TBC-ML-Pipeline

Machine learning pipeline for predicting the effective thermal conductivity of Thermal Barrier Coating (TBC) systems composed of Yttria-Stabilized Zirconia (YSZ) over a CMSX-4 single-crystal nickel superalloy substrate.

---

Overview

The pipeline consists of three sequential stages:

Stage	Script	Description
1	generate_cards.py	Generates randomized YSZ material cards as ABAQUS .inp files
2	abaqus_ml_pipeline.py	Runs all simulations in ABAQUS, extracts thermal results from .odb files
3	train_model.py	Trains a Random Forest model to predict effective thermal conductivity

---

TBC-ML-Pipeline/
├── run_pipeline.py                # Main orchestrator; handles environmental auto-detection
├── src/                           # Logic layer (The creative intellect of the project)
│   ├── generate_cards.py          # Stage 1: Randomized ABAQUS .inp generation
│   ├── abaqus_ml_pipeline.py      # Stage 2: Simulation execution and ODB data parsing
│   ├── train_model.py             # Stage 3: Random Forest training and evaluation
│   ├── comp.py                    # Mathematical comparison and validation logic
│   └── tbc_plots.py               # Logic for generating analytical visualizations
├── data/                          # Fixed input parameters
│   └── materials.json             # Reference physical properties for YSZ and CMSX-4
├── results/                       # Artifact layer (Transient outputs of the logic)
│   ├── dataset.json               # Primary simulation database (512 samples)
│   ├── ml_results.json            # Model metrics and feature importance weights
│   └── ml_results_description.txt # Textual interpretation of simulation outcomes
├── requirements.txt               # Dependency manifest
├── .gitignore                     # Rules to exclude machine-specific/binary bloat
└── README.md                      # Project documentation and usage instructions

---

Material System

YSZ (Yttria-Stabilized Zirconia) — Thermal Barrier Coating

• Thickness: 0.5, 1.0, 1.5, 2.0 mm.
• Properties randomized ±20% per variant:
  • Thermal conductivity (dense or porous curves).
  • Specific heat capacity.
  • Density.

CMSX-4 — Single-Crystal Nickel Superalloy Substrate

• Fixed thickness: 10.0 mm.
• Temperature-dependent properties: Thermal conductivity, specific heat, and density.
• No randomization applied.

---

Simulation Setup

• Solver: ABAQUS/Standard (steady-state heat transfer).
• Element type: DC2D4 (bilinear quadrilateral). The finalized mesh uses 14 total elements (4 for the YSZ layer, 10 for the CMSX-4 substrate).
• Boundary conditions:
  • Hot side (YSZ bottom): T = 1400 K.
  • Cold side (CMSX-4 top): T = 600 K.
• Effective thermal conductivity computed as:

k_eff = (q * L_total) / ΔT

where q is area-weighted average heat flux [W/mm²], L_total is total coating thickness [mm], and ΔT = 800 K.

---

Dataset

• Total variants generated: 512 (4 thicknesses × 128 random samples).
• Successful simulations: 512 / 512.
• Training set: 412 samples.
• Test set: 100 samples.

ML Features

Feature	Description
ysz_thickness_mm	YSZ layer thickness
ysz_density	YSZ density [tonne/mm³]
ysz_k_avg	Average YSZ thermal conductivity [W/mm·K]
ysz_cp_avg	Average YSZ specific heat [mJ/tonne·K]

Target: k_effective — effective thermal conductivity of the full TBC system [W/mm·K].

---

Results Summary

Feature Importance (Random Forest MDI)

Feature	Importance
YSZ k_avg	84.19%
YSZ Thickness	15.62%
YSZ Cp_avg	0.10%
YSZ Density	0.09%

Model Performance

Metric	Training	Test
R²	0.9994	0.9938
MAE	6.50×10⁻⁵	2.06×10⁻⁴
RMSE	1.05×10⁻⁴	3.09×10⁻⁴

Full result details can be found in results/ml_results_description.txt.

---

Usage

Recommended: single entry point

python run_pipeline.py

run_pipeline.py automatically checks whether ABAQUS is installed and routes accordingly:

Situation	What happens
ABAQUS found on PATH	Full pipeline: generate cards → simulate → train model
ABAQUS not found	Stages 1–2 skipped; pre-built results/dataset.json is used for training

No manual configuration required.

---

Manual stage-by-stage execution

-Stage 1 — Generate ABAQUS input files (requires ABAQUS)

python src/generate_cards.py data/materials.json out_dir

Output: out_dir/ containing 512 .inp files and variants_manifest.json.

Stage 2 — Run simulations and extract results Output: results/dataset.json

Stage 3 — Train ML model Output: results/ml_results.json, results/ml_results.png

If running Stage 3 manually without ABAQUS, ensure results/dataset.json exists.

---

Requirements

pip install -r requirements.txt

See requirements.txt for the full list. ABAQUS is required for Stage 2 and is not installable via pip.

---

Units

All quantities follow the ABAQUS consistent unit system (mm-t-s) used in this project:

Quantity	Unit
Length	mm
Force	N
Mass	tonne
Stress	MPa
Thermal conductivity	W/(mm·K)
Specific heat	mJ/(tonne·K)
Temperature	K