PowderBench

Physics-based synthetic data for SEM powder characterization.

PowderGen uses 3D particle scans and rigid-body simulation to create multi-modal synthetic powder scenes representative of real SEM images. PowderBench contains these scenes, and paired multi-modal labels, enabling improved model performance and expanded characterization capabilities in scientific computer vision.

Rendered scene of spherical powder particles SEM-style translation of the spherical particle scene Rendered scene of irregular powder particles SEM-style translation of the irregular particle scene
Rendered ScenesSEM-style Translations

Synthetic SEM Images for Model Training

540synthetic scenes
9powder types
334real SEM benchmark images
4label modalities

One simulated scene, multiple labels

Each scene pairs rendered and SEM-style images with labels that are difficult to obtain from a single micrograph. The dataset also includes amodal masks and 3D particle geometry.

From rendered scenes to SEM-style images

Rendered synthetic powder scene with enlarged detail
Rendered scene
Ground-truth particle instance masks with enlarged detail
Instance masks
Translation using the rendered image alone, with enlarged detail
Render-only translation
Multi-channel translation with enlarged detail
Multi-channel translation

Render-only translation uses the rendered image; multi-channel translation also uses an instance overlay, surface normals, ambient occlusion, and depth to help preserve particle boundaries.

Better segmentation with fewer real labels

mAP@50:95

AP@95

Real onlyReal + renderedReal + translatedRendered onlyTranslated only

Mask R-CNN performance on 67 held-out real SEM images, comparing real-only training with training supplemented by 540 synthetic PowderBench images. Synthetic data provides the largest gains with fewer real training images. Error bars show the reported variation across three training seeds.

Beyond visible particle masks

Synthetic particle with visible and occluded regions highlighted as an amodal mask

Amodal segmentation

Estimate the full projected particle shape, including occluded regions.

Synthetic scene with color-coded surface depth labels

Depth estimation

Predict visible surface depth from a single image.

Synthetic 3D particle scene with one particle highlighted

3D reconstruction

Recover particle geometry from a single view.

These scenes provide amodal masks, visible-surface depth maps, and 3D particle geometry. Results are evaluated quantitatively on held-out synthetic scenes, with small expert assessments on real images for amodal segmentation and depth; real-image 3D performance remains unverified.

Depth predictions on real SEM images

Real SEM micrograph used as model input
Real SEM image
Depth prediction from a pretrained model
Pretrained Depth Anything V2 Small
Depth prediction after PowderBench fine-tuning
After PowderBench fine-tuning

Depth Anything V2 Small was fine-tuned on PowderBench and assessed on real micrographs using a blinded expert comparison. These are qualitative predictions, not measured depth; the study assessed 10 real SEM images.

How label precision affects segmentation

Estimated curves

mAP@50:95

Estimated mAP at 50 to 95 across label-jitter levels for rendered and translated data on real and synthetic tests

AP@95

Estimated AP at 95 across label-jitter levels for rendered and translated data on real and synthetic tests
Rendered · real testRendered · synthetic testTranslated · real testTranslated · synthetic test

Synthetic masks are perturbed by 0–32% IoU deviation. Real-image test performance remains relatively stable through 4% deviation, then declines as masks become less precise.

Project status

Results from Stephen Price, Danielle L. Cote, and Elke A. Rundensteiner, PowderGen: Physics-Based Data Generation for Multi-Modal SEM Powder Characterization, Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2027. Research conducted while at Worcester Polytechnic Institute.

Discuss a scientific imaging project.

Talk with Atomix about microscopy, synthetic data, or model evaluation.

Discuss a project