Weld Pool Geometry AI Analysis: Real-Time Quality Prediction from Molten Pool Data

The Weld Pool as a Quality Signal

Every weld defect has a precursor. Porosity begins as dissolved gas trapped before solidification. Lack of fusion forms when the pool fails to wet the base metal at the joint sidewall. Hot cracking initiates in a mushy zone where tensile stress exceeds the strength of partially solidified metal. In each case, the weld pool geometry — the shape, size, temperature gradient, and dynamic behaviour of the molten zone — contains measurable information about what is about to go wrong.

Weld pool geometry AI analysis is the discipline of extracting that information in real time, before the defect solidifies and becomes a non-conformance requiring rework or rejection. Thermal cameras, operating in the near-infrared and mid-wave infrared bands, are uniquely suited to this task: they see through the bright arc plasma to the pool periphery, measure calibrated temperatures at 50–200 Hz frame rates, and feed a continuous stream of geometric and thermal data to an AI inference engine.

This post explains the physics of weld pool geometry as a quality predictor, the specific geometric features that matter most, how AI models are trained and deployed to act on those features, and how a production welding engineer can implement real-time pool monitoring on a live line.

Why Pool Geometry Predicts Defects

The weld pool is a transient liquid-solid system governed by heat transfer, fluid dynamics, and surface tension. Its geometry at any instant is determined by the balance between energy input from the heat source and energy dissipation into the surrounding base metal and atmosphere.

When that balance shifts — due to parameter drift, consumable variability, fit-up deviation, or material property variation — the pool geometry changes first. The defect follows.

Pool Length and Width: Heat Input Proxy

Pool length (in the welding direction) and width (transverse) are the most direct geometric indicators of instantaneous heat input. Per AWS D1.1/D1.1M:2025, heat input governs HAZ toughness, distortion, and solidification microstructure. A pool that is too long indicates excessive heat — risk of burn-through, excessive grain growth, and reduced HAZ toughness. A pool that is too short indicates insufficient fusion.

For a given WPS (welding procedure specification), production engineers define a control corridor: acceptable pool dimensions at the nominal travel speed and current. Any excursion outside the corridor triggers an alert.

Pool Aspect Ratio: Penetration Indicator

The ratio of pool length to pool width (elongation ratio) correlates with penetration depth and mode. In GMAW and GTAW, a circular pool indicates low-energy shallow penetration; an elongated tear-drop pool indicates higher energy with deeper penetration. For groove welds requiring full penetration per ISO 5817:2023, the AI model learns the aspect ratio corridor associated with verified penetration from the weld procedure qualification record (WPQR).

Pool Tail Geometry: Solidification Crack Risk

The trailing edge of the weld pool — the “tail” — is where solidification initiates. A broad, flat tail distributes solidification stress across a larger area. A narrow, pointed tail concentrates stress at a single point, dramatically increasing the risk of centreline solidification cracking (hot cracking). Pool tail angle is a real-time indicator that AI models correlate with cracking propensity, particularly in austenitic stainless steels, nickel alloys, and dissimilar-metal joints.

Thermal Gradient (ΔT/Δx): Microstructure Predictor

The spatial temperature gradient across the pool and into the HAZ predicts the solidification rate and, consequently, the grain structure of the weld metal. A steep gradient (rapid temperature drop at the pool boundary) drives fine columnar grain growth. A shallow gradient (slow temperature drop) promotes coarse grains and reduces toughness. AI models trained with EBSD metallographic data from coupon welds learn the gradient signatures associated with acceptable microstructure for a specific material and process.

What a Thermal Camera Actually Measures at the Weld Pool

Spatial Resolution and Frame Rate Requirements

Resolving weld pool geometry requires adequate spatial resolution. A pool in a typical GMAW pass is 8–15 mm wide. An infrared camera with a 640×512 pixel array and a close-focus lens calibrated to cover a 50×40 mm field of view gives approximately 12 pixels per millimetre — sufficient to detect 1 mm changes in pool width reliably.

Frame rate must be high enough to capture dynamic events. Pool oscillation frequencies in GMAW are typically 20–80 Hz. Cameras operating at 100+ Hz capture the full oscillation dynamics; cameras at 25 Hz will alias the signal but still provide stable time-averaged pool geometry.

Near-Infrared vs Mid-Wave Infrared

Near-infrared (NIR, 0.9–1.7 µm) cameras are well suited to pool observation because:

• Arc plasma emission is partially attenuated in the NIR band with appropriate filtering
• Silicon-based optics are inexpensive and thermally stable
• High-speed NIR imaging at 200+ Hz is achievable with InGaAs detector arrays

Mid-wave infrared (MWIR, 3–5 µm) cameras offer calibrated temperature measurement across the full pool and HAZ, enabling quantitative thermal gradient extraction. The tradeoff is cost and the need for anti-reflective coatings compatible with the MWIR band.

HeatCore AI uses a proprietary dual-band approach: a NIR channel for high-frame-rate pool geometry extraction and a MWIR channel for calibrated temperature gradient analysis. Data from both channels are timestamped and fused per weld pass.

Emissivity and Calibration

Quantitative temperature measurement from the weld pool requires emissivity calibration. Liquid steel emissivity at 1500–1600°C is approximately 0.35–0.45 and varies with surface oxidation and alloy content. For geometry analysis (rather than absolute temperature), emissivity variation has limited impact: pool boundaries are identified by temperature gradient rather than absolute value, making geometric measurements robust to emissivity uncertainty.

For thermal gradient analysis requiring calibrated temperatures, two-colour pyrometry or emissivity-corrected single-channel measurement is used. HeatCore AI applies per-material emissivity tables from WPQR qualification coupons.

AI Models for Real-Time Pool Analysis

Feature Extraction Pipeline

The AI pipeline for weld pool geometry analysis operates in three stages:

1. Preprocessing: Raw thermal frames are denoised, lens-distortion corrected, and normalized to a consistent temperature scale. A Region of Interest (ROI) tracker locks onto the pool and arc location using the previous frame’s prediction, compensating for torch movement and workpiece vibration.

2. Segmentation: A convolutional neural network (CNN) performs semantic segmentation of each frame, classifying pixels as: solid base metal, HAZ, mushy zone (partially solidified), liquid pool, and arc/plasma. The liquid pool boundary is the primary output used for geometric feature extraction.

3. Feature extraction and classification: From the segmented pool boundary, geometric features are computed: length, width, area, aspect ratio, tail angle, and centroid offset from the weld seam centreline. These features are fed to a second model — a recurrent network (LSTM or Transformer) — that integrates the time-series of pool geometry across the pass and produces a real-time quality score and defect-class prediction.

Training Data Requirements

Training a reliable pool geometry model requires labelled data from the specific process and material family it will monitor. The labelling process involves:

• Running qualification welds with intentional parameter variations (travel speed ±20%, current ±15%, voltage ±10%)
• Recording synchronized thermal video and weld parameters
• Performing destructive and non-destructive evaluation of the welds (radiographic testing per ISO 17636-2, ultrasonic testing, cross-section metallography)
• Labelling each thermal sequence with the defect class (no defect, porosity, lack of fusion, undercut, hot crack) confirmed by NDT

A minimum dataset of 200–500 labelled weld passes per material-process combination is typically required for a model with production-grade precision and recall.

Online Adaptation and Drift Detection

Production conditions drift. Wire lot changes, torch wear, shielding gas variability, and base metal surface condition all introduce gradients the training data may not fully cover. HeatCore AI includes an online adaptation layer that monitors the statistical distribution of pool geometry features and flags distribution shift — signalling that the model’s confidence has degraded and recalibration or retraining is warranted.

This is the AI equivalent of a calibration interval: the system actively monitors its own reliability rather than assuming static accuracy over time.

Implementation: Deploying Weld Pool Monitoring in Production

Step 1: Camera Mounting and Field of View

The camera must have a clear line of sight to the pool, typically at 30–45° from vertical, positioned 150–300 mm from the arc. For robotic welding, the camera mounts on the robot arm or a fixed bracket with the robot TCP in view. For manual welding stations, a fixed overhead bracket with a motorized tilt axis that follows the torch is used.

Key mounting requirements:

• Shielding gas envelope integrity: Camera enclosure must not disrupt shielding gas coverage
• Thermal protection: Camera body temperature must remain within operating spec (typically <50°C ambient); active cooling is required near hot parts
• Vibration isolation: Arm-mounted cameras require vibration damping to avoid image blur at high frame rates

Step 2: Process Baseline and Corridor Definition

Before AI-based defect prediction is activated, the system runs in logging-only mode for a baseline period (typically one production shift). During this period:

• Pool geometry statistics are accumulated: mean and standard deviation of length, width, aspect ratio, tail angle
• Weld quality is confirmed by standard post-weld inspection (VT, MT, RT as appropriate for the WPS)
• The baseline statistics become the control corridor: alert thresholds are set at ±2σ from the mean for each geometric feature

This corridor is WPS-specific and material-specific. A new WPS requires a new baseline. The baseline is stored in the welding QMS as a quality record, traceable to the WPQR that qualified it.

Step 3: Real-Time Alert and Response

During production, each weld pass is monitored in real time. When a geometric feature exits the control corridor:

1. An alert is logged with timestamp, weld ID, feature name, observed value, and corridor limits
2. If the excursion persists for more than a configurable number of frames (typically 0.5–2 seconds), the alert escalates to a non-conformance event
3. The welder or robot controller receives a signal (visual, audible, or closed-loop parameter correction)
4. The affected weld segment is flagged in the digital weld record for post-weld NDT verification

This workflow integrates with the the HeatCore QMS workflow NCR management module, automatically opening a non-conformance record linked to the weld ID, thermal video clip, and parameter log.

Step 4: Closed-Loop Correction (Robotic Lines)

On robotic lines with accessible control interfaces (Fanuc, KUKA, ABB, Yaskawa), HeatCore AI can output correction signals directly to the robot controller. The most common corrections are:

• Travel speed correction: If pool length exceeds the upper corridor limit, travel speed is increased by 2–5% to reduce heat input
• Wire feed rate correction: If pool area drops below the lower corridor limit (indicating insufficient fill), wire feed is increased within the WPS parameter window
• Torch position correction: If pool centroid drifts from the seam centreline, the robot path is corrected laterally within the joint preparation tolerance

Closed-loop correction requires rigorous safety validation and must be qualified as part of the WPS qualification, with documented parameter windows within which autonomous correction is permitted. Corrections outside the window are not applied — the pass is flagged for review.

Integration with Quality Records and Traceability

Every pool geometry measurement is timestamped and linked to:

• Weld ID (unique per joint per pass)
• Operator or robot program ID
• Batch/heat number of base material and consumables
• WPS reference
• Parameter log (current, voltage, travel speed, wire feed)

This creates a complete digital weld record that satisfies the traceability requirements of ISO 3834-2:2021 and is ready for audit by certification bodies, customer quality teams, and regulatory inspectors. For pressure equipment manufacturers, this record set supports conformity assessment under the Pressure Equipment Directive (PED 2014/68/EU).

The pool geometry time-series is compressed and stored as a per-weld data object, not as raw video — reducing storage requirements by 95% while retaining all quality-relevant signals.

Which Defects Are Most Predictable from Pool Geometry?

Based on published research and production validation data, the defect classes most reliably predicted from pool geometry features are:

Porosity is the least predictable from geometry alone because the gas bubble formation and entrapment event is sub-millimetre and occurs within the pool interior — not visible at the pool boundary. Pool oscillation analysis improves porosity detection probability, but acoustic emission monitoring or X-ray fluoroscopy is required for high-confidence detection.

HeatCore AI: Purpose-Built for Weld Pool Geometry Analysis

Therness HeatCore AI integrates all the capabilities described in this post into a production-ready system:

• Dual-band thermal + NIR imaging with calibrated temperature measurement
• On-edge AI inference (no cloud dependency) with <50 ms detection latency
• WPS-linked baseline management and control corridor definition
• Real-time alert generation and the HeatCore QMS workflow NCR integration
• Closed-loop robot parameter correction API (Fanuc, KUKA, ABB, Yaskawa)
• Digital weld record export in ISO 3834-compliant format

HeatCore AI has been validated on GMAW, GTAW, FCAW, and laser welding processes across carbon steel, stainless steel, and aluminium alloy material families.

For more on how HeatCore AI captures thermal data across the full weld pass — not just the pool — read our deep dive on HeatCore AI thermal weld monitoring. For the regulatory traceability context, see our guide to ISO 3834 and EN 1090 welding traceability with thermography. For how pool geometry integrates with broader quality records, see digital welding quality records for WPS, PQR, and traceability.

Summary

Weld pool geometry AI analysis transforms the molten pool from a transient thermal event into a continuous quality signal. By measuring pool length, width, aspect ratio, tail geometry, and thermal gradient at production frame rates, AI models predict weld defects before solidification — giving process engineers the ability to respond in real time rather than discovering non-conformances at final inspection.

The technology is mature enough for production deployment on robotic and semi-automated lines today, with validated models for the most common defect classes in structural, pressure vessel, and automotive welding applications. The key implementation requirements are adequate camera resolution, a process baseline period, WPS-aligned control corridors, and integration with the welding QMS for traceability.

For welding engineers ready to move from post-weld inspection to in-process quality control, weld pool geometry monitoring is the highest-leverage entry point: it provides actionable quality signals from the process itself, without adding post-weld inspection steps or disrupting production flow.