WAAM Thermal Forecasting

Leakage-aware thermal-event forecasting, physics-guided modelling, inverse PINNs, and safety-gated adaptation for Wire Arc Additive Manufacturing.

Overview

This repository contains an end-to-end research pipeline for forecasting and reconstructing thermal behaviour in multi-layer WAAM using real four-channel thermocouple recordings from three independent steel-wall builds.

The project was designed around the realities of WAAM experimental data:

• logger restarts and uncertain recording gaps;
• sensor reconnection and channel-quality changes;
• unequal deposition strategies across builds;
• strong heat accumulation across layers;
• only three independent physical builds despite many sequential thermal cycles;
• the risk of temporal and build-level data leakage;
• the fact that a complex PINN may not outperform a simple causal baseline on a new build.

The final repository therefore does not present a single over-claimed neural model. It separates three different WAAM prediction tasks and evaluates each under an appropriate validation protocol.

WAAM experimental programme

Build	Current	Voltage	Travel speed	Heat input	Deposition strategy	Approx. layers
W1	140 A	24 V	3.40 mm/s	0.6 kJ/mm	Two overlapping opposite-direction passes per layer	27
W2	180 A	23.5 V	2.10 mm/s	1.2 kJ/mm	One deposition pass; arc off during return	26
W3	220 A	24 V	1.7167 mm/s	1.8 kJ/mm	One deposition pass; arc off during return	24–25

Common conditions:

• ST37 substrate: 150 × 100 × 10 mm;
• wall deposited along the 150 mm longitudinal direction at the transverse centreline;
• nominal wall height: approximately 50 mm;
• nominal wall width: approximately 20 mm;
• shielding gas flow: 18 L/min;
• ambient temperature: approximately 30 °C;
• target interpass temperature: approximately 200 °C, with some restarts near 185 °C;
• cooling intervals increased from roughly 60 s in lower layers to as much as 900 s in upper layers.

Thermocouple geometry

All four thermocouple junctions were located at the longitudinal midpoint, x = 75 mm, and approximately at substrate mid-thickness, z = -5 mm when the substrate top surface is z = 0.

Sensor	Coordinate (x, y, z) mm	Role
D1	(75, -20, -5)	near, D side
D2	(75, -40, -5)	far, D side
U1	(75, +20, -5)	near, U side
U2	(75, +40, -5)	far, U side

The D/U signs are a modelling convention. Physical left/right naming is not asserted.

Dataset scope and audit decisions

• 76 accepted thermal cycles across the three successful builds;
• 71 event-to-next-event forecasting records after respecting sequence boundaries;
• 3 independent physical builds—the sequential rows are not treated as independent experiments;
• the first failed/collapsed 220 A attempt is retained only as auxiliary out-of-distribution provenance;
• two unusable logger files are retained under data/auxiliary/rejected/ and excluded from quantitative modelling;
• no forecasting target crosses a logger restart or sensor-reconnection boundary;
• W3 uses time-varying sensor masks because channel quality changes between recording segments;
• one W3 event with a single formal derivative vote is explicitly audited and subjected to strict sensitivity analyses.

Full provenance is documented in docs/Data_Provenance.md.

Research pipeline

Stage	WAAM-focused purpose
Stage 1	logger parsing, channel QC, smoothing, event detection, sequence protection
Stage 2	strictly causal feature construction and leakage-aware ML baselines
Stage 3	lumped thermal-memory models and physics-guided residual learning
Stage 4.0–4.1	canonical WAAM geometry, APDL/FEM audit, PINN-readiness assessment
Stage 4.2	within-build sensor-constrained inverse PINN on W2
Stage 4.3	strict leave-one-build-out conditional PINN transfer
Stage 4.4	chronological few-shot adaptation using early layers of a new build
Stage 4.5	uncertainty-gated PINN residual correction with mandatory causal fallback
Stage 4.6	final evidence synthesis, uncertainty audit, model card, and claim lock

Key results

The rows below correspond to different prediction tasks and should not be ranked against one another without considering the target and validation design.

Task	Validation	Selected model	MAE	Interpretation
Next-event peak forecasting	Leave-one-build-out	PhysicsResidualRidgeCausal	19.23 °C	Final primary deployable research model; 20.3% lower MAE than persistence
Next-event peak with known future dwell	Leave-one-build-out	PhysicsMemoryScheduleAware	18.14 °C	Lower MAE, but requires the next schedule interval in advance
W2 held-out event-trace reconstruction	Complete events withheld	Sensor-Constrained PINN	16.12 °C	2.43% better than matched data-only network; spatial field remains unvalidated
New-build continuation after five early events	Chronological few-shot test	Uncertainty-Gated Hybrid	25.62 °C	Small pooled improvement over 26.62 °C causal baseline; not statistically conclusive across three builds

Primary next-event result

The strictly causal PhysicsResidualRidgeCausal model achieved:

• MAE: 19.23 °C
• RMSE: 26.24 °C
• bias: −1.89 °C
• R²: 0.736
• MAE improvement over persistence: 20.3%

It reduced MAE on each held-out build separately. W3 nevertheless retained a negative held-out R², so its result is reported as an MAE improvement rather than reliable trend reconstruction.

What the PINN experiments showed

The within-build W2 inverse PINN slightly improved on a matched data-only network and recovered physically plausible effective parameters. However, strict cross-build testing exposed substantial distribution shift: the conditional PINN was worse than the causal sensor baseline on every completely unseen wall. Few-shot adaptation reduced the zero-shot PINN error but still did not beat the baseline consistently.

The final hybrid therefore preserves the causal prediction and permits only a bounded PINN residual correction when early-build out-of-fold evidence supports it. For W2 the gate closed completely and used alpha = 0.

This negative-to-safe-hybrid progression is a central result of the project, not a hidden failure.

Quick start

1. Clone and create an environment

git clone <YOUR_REPOSITORY_URL>
cd WAAM-Thermal-Forecasting
python -m venv .venv

Windows PowerShell:

.venv\Scripts\Activate.ps1
python -m pip install --upgrade pip
pip install -e .

Linux/macOS:

source .venv/bin/activate
python -m pip install --upgrade pip
pip install -e .

2. Verify the locked repository

pytest -q
python scripts/validate_repository.py

Expected test result:

51 passed

3. Run the main causal WAAM pipeline

python scripts/run_stage1.py
python scripts/run_stage2.py
python scripts/run_stage3.py
python scripts/run_final_validation.py

Or:

python scripts/run_pipeline.py --profile core

4. Run the PINN and safe-adaptation research stages

python scripts/run_stage4_2.py
python scripts/run_stage4_3.py
python scripts/run_stage4_4.py
python scripts/run_stage4_5.py
python scripts/run_stage4_6.py

A full retraining run is substantially slower than the core causal pipeline. Precomputed audited outputs are included for inspection and test reproducibility.

See docs/REPRODUCIBILITY.md for stage dependencies, expected outputs, and practical runtime guidance.

Repository structure

WAAM-Thermal-Forecasting/
├── assets/                  # GitHub banner, geometry, workflow and key figures
├── configs/                 # audited experimental and model configurations
├── data/
│   ├── raw/                 # successful WAAM logger segments
│   └── auxiliary/           # failed/rejected recordings, excluded from main fitting
├── docs/                    # experimental setup, data dictionary, methods, results and limits
├── references/              # legacy APDL/FEM provenance and build photographs
├── results/                 # locked stage outputs and manuscript-ready evidence
├── scripts/                 # stage runners and repository validation
├── src/waam_thermal/        # reusable implementation
└── tests/                   # 51 automated scientific and software checks

Scientific claim boundary

This repository supports the following claim:

An audited proof-of-concept for causal next-event thermal-peak forecasting in multi-layer WAAM, complemented by within-build sensor-constrained PINN reconstruction and a safety-gated early-build residual hybrid.

It does not claim:

• industrial deployment readiness;
• validated full 2D or 3D temperature fields away from the four thermocouples;
• sensor-free operation;
• generalisation beyond the three independent successful builds;
• statistically conclusive superiority of the hybrid over the causal fallback;
• uniquely identified heat-source or boundary-condition parameters.

Detailed limitations are in docs/LIMITATIONS.md and the final model card in docs/MODEL_CARD.md.

Citation

Citation metadata are provided in CITATION.cff. GitHub will expose a Cite this repository button after publication.

Suggested citation:

Mori, A. (2026). WAAM Thermal Forecasting: Leakage-Aware Causal Modelling, Sensor-Constrained PINNs, and Safety-Gated Adaptation (Version 1.5.0) [Computer software and research dataset].

Licences

• Source code: MIT License
• Experimental data, figures and photographs: CC BY-NC 4.0

Before public release, confirm that the experimental data and photographs may be redistributed under the selected data licence.

Author

Ali Mori — WAAM experiments, welding metallurgy interpretation, thermal-data audit, modelling, validation, and project documentation.

Acknowledgement of research scope

The repository is intentionally transparent about limited independent sample size, measurement heterogeneity and negative model-transfer results. Its value lies in reproducible WAAM data engineering, leakage-aware validation, physics-guided modelling, and explicit separation between promising research behaviour and deployment-ready evidence.