uncaught: 28 ppm HCl (5.6x OSHA PEL; TCS pyrophoric release) as 0.3 ppm.
    CVD_REACTOR_TEMPERATURE uncaught: 750 deg C (pyrophoric Si powder) shown as 1,100 deg C.
    TCS_STC_MOLAR_RATIO uncaught: 8.2 mol% STC (above 1.5% spec) shown as 0.8 mol%.
    H2_MOISTURE_CONTENT uncaught: 165 ppm H2O (1,650x spec) shown as 0.06 ppm."""

    def __init__(self, scan_id, score, context, unit_id, flagged_region=None):
        self.scan_id = scan_id
        self.score = score
        self.context = context
        self.unit_id = unit_id
        self.flagged_region = flagged_region
        super().__init__(
            f"Adversarial TCS image: context={context.value} "
            f"score={score} unit={unit_id} scan_id={scan_id}"
        )


async def scan_tcs_image(image_bytes, context, unit_id, client):
    image_hash = hashlib.sha256(image_bytes).hexdigest()
    payload = {
        "image": base64.b64encode(image_bytes).decode(),
        "source": f"tcs:{context.value}:{unit_id}",
        "metadata": {
            "unit_id": unit_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) >= TCS_THRESHOLD:
        raise AdversarialTCSImageError(
            scan_id=result["scan_id"],
            score=result["score"],
            context=context,
            unit_id=unit_id,
            flagged_region=result.get("flagged_region"),
        )
    return result


async def main():
    async with httpx.AsyncClient() as client:
        with open("cvd_reactor_temp_screenshot.png", "rb") as f:
            image_bytes = f.read()
        result = await scan_tcs_image(
            image_bytes,
            TCSProcessContext.CVD_REACTOR_TEMPERATURE,
            unit_id="CVD-REACTOR-03-TEMP",
            client=client,
        )
        print(f"Clean scan: {result['scan_id']} score={result['score']}")


asyncio.run(main())

Frequently asked questions