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

Shih-Chieh Chou1 I-Chieh Cheng2 Bo-Shiuan Li1

1Department of Mechanical and Electro-Mechanical Engineering, National Sun Yat-sen University
2Department of Materials and Optoelectronic Science, National Sun Yat-sen University

TL;DR: IM2PROP is a tri-branch network that predicts the Micro-Vickers hardness of duplex stainless steel from optical micrographs; its best configuration lowers cross-validated error by 12.7% against an RGB-only baseline.

Framework overview: a micrograph goes through image processing into a phase mask; the micrograph feeds a MicroNet encoder, the mask feeds a PhaseCNN and the phase ratios, and all three meet in a regression head that predicts hardness.
Three complementary branches — a MicroNet-pretrained encoder, a PhaseCNN with spatial attention, and macroscopic phase ratios — are fused to predict hardness.

Abstract

Predicting local mechanical properties from microstructural images is essential for reducing the experimental burden in materials development, yet existing deep learning approaches predominantly rely on synthetic training data, while domain-specific pretraining remains underexplored for image-based property regression. IM2PROP is a tri-branch regression framework that predicts the Micro-Vickers hardness of S32205 duplex stainless steel directly from real optical micrographs, fusing a MicroNet-pretrained ResNet50 encoder, a phase-driven spatial attention mechanism derived from binary phase masks, and macroscopic phase ratios. Its 94 samples are cropped with the expanding cavity model, so each image covers the plastic zone the indent actually deformed. A full-factorial ablation evaluated with 5-fold cross-validation over 360 training runs shows that the best configuration reaches a mean absolute error of 0.175 GPa, 12.7% below the RGB-only baseline.

Sizing Each Sample

A hardness value describes only the material the indent deformed, so each micrograph is cropped to that region. The expanding cavity model of Mata et al. treats the indentation as a spherical cavity inflating in an elastic–perfectly-plastic solid, which turns the indent’s two measured diagonals into the radius of its plastic zone at the surface. Averaged over all 94 indents that radius is 121 µm, so every sample is a 720 × 720 px square centred on the indent, large enough to hold the whole zone. The micrographs deliberately leave out the indent itself, whose size alone would give away the hardness. The steps below follow Appendix A of the paper for one representative indent.

A Vickers indent drawn to scale, its equivalent contact circle, the expanding plastic zone and the 720 by 720 pixel crop that contains it, ending on the real micrograph from the paper’s Fig. 4. d1 = 87.35 µm d2 = 87.53 µm as = 34.88 µm R = 54.94 µm Cs = 122.5 µm mean Cs = 121 µm 720 × 720 px Paper Fig. 4
Worked example from Appendix A: d1 = 87.35 µm, d2 = 87.53 µm, F = 1 kgf.

From Micrograph to Phase Mask

Six image-processing stages — median denoising, background normalization, CLAHE contrast enhancement, Otsu binarization, morphological refinement and small-region filtering — turn each micrograph into a binary map of its two phases. The mask feeds the PhaseCNN branch, and its pixel counts give the phase ratios. The three specimens below sit at the low, typical and high end of the austenite fraction; drag any divider to compare all three.

Phase mask of specimen HO_48. Optical micrograph of specimen HO_48. Micrograph Mask
HO_48 · 2.66 GPaaustenite 48.4% · ferrite 51.6%
Phase mask of specimen HO_38. Optical micrograph of specimen HO_38. Micrograph Mask
HO_38 · 2.42 GPaaustenite 62.8% · ferrite 37.2%
Phase mask of specimen HO_73. Optical micrograph of specimen HO_73. Micrograph Mask
HO_73 · 2.42 GPaaustenite 77.5% · ferrite 22.5%

austenite (FCC, light phase) ferrite (BCC, dark phase)

Architecture

Three branches read the same specimen at different scales. Three components can be switched off independently — phase ratios (R), phase attention (A) and PhaseCNN features (F) — while the rest of the network stays identical: a disabled branch is zeroed rather than removed, and a disabled attention gate passes the encoder’s features through unchanged. That gives the eight configurations in the results.

Detailed architecture: ResNet50 encoder on the RGB image, PhaseCNN on the binary phase mask with a phase-attention gate, and phase ratios, concatenated to 2178 dimensions and passed to a regression head.

  • MicroNet encoder. A ResNet50 pretrained on micrographs and kept frozen; 2048-d RGB features.

  • PhaseCNN (F). Three convolution layers on the phase mask; 128-d phase features.

  • Phase attention (A). A 1×1 convolution and sigmoid that re-weights the encoder’s feature map.

  • Phase ratios (R). Austenite and ferrite fractions; 2-d.

  • Fusion. Concatenated to 2178-d and mapped to hardness in GPa by a five-layer head.

Where the Model Looks

The phase-attention gate turns the PhaseCNN’s reading of the phase mask into a map between 0 and 1 that re-weights the encoder’s features. Each slide puts that map beside the micrograph it was computed from. The three specimens have the lowest, median and highest measured hardness among the 19 held-out specimens of one cross-validation fold of the best configuration (Combo 4, 80 epochs), and each map’s colors are stretched to its own range, so compare regions within a slide rather than across slides.

Optical micrograph of specimen HO_92.Micrograph
Phase-attention map over specimen HO_92.Phase attention
HO_92 · measured 2.27 GPa

LowHigh attention

1 / 3

Results

Mean absolute error in GPa from 5-fold cross-validation repeated three times, pooled over all 94 specimens. Lower is better.

# Configuration R A F 30 epochs 50 epochs 80 epochs
1Baseline (RGB only)offoffoff0.23120.22270.2011
2+ Phase featuresoffoffon0.23780.21970.1933
3+ Attentionoffonoff0.20340.21920.2261
4+ Attention + phase featuresoffonon0.18720.21660.1754
5+ Phase ratiosonoffoff0.20000.18280.1827
6+ Ratios + phase featuresonoffon0.23350.22750.2272
7+ Ratios + attentionononoff0.20300.22410.1760
8Full IM2PROPononon0.19490.20240.1807

R phase ratios A phase attention F PhaseCNN features

Lowest error: + Attention + phase features at 80 epochs, 0.1754 GPa — 12.7% below the RGB-only baseline at the same setting (0.2011 GPa).

Chart of cross-validated MAE for all eight configurations at 30, 50 and 80 epochs.
Cross-validated MAE for every configuration and training length, as rendered by reproduce.sh.

Reproduce

One script recomputes every number in the results table, and the chart, from the 94 source images — about 15 hours on a single CUDA GPU, or about 5 with --repeats 1. Full instructions are in the README.

git clone https://github.com/JayChou04/IM2PROP.git
cd IM2PROP && uv sync
bash reproduce.sh

Paper

The latest version lives in the GitHub repository and is updated as the work evolves.

All nine pages of the paper; opens the PDF on GitHub.

BibTeX

@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/}
}

Acknowledgements

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

Page design inspired by Zip-NeRF, BakedSDF, Adaptive Compliance Policy and ZipMap.