Materials Science Research at Vision Research Laboratory

Large-Scale Microstructure Analytics

Physics-aware AI and open data infrastructure for 3D materials microstructure.

Material behavior begins with microstructure: the three-dimensional arrangement of grains, crystal orientations, phases, boundaries, and defects that controls how a component deforms and fails. Our interdisciplinary UCSB team combines advanced materials characterization, mechanics, computer vision, and the BisQue data platform to make these complex measurements easier to acquire, share, analyze, and interpret. The projects below trace that effort from physics-guided image reconstruction and an open 3D dataset to compact symmetry-aware learning and ongoing topology-based tools for connecting microstructure to stress, slip, and crack evolution.

Overview

  • Q-RBSA — high-resolution 3D EBSD reconstruction from sparse sections
  • Ti-6Al-4V 3D Dataset — an open experimental volume and primer for generative models
  • SG-SRAN — compact, symmetry-aware EBSD super-resolution
  • Reeb Atlases — interpretable topology linking microstructure to deformation and failure

Q-RBSA Past · 2023

High-resolution 3D EBSD from sparse sections.

Paper — arXiv 2023 Source Code on GitHub

Three-dimensional electron backscatter diffraction (EBSD) reveals how crystal orientations vary through a material, but repeatedly mapping every serial section is often the rate-limiting step. Q-RBSA addresses that bottleneck by predicting the missing orientation planes between sparsely measured sections, offering a route to larger or more finely sampled experiments without treating crystallographic measurements as ordinary RGB images.

Three-panel Q-RBSA pipeline. Panel A shows electron-beam and EBSD-detector acquisition of serial sections from a three-dimensional specimen. Panel B shows sparsely sectioned 3D EBSD data entering a deep-learning framework and leaving as a denser high-resolution stack. Panel C expands the framework: sparse 2D EBSD sections pass through the Q-RBSA network, predicted orientation planes are assembled into a 3D volume, and the final output is shown as a colorful inverse-pole-figure rendering.
Q-RBSA takes sparsely sectioned 2D EBSD planes, predicts intermediate high-resolution orientation maps with a quaternion residual-block self-attention network, and assembles them into a dense 3D EBSD volume. Figure: Jangid et al. (2023).

Quaternion-valued convolutions learn local relationships in orientation space, while self-attention captures longer-range correlations across the microstructure; the predicted planes are then assembled into a dense 3D EBSD volume. A BisQue inference module also made the workflow accessible through a browser. The project was enabled by collaboration across computer vision, 3D materials characterization, solid mechanics, and scientific data infrastructure.

  • Predicts missing planes between sparse sections
  • Quaternion-valued convolutions in orientation space
  • Self-attention for longer-range correlations
  • BisQue inference module in the browser

Open 3D Titanium Microstructure Dataset Past · 2023

An experimental Ti-6Al-4V volume and primer for physics-based generative models.

Open Dataset in BisQue Paper — Dataset & Primer 2023 3D Visualization Guide

High-quality three-dimensional microstructure data are expensive to collect and remain scarce for machine-learning research. This project released an experimental Ti-6Al-4V volume acquired with TriBeam serial sectioning and EBSD, lowering the barrier for researchers who want to study realistic synthetic microstructures, uncertainty, or rare events without first reproducing a major characterization campaign.

8,893
grains in the released volume
0.6 µm
in-plane resolution
1 µm
serial-sectioning resolution
HDF5
DREAM.3D data format
Left-to-right microstructure data-collection pipeline. A laser serially sections a Ti-6Al-4V specimen; an electron beam records a stack of Kikuchi diffraction patterns; indexing assigns a crystal orientation to every voxel; inverse-pole-figure mapping produces a colorful three-dimensional Ti-6Al-4V orientation volume labeled with its physical dimensions.
The data-collection pipeline follows Ti-6Al-4V from laser serial sectioning and Kikuchi diffraction patterns through indexing into voxelwise crystal orientations and an inverse-pole-figure-colored 3D volume. Figure: Jangid et al. (2023).

The release includes DREAM.3D/HDF5 data, crystallographic orientations, grain labels, processing information, and a browser-based BisQue visualization path. A companion primer explains the material-specific representations and physical constraints needed to build useful generative models. It shows the value of combining materials expertise with open data infrastructure so that complex 3D measurements are findable, inspectable, and reusable.

SG-SRAN Recent · 2026

Compact, symmetry-aware EBSD super-resolution via invariant latent-space learning.

Paper — arXiv 2026 Project Page Source Code on GitHub

SG-SRAN advances EBSD super-resolution by learning in a latent space that respects crystal symmetry. It matches large super-resolution backbones with a tiny fraction of their trainable parameters, making it both scientifically grounded and lightweight enough for practical deployment. Its BisQue integration makes high-quality orientation reconstruction accessible through a browser.

27–49k
trainable parameters
15–16M
parameter backbones matched
FCC + HCP
crystal-structure benchmarks
Unseen
alloys handled without retraining
Two-part SG-SRAN architecture. The upper panel transforms low-resolution EBSD maps through an invariant encoder, low-resolution convolution, routed upsampler, high-resolution refinement, and dictionary decoder to produce four-times upsampled orientation maps. The lower panel expands the upsampler: a local feature-distance cluster is packed into slots, proposal features and subpixel position embeddings enter a router, and the selected owner generates a high-resolution routed patch.
The full pipeline from invariant encoding through routed upsampling and dictionary decoding, together with the local feature-cluster and slot-proposal mechanism used to assign each high-resolution subpixel.

The method is built from three parts, so that every step respects the physics of crystal orientation:

Symmetry-invariant encoder

A frozen, locally isometric encoder maps symmetry-equivalent orientations to the same latent representation.

Feature-consistent local support

Information is aggregated only within consistent neighborhoods, so it is not mixed across grain boundaries.

Dictionary-based decoder

A dictionary-based decoder returns valid crystal orientations.

Across face-centered-cubic and hexagonal close-packed benchmarks, the method achieves the lowest p68 errors and highest inverse-pole-figure fidelity in the reported comparisons, and transfers to alloys not seen during training.

Four-row comparison of Ti-6Al-4V inverse-pole-figure maps labeled Test 0 through Test 3. Ten columns show the low-resolution input, Atindama inpainting, EDSR, HAN, RCAN, SAN, Q-RBSA, QEDSR, SG-SRAN, and the high-resolution target. Each method is shown at a common extent; a 50-micrometer scale bar appears with Test 0, and an HCP IPF-Z color key appears at right.
Ti-6Al-4V IPF-Z maps compare the low-resolution input and learned baselines with SG-SRAN and the high-resolution target across four held-out tests. The common HCP color key and spatial extent make differences in grain interiors and boundaries directly visible.

Reeb Atlases Ongoing

An interpretable bridge from complex 3D microstructures to deformation and failure mechanisms.

Rather than compressing an entire volume into a few global statistics, this ongoing project constructs physically meaningful fields — such as defect proximity, slip response, local misorientation, and distance to a free surface — and summarizes how connected regions appear, merge, and split using Reeb graphs. The goal is to help researchers identify and compare candidate regions for stress or slip localization, crack initiation, and crack-growth pathways across processing conditions.

Detailed two-row architecture diagram. The first row begins with a stack of colorful serial EBSD sections and a reconstructed 3D IPF volume, followed by a shared physical complex containing grains, boundaries, orientations, defects, and surfaces. Two Reeb graphs summarize volume/load-conditioned fields and boundary-network connectivity; a highlighted branch maps back to a marked defect in the EBSD image. The row ends with localized stress, slip, and crack-path analysis verified by CPFE, HR-DIC, and experiments. The second row is a feedback loop linking scientist queries, process-conditioned generation, topological auditing, and prediction and verification.
Proposed architecture for diagnosis, prediction, and intervention. Serial EBSD sections are registered into a 3D orientation volume and a shared physical complex. Volume- and boundary-conditioned Reeb graphs retain exact spatial back-mapping, support mechanism localization, and close an auditable loop through CPFE, HR-DIC, and experimental verification.

The architecture under development has four stages:

Step 1 — Process and data

Serial EBSD sections are registered into a 3D orientation volume, keeping coordinates and provenance from the TriBeam measurements, processing conditions, and mechanical tests.

Step 2 — Physical complex

Grains, the boundary network, phases, orientations, defects, and surfaces are held in one shared spatial object.

Step 3 — Reeb atlas

Volume- and boundary-conditioned Reeb graphs expose persistent structures, bottlenecks, bridges, anisotropy, and cross-field motifs, while exact back-mapping connects every graph feature to a real location in the specimen.

Step 4 — Property and mechanism

Candidate regions for stress or slip localization, crack initiation, and crack-growth pathways are identified and compared, with targeted verification through CPFE, HR-DIC, and experiments.

A planned interactive loop will let scientists propose microstructural changes, generate candidates, and audit them using topology, crystal-plasticity finite-element analysis, and experiments. This collaboration joins 3D microscopy, mechanics, topology, learning, and data infrastructure around an interpretable alternative to black-box voxel prediction.

Contact

B.S. Manjunath
Professor · UCSB Vision Research Lab
manj [at] ucsb [dot] edu

Microstructure analytics is developed in the UCSB Vision Research Lab, led by Professor B.S. Manjunath (Department of Electrical and Computer Engineering, UCSB), in collaboration with Tresa M. Pollock, Samantha H. Daly, and McLean P. Echlin in UCSB materials science and mechanical engineering.