From Image to Properties (IM2PROP): Deep Learning for Microstructure Property Prediction

Project page

Code and data to reproduce the ablation study from the paper. IM2PROP predicts Micro-Vickers hardness of S32205 duplex stainless steel from optical microscope images by fusing three branches: a MicroNet-pretrained ResNet50 encoder, a PhaseCNN over the binary phase mask, and the macroscopic phase ratios.

Method, dataset construction and discussion are in the paper; this repository covers reproduction.

flowchart LR
    OM["OM image"]:::io --> ENC["ResNet50<br/>MicroNet, frozen"]:::rgb --> CAT
    MASK["Phase mask"]:::io --> PCNN["PhaseCNN"]:::phase --> CAT
    PCNN --> ATT["Attention<br/>1x1 conv + sigmoid"]:::attn --> ENC
    RATIO["Phase ratios"]:::io --> CAT["Concat 2178"]:::fuse
    CAT --> HEAD["Regression head"]:::fuse --> OUT["Hardness (GPa)"]:::io

    classDef io fill:#d9d9d9,stroke:#666666,color:#000000
    classDef rgb fill:#cfe2f3,stroke:#3d85c6,color:#000000
    classDef phase fill:#d9ead3,stroke:#6aa84f,color:#000000
    classDef attn fill:#f4cccc,stroke:#cc0000,color:#000000
    classDef fuse fill:#d9d2e9,stroke:#674ea7,color:#000000

Loading

Three components are independently switchable — phase ratios (R), spatial attention (A), PhaseCNN features (F) — giving the 8 combinations in the results table. Disabled branches are zeroed rather than removed, so the architecture is identical across all of them.

Install

git clone https://github.com/JayChou04/IM2PROP.git
cd IM2PROP
uv sync

A CUDA GPU is required for a practical runtime. torch and torchvision are pinned to CUDA 12.8 builds from PyTorch's index, which publishes wheels for Linux (x86_64, aarch64) and Windows only. On macOS or a CPU-only machine, remove the [[tool.uv.index]] block and the torch / torchvision lines under [tool.uv.sources] in pyproject.toml, then re-run uv lock && uv sync.

Reproduce

bash reproduce.sh

Regenerates every figure from the 94 source images:

Stage	Step	Time
1	Extract phase masks and phase ratios	~1 min
2	Ablation sweep, 360 runs	~15 h
3	Verify the run matrix is complete	instant
4	Pool cross-validation folds	~1 min
5	Render figures and tables into docs/assets/	~1 min

Resume-safe — completed runs are skipped, so an interrupted job restarts with the same command. --smoke runs a single short job end to end; --dry-run lists the planned runs without executing them. No Weights & Biases account is needed; the sweep logs offline.

Shorter runs

--repeats trades precision for time:

Command	Runs	Time
bash reproduce.sh --repeats 1	120	~5 h
bash reproduce.sh --repeats 2	240	~10 h
bash reproduce.sh	360	~15 h

Repeat seeds are drawn in a fixed order, so a shorter run is a prefix of a longer one: raising the count later reuses the completed runs rather than discarding them. Pass the same --repeats value when checking status, so the completeness gate expects the right number.

Evaluation

With 94 specimens, a single held-out split leaves too few samples for a stable score, so the sweep uses 5-fold cross-validation with 3 repeats:

8 combos x 3 epoch settings x 3 repeats x 5 folds = 360 runs

Check progress or completeness at any time:

uv run scripts/sweep_status.py --report

Results

Cross-validated MAE in GPa, pooled over all 94 specimens and averaged across the 3 repeats.

Combo	R	A	F	30 epochs	50 epochs	80 epochs
1 Baseline	0	0	0	0.2312	0.2227	0.2011
2 +Phase feat	0	0	1	0.2378	0.2197	0.1933
3 +Attention	0	1	0	0.2034	0.2192	0.2261
4 +Attn+feat	0	1	1	0.1872	0.2166	0.1754
5 +Ratios	1	0	0	0.2000	0.1828	0.1827
6 +Ratio+feat	1	0	1	0.2335	0.2275	0.2272
7 +Ratio+attn	1	1	0	0.2030	0.2241	0.1760
8 Full IM2PROP	1	1	1	0.1949	0.2024	0.1807

The lowest MAE is Combo 4 (+Attention +Phase features) at 80 epochs, 0.1754 GPa, against 0.2011 for the RGB-only baseline at the same setting — a reduction of 0.0256 GPa, or 12.7%. Six of the twenty-one combination-by-epoch cells score above the baseline, so the table is best read cell by cell.

Longer training helps consistently: 80 epochs beats 30 in 20 of 24 combination-by-repeat comparisons.

Citation

@unpublished{chou2026im2prop,
  title  = {From Image to Properties: Deep Learning for Microstructure Property Prediction},
  author = {Chou, Shih-Chieh and Cheng, I-Chieh and Li, Bo-Shiuan},
  note   = {Manuscript},
  year   = {2026},
  url    = {https://jaychou04.github.io/IM2PROP/}
}

Supported by the College Student Research Grant, National Science and Technology Council (NSTC), Taiwan, under Grant No. 114-2813-C-110-052-E.

License

MIT — see LICENSE.