Mechanisms Engineering Test Loop (METL)

• METL is a full-scale experimental platform that emulates sodium-cooled fast reactor dynamics with robust thermal-hydraulic and control systems.
• It integrates over 200 sensor channels and multi-layer control architectures to capture detailed physical behavior and enable precise data acquisition.
• The platform supports AI-driven cybersecurity evaluations through structured attack taxonomies and machine learning benchmarks with validated performance metrics.

The Mechanisms Engineering Test Loop (METL) at Argonne National Laboratory is a full-scale experimental platform engineered to emulate the thermal-hydraulic and control dynamics of a sodium-cooled fast reactor primary loop. Designed for both advanced nuclear reactor research and the evaluation of AI-driven cybersecurity for cyber-physical systems, METL provides a cyber-physical environment with authentic physical behavior, complex instrumentation, and robust control architectures. Recent work has established METL as a cornerstone for benchmarking artificial intelligence approaches to cybersecurity, with a structured evaluation framework and comprehensive attack taxonomy applied to multivariate operational datasets (Blakely et al., 1 Dec 2025).

1. Physical Architecture and Core Subsystems

METL comprises a closed-loop, liquid-sodium flow architecture mimicking the core features of a fast-reactor coolant system. The primary loop includes the following components:

• Test Vessel ("Core"): Operates at liquid-sodium temperatures up to approximately 1200 °F (649 °C), forming the centerpiece of experimental campaigns.
• Electromagnetic Pumps: Circulate liquid sodium at controlled flow rates, driven by external power supplies and governed by programmable logic controllers (Emerson/NI cRIO).
• Cold-Trap Purification Subsystem: A bypass purifier removes dissolved gases and impurities, regulated by isolation valves and dampers.
• Shell-and-Tube Heat Exchanger: Transfers heat from the sodium loop to a secondary air or inert-gas stream, simulating reactor decay-heat removal.
• High-Temperature Piping: Fabricated from Inconel or stainless steel, interconnecting pumps, vessel, cold trap, and heat exchanger; extensive valve networks facilitate flow routing, isolation, and maintenance configurations.

Instrumentation racks and industrial control cabinets (Emerson DeltaV, NI cRIO) provide data acquisition and real-time control, while an operational-technology (OT) segmented network physically isolates plant control from information-technology (IT) infrastructure. Read-only REST API endpoints ensure data integrity for downstream analysis and testbed integration.

2. Instrumentation, Sensing, and Data Acquisition

214 sensor channels were employed in AI-based cybersecurity experiments, selected from a total signal inventory exceeding 11,899 points. Key sensor modalities and acquisition parameters include:

• Temperature Sensing: Bulk-sodium thermocouples and resistance temperature detectors (RTDs) cover 100 °C–650 °C; air-side temperature monitoring spans 0 °C–150 °C.
• Flow Measurement: Magnetic flow meters (0–100 L/min) monitor sodium circulation.
• Pressure Transducers: Gauge pressures (0–2 MPa) at critical loop locations.
• Level Gauges: Cold-trap sodium levels with 0–100% span.
• Valve Position and Damper Sensors: Continuously indicate actuation states.

Sampling is event-responsive, defaulting to 1/60 Hz in steady-state operation, ramping to 1 Hz during valve actuations and maintenance. Data are acquired via REST API, timestamped using ISO 8601 convention, and delivered in JSON or CSV formats. Embedded watchdogs in cRIO firmware provide built-in sensor fault detection, while critical sensors in the cold trap employ redundancy for cross-validation.

3. Control System Structure and Feedback Logic

METL's control architecture is stratified into four layers:

• Field Layer: NI cRIO hardware interfaces with analog/digital inputs and outputs across the physical plant.
• Control Layer: Emerson DeltaV (or analogous PLC) implements core PID loops, responsible for:
  • Pump speed regulation (flow control)
  • Loop temperature maintenance (coolant management)
  • Cold-trap bypass (impurity management)
  • Heat-exchanger operations (fan and valve actuation)
• Supervisory Layer: Windows-based human-machine interface (HMI) enables operator oversight, recipe execution, and manual override, with all control actions constrained by PLC logic.
• Data-Extraction Layer: Hardened gateway serves read-only, real-time data streams to external proxies, isolating plant logic from experimental testbed modifications.

Feedback mechanisms include flow-controlled pump speed (magnetic flow meter as process variable), temperature-regulated coolant flow, and impurity-driven cold-trap cycling. Safety interlocks, enforced in PLC logic, prohibit non-permissive valve states during hazardous conditions.

4. Fluid Dynamics and Thermal-Hydraulic Modeling

The governing equations for METL dynamics reflect canonical models in thermofluids engineering:

• Mass Continuity:

$$\frac{\partial\rho}{\partial t} + \nabla \cdot (\rho \mathbf{v}) = 0$$

where $\rho$ is sodium density and $\mathbf{v}$ velocity.

• Momentum (Incompressible Navier–Stokes):

$$\rho \left( \frac{\partial \mathbf{v}}{\partial t} + \mathbf{v} \cdot \nabla \mathbf{v} \right) = -\nabla p + \mu \nabla^2 \mathbf{v} + \rho \mathbf{g}$$

• Bernoulli (Steady, Inviscid):

$$p + \frac{1}{2} \rho |\mathbf{v}|^2 + \rho g h = \text{const}$$

• Thermal-Energy Balance:

$$\rho C_p \left( \frac{\partial T}{\partial t} + \mathbf{v} \cdot \nabla T \right) = k \nabla^2 T + \dot{q}^{\prime\prime\prime}$$

• Convective Wall Heat Transfer:

$\dot{q} = h A (T_s - T_f)$

with $h$ as convective transfer coefficient.

• Equation of State (Liquid Sodium Approximation):

$p \approx \frac{ \rho R_s T }{1 + \beta (T - T_0)}$

These models underpin not only design and operation, but also the validation of synthetic data modifications used in cybersecurity benchmarking.

5. Operational Envelope and Standard Procedures

METL supports a diverse regime of thermophysical conditions:

• Working Fluid: Liquid sodium, $C_p \approx 1.3$ kJ/kg·K, $\rho \approx 850$ kg/m³ at 550 °C.
• Flow Rate: 10–100 L/min.
• Pressure: 0.1–1.5 MPa (gauge).
• Temperature: Sodium up to 650 °C; cold trap at 200–300 °C; air-side exchanger at 20–150 °C.

Standard operating procedures include argon-purged sodium fill, controlled ramp-up ($\leq 5$ °C/min), extended steady-state data collection, dynamic sample-rate elevation during maintenance (>1 Hz), and staged cooldown phases. This comprehensive environmental control is foundational for high-fidelity data generation and repeatable experimentation.

6. Dataset Structure and Preprocessing Protocols

METL's data-handling paradigm prioritizes reproducibility and high-quality time series analytics:

Attribute	Value/Description	Purpose/Context
Channels	214 channels (temperature, flow, pressure, etc.)	Input to ML, anomaly detection
Storage Format	Apache Parquet (columnar, daily partitions)	Efficient, queryable storage
Acquisition Mode	Batch (file-watcher); streaming (MQTT; unused in referenced study)	Flexible ingestion for experiments
Sampling/Windowing	Aligned to 30 s intervals (via interpolation/forward-fill); 50-sample windows	Standardizes ML feature extraction
Normalization	Min–Max scaling with 30 d historical min/max	Channel-level data normalization
Missing Data	Gaps >5 min flagged as “outage;” excluded from modeling	Robustness in model training/testing

Preprocessing aligns, normalizes, and windows data, with explicit exclusion of prolonged outages and precise flagging of operational modes. Each record encodes timestamp, channel ID, raw value, units, and sample rate flag, forming the core corpus for machine learning pipelines.

7. Integration with AI-Driven Cybersecurity Evaluation

METL operational data underpins a full-cycle, reproducible AI cybersecurity evaluation pipeline (Blakely et al., 1 Dec 2025):

• Data Proxy and Attack Injection: Real-time streams from METL are siphoned into the ARSS Data Proxy, where multivariate batch datasets are stored and subjected to attack injections via an extensible transformer library (19 transformer types, 15 attack scenarios, severity tiers). This methodology preserves ground-truth labeling by never altering the live physical process.
• Machine Learning Benchmarks: Four paradigms—Change Point Detection, LSTM-based Anomaly Detection, Dependency Violation Analysis, and Autoencoder Reconstruction—are evaluated against paired clean/attacked data streams in lock-step. The experiment protocol includes 300 trials with stratified training (70%), validation (15%), and testing (15% plus attacks) splits, with anomaly scoring and ROC/AUC computed for all scenario-tier combinations.
• Benchmark Results: Change Point Detection demonstrated superior mean AUC (0.785), followed by LSTM Anomaly Detection (0.636), Dependency Violation (0.621), and Autoencoder methods (0.580). Attack detectability varied systematically (e.g., multi-site coordinated attacks: AUC=0.739; precision trust decay: AUC=0.592).

This architecture, including read-only HMI operation and offline dual-stream evaluation, enforces rigorous experimental integrity and sets a robust standard for cyber-physical anomaly detection in critical infrastructure research.

---

METL thus serves as a dual-purpose experimental asset: both as a premier thermal-hydraulic testbed for advanced reactor systems and as a high-fidelity, data-rich substrate for systematic AI-driven cybersecurity benchmarking and research (Blakely et al., 1 Dec 2025).