The State of Play

AI-driven microscopy for materials science has consolidated into the leading-edge tier with autonomous systems now operational at scale and agentic control deployed on real scientific infrastructure. Production deployments are routine at research institutions and industrial sites; agentic workflows have moved from theory to functioning platforms with demonstrated real-hardware capability at national user facilities (Stanford/SLAC synchrotron); and vendor ecosystem consolidation is complete with AI as standard functionality. Foundation models and domain-adaptive learning techniques excel at extracting quantitative physical parameters from complex microscopy data—grain boundary kinetics, defect types and concentrations, microstructure topology, crystal structure reconstruction—with minimal manual intervention. The Chinese Academy of Sciences' Aeye-1 autonomous AI-TEM represents the operational inflection point: fully unmanned end-to-end operation from sample transfer through analysis without human intervention, achieving 300× faster image analysis and processing 168 samples per day. Concurrent advances in multimodal AI, generative models for data augmentation, and electron ptychography push capabilities to sub-picometer atomic precision (18 picometer information limit on samples 3× thicker than conventional limits). Deployments now span semiconductor manufacturing (nanometer-scale CMP metrology), aerospace alloy characterisation (1 million precipitate features, reduced from 7+ years to 13 days), orthopedic implant QA (Vision Transformer on 8,493 production SEM images, 90.7% accuracy), pharmaceutical formulation analysis (DigiM FIB-SEM for drug-release prediction on FDA-validated samples), crystal structure reconstruction, and materials discovery. Market growth confirms the shift: microscopy-AI market $1.5B (2025, 15.4% CAGR); electron microscope market $3.17B (9.9% CAGR to $6.13B by 2033), with AI as the primary transformation driver. Yet critical barriers persist despite operational deployments. Agentic LLM-controlled systems (Stanford/SLAC synchrotron case excepted) show severe generalization limits—surrogate models trained on agent benchmarks fail to predict performance on unseen tasks, narrowing real-world applicability to pre-tested workflows. Enterprise AI failure rates (80%+ for projects, 95% for generative-AI pilots), scientific integrity crisis—researchers distinguish AI-fabricated from authentic microscopy images at chance rates (40-51%)—and critical voices from field pioneers (2017 Nobel laureate in cryo-EM warns AI contributions are overestimated and strip scientific creativity) constrain adoption beyond specialized high-value applications and raise reproducibility concerns.

Autonomous systems have reached operational maturity across multiple modalities. Aeye-1 (Chinese Academy of Sciences, Dalian Institute of Chemical Physics) represents the breakthrough: fully unmanned AI-TEM autonomous across complete workflow—sample transfer, imaging, analysis—without human intervention. Performance metrics: 300× faster image analysis than manual, processes 168 samples/day generating 4,000+ images with automatic professional reports including comprehensive microstructural quantification. Deployment: production use in catalyst materials analysis (molecular sieves), energy, and chemical engineering applications. Passed formal scientific achievement evaluation (May 2026) by China Petroleum and Chemical Industry Federation, deemed "highly innovative" and "internationally leading." Agentic AI has demonstrated real-world deployment at national user facilities: Stanford/SLAC synchrotron beamline (SSRL BL17-2) deployed an agentic AI X-ray Scientist (LLM control via Model Context Protocol, Claude Opus) for autonomous single-crystal diffraction alignment, achieving sub-5° orientation error on real hardware with unmodeled clamping errors and noise—confirming autonomous structural analysis capability beyond simulation (Nature Machine Intelligence). Commercial deployments now extend to specialized high-value applications: DigiM Solution deployed FIB-SEM and X-ray microscopy with ML semantic segmentation for pharmaceutical microstructure analysis (PLGA formulations, API distribution, porosity mapping); commercial service across 50+ labs including synchrotron facilities; FDA/University of Connecticut validation studies show capability to predict drug-release profiles without 30-90 day dissolution tests, demonstrating real regulatory/development adoption.

Methodological advances expand capability boundaries. Electron ptychography (Tsinghua University, June 2026) achieves 18-picometer information limit and 0.39-picometer atomic position precision on silicon samples up to 85 nm thick—3× thicker than conventional multislice methods—broadening applicability to complex materials. Multimodal AI frameworks (arXiv, June 2026) integrate image contrast with metadata (composition, beam energy, detector geometry) for atomic-resolution STEM defect classification, achieving 98% accuracy on simulated data and near-human agreement on experimental images. Generative models (latent diffusion) synthesize realistic TEM images with controlled defect labels for data augmentation, improving detection in small-dataset scenarios. Crystal structure reconstruction using diffusion models (Paul Scherrer Institute, June 2026) enables hydrogen position recovery with 97% success rate, correcting database errors and enabling broader materials property simulation. These advances push research-tier methodology toward production deployment.

Vendor ecosystem consolidation accelerates. ZEISS and Leica dominate with entrenched positions; all major vendors now ship AI as standard. ZEISS containerised deployments (Smith & Nephew implant inspection 5-7 min vs. 45-60 min; Festo production workflows); Leica Aivia 15 (Hokkaido University); Thermo Fisher Metrios 6 and Scios 3 FIB-SEM (structural biology workflows); Molecular Devices CellXpress.ai (Emory, UCLA organoid imaging); Bruker Python-based AFM control with real-time defect detection and Q1 2026 >20% organic bookings growth in AI-driven semiconductor metrology; MiViA metallographic grain-size automation. Market size confirms adoption: microscopy-AI $1.5B (2025, 15.4% CAGR); AFM market $569M (2025) → $1.022B (2032, 8.8% CAGR) with AI-based automatic image analysis as key driver; electron microscope market USD 5.64B (2026) → USD 10.80B (2035, 7.5% CAGR) with AI-automation as primary driver; AI-based microscopy software fastest-growing segment at 19.4% CAGR. Tescan-Shimadzu acquisition (July 2026) consolidates 4,000+ installed systems across 80+ countries, signaling sustained vendor commitment.

Production deployments span materials and scales. NIMS Japan: 7+ years → 13 days for aerospace alloy characterisation (1 million precipitate features, ML-enabled automated SEM). Theia Scientific: 43× speedup in TEM grain boundary analysis. Ceramic orthopedics: Vision Transformer deployed on 8,493 production SEM images (5 years in-service), 90.7% accuracy for fracture-origin classification, validates low-magnification pre-screening. AA6xxx aluminum alloys: deep learning on EBSD/SEM/EDS data quantifies microstructural changes in cryogenic friction stir processing. Phase-change memory analysis: unsupervised deep learning reconstructs 3D elemental maps under low-dose conditions, overcoming experimental constraints. Topological defect prediction: deep learning reduces millisecond-scale predictions vs hours for traditional simulation of nematic liquid crystal wrinkling, enabling optical device design optimization. Semiconductor advanced-node fabs: ASML, Applied Materials, and Lam Research deploy AI-powered electron-microscopy-based defect detection in real-time monitoring of 3nm and 2nm processes, integrating automated analysis into fab equipment control workflows. Lawrence Livermore released LIST open-source toolkit for high-throughput SEM nanomaterial analysis. Research institutions routinely deploy domain-adaptive learning: MIT multihead-attention defect classifier (2000 materials, 6 simultaneous defect types, <0.2% sensitivity); ORNL AtomAI for atomic-resolution microscopy; PNNL generalizable grain-boundary segmentation (0.34 µm accuracy). Georgia Tech advancing agentic AI with real-time experiment adaptation. Over 200 publications across 15+ disciplines use desktop SEM; Materials Project database 650,000+ users.

Barriers persist despite operational autonomy and methodological advance. Agentic system generalization remains unresolved: empirical benchmark study (arXiv, August 2026) evaluated LLM-controlled microscopy across 53 benchmark tests (1,949 runs, 105 agent configurations, five LLMs); surrogate models trained on agent architecture and test results failed to predict agent performance on new unseen tasks, qualifying autonomous microscopy applicability to pre-tested workflows only and limiting transfer to novel experimental designs. Scientific integrity crisis unresolved: Nature Nanotechnology surveys confirm researchers distinguish AI-fabricated from authentic microscopy images at 40-51% accuracy (effectively chance); 20-30% baseline errors endemic to standard characterisation. Critical assessment from field pioneers: 2017 Nobel laureate in cryo-EM warns that AI contributions to microscopy-dependent sciences are overestimated and that AI adoption strips researchers of creativity and scientific rigor. Enterprise AI failure rates (80%+ projects, 95% generative-AI pilots) constrain scalability—most organisations cannot justify integration costs for workflows that remain brittle outside controlled conditions. Most industrial deployments narrowly scoped (KIBi building-materials aggregate analysis remains one of few true production transitions; DigiM's pharmaceutical adoption, while commercially successful, remains confined to specialized formulation-characterisation workflow). Meaningful adoption confined to specialized high-value applications (semiconductor metrology, aerospace inspection, pharmaceutical QA, materials research).

• 2019: Research-stage demonstrations of machine learning for grain boundary detection and characterisation; conceptual frameworks for autonomous microscopy emerging from national laboratories; recognition by National Academies of data analytics integration as a critical frontier in advanced microscopy.

• 2020: AI-based image segmentation deployed at major national laboratories (Brookhaven, Oak Ridge, Lawrence Berkeley) for nanoparticle tracking and optimization tasks; 5.8× improvement in scanning probe microscopy via ML-optimized scanning; commercial 3D reconstruction tools (Dragonfly) adopted in industry-academic partnerships; first deep learning frameworks for automated grain segmentation in geological materials; acknowledgment that autonomous TEM systems remain far from fully independent operation.

• 2021: Deep learning defect segmentation in TEM reaches F1 scores of 0.8 for materials characterization; Leica launches Aivia 10 autonomous image analysis software signaling vendor ecosystem maturation; super-resolution microscopy remains limited to expert users with significant automation and cost barriers to routine adoption; field remains solidly at research stage with no production-scale deployments.

• 2022-H1: Georgia Tech and Science journal publish real-time atomic-scale grain boundary visualization using AI tracking; ZEISS launches ZEN Core with integrated AI models for materials analysis; Argonne develops FANTASTX for automating structure extraction from microscopy; ETH researchers reduce STEM analysis time from 30 minutes to seconds using ML. Vendor ecosystem accelerates but critical barriers remain: tool complexity, automation challenges, and cost prevent broader adoption. Field maintains research-stage tier despite increasing institutional support and commercial interest.

• 2022-H2: Nature Machine Intelligence publication of AtomAI open-source framework for atomic/mesoscopic image segmentation; Leica releases Aivia 11 with deep learning segmentation; PNNL/NETL validate CNN-based grain boundary automation in engineering materials; advances in semi-supervised and self-supervised learning for electron and scanning probe microscopy. Research momentum sustained but no evidence of production-scale autonomous systems; adoption barriers in tool complexity, sample preparation, and cost persist.

• 2023-H1: Brookhaven CFN deploys AI-driven autonomous discovery workflows discovering new self-assembled nanostructures; Oak Ridge/University of Sydney advance atom probe crystallography with AI/data mining; FAST toolkit demonstrates autonomous scanning microscopy with <25% sample coverage. Imperial College develops GAN-based artifact removal with no-code GUI; Drexel/AFRL/Johns Hopkins demonstrate automated grain boundary analysis. PNNL raises research integrity concerns about generative AI-fabricated microscopy images. Tool complexity and automation barriers persist despite advancing research capability and growing commercial product ecosystem.

• 2023-H2: Lawrence Berkeley demonstrates autonomous STM with AI/ML for atomic defect mapping in 2D materials; Columbia develops deep learning for grain boundary detection in TEM with statistical validation; University of Illinois deploys generative models (CycleGAN) for synthetic STEM training data. Leica launches Autonomous Microscopy powered by Aivia with 90% detection rates and 70% time savings. Foundation models adapted to materials microscopy but require significant domain expertise. Critical questions emerge about AI generalization beyond training data and preventing AI-generated image misconduct. Research stage maintained despite growing technical capability and commercial product availability.

• 2024-Q1: ORNL and university researchers advance deep learning methods for defect segmentation in 2D materials (MX2 phases) and grain boundary analysis (EBSD). AFM research breakthroughs enable sub-probe resolution on nanoparticles. Constrained matrix factorization methods demonstrate pixel-level grain boundary resolution. Critical reassessment of AI claims in autonomous materials discovery highlights gap between computational predictions and actual synthesis success, tempering some earlier optimism. Research momentum continues but no evidence of production deployment or autonomous operation emerges.

• 2024-Q2: Science journal publication demonstrates AI-enhanced grain boundary characterization discovering new topological phases in titanium. ML frameworks achieve 10,000x speedup in grain boundary segregation prediction for refractory alloys. ZEISS deploys production-level containerized AI models for microscopy across platforms. Autonomous SPM achieves robust room-temperature operation with adaptive defect detection. However, evaluation of LLMs for microscopy reveals constraints in advanced technical design; AI-guided materials discovery work highlights data integration gaps despite successful phase discovery. Field shows vendor ecosystem maturation and incremental capability gains with adoption barriers persisting.

• 2024-Q3: Northwestern ML classifier achieves 95% precision on real-time STEM image classification. MIT computer vision technique demonstrates 85x speedup for materials characterization with 98.5% accuracy. AI denoising enables atomic-resolution TEM observation at 10 ms temporal resolution, revealing previously unobservable nanoparticle dynamics. ZEISS launches VersaXRM 730 with DeepRecon Pro AI module. National Research Council Canada documents 3-year industrial deployment automating aluminum component microscopy with 10x speedup. Peer-reviewed methods for atomic column localization advance automated structural analysis. Temporal resolution and high-throughput capabilities accelerate; adoption barriers and synthesis framework gaps persist. Research stage maintained with no evidence of autonomous production deployment.

• 2025-Q1: ZEISS launches arivis Cloud for cloud-based AI model training in materials microscopy, expanding vendor ecosystem accessibility. ML workflows for grain segmentation in nanoparticles demonstrate 4 images/minute processing with noise robustness. Critical integrity risks crystallized: fraudulent publications proliferate with AI-generated images; expert researchers report inability to distinguish AI forgeries from authentic microscopy data, threatening scientific credibility. LLM agents for microscopy control achieve only 65% success with documented "sleepwalk" safety failures. Independent expert assessment (NREL/PNNL/Purdue) confirms autonomy and robustness barriers remain unresolved. Research capabilities continue advancing but safety, integrity, and autonomy gaps block progression toward production deployment.

• 2025-Q2: Vendor ecosystem expands: Leica releases Aivia 15 with 69% faster 3D processing and accessible deep learning interface; ZEISS demonstrates production containerized deployment across platforms. Research advances in specific methods: dictionary-learning EBSD reconstruction achieves fidelity from 10% probe positions, reducing beam damage. Institutional adoption broadens with Hokkaido University deploying Aivia. Advanced microscopy characterization continues with near-atomic-scale TEM grain boundary studies. Fundamental barriers persist: autonomy unachieved, integrity risks unresolved, tool complexity and cost limit adoption. Field remains research-stage with no evidence of autonomous production systems.

• 2025-Q3: Domain-specific capability breakthroughs: ML-enabled autonomous STEM achieves fabrication of tailored quantum defect structures in 2D materials with precise atomic control; MicroscopyGPT vision-language model trained on simulated STEM data predicts full atomic configurations including lattice parameters and element types from real STEM images; deep learning frameworks enable high-throughput automated perovskite solar cell material segmentation. Integrity crisis intensifies: Nature Nanotechnology commentary documents that experienced researchers still cannot reliably distinguish AI-generated from authentic nanomaterial microscopy images, threatening scientific credibility. Autonomy, tool complexity, cost barriers, and research integrity risks remain unresolved. Field maintains research stage despite advancing domain-specific capabilities and vendor ecosystem maturity.

• 2025-Q4: Workflow automation accelerates: TEM data analysis reduced days-to-minutes via physics-guided AI; Duke ATOMIC platform achieves 99.4% accuracy on 2D materials with zero-shot foundation models. Vendor ecosystem expands with production containerized deployments (ZEISS) showing 10-fold efficiency gains on named customers (Smith & Nephew, Festo). Market growth confirmed: $1.16B (2025) at 15.5% CAGR, forecast $2.04B by 2029. Scientific integrity crisis crystallizes: surveys document experts distinguish AI-generated from authentic microscopy images at chance rates (40-51%); 20-30% baseline errors in standard characterization; fraudulent papers proliferate. Autonomy, integrity risks, tool complexity, and economic barriers remain unresolved. Research stage maintained despite capability and vendor maturity advances.

• 2026-Jan: Data infrastructure expansion: Materials Project database reaches 650,000 registered users, providing AI-ready datasets for materials science applications. Industry analysis confirms energy-transition adoption focus (perovskite/tandem solar, green hydrogen) driving double-digit software spend growth. Integrity challenges persist: perspective papers document AI-generated microscopy images indistinguishable from authentic data, with 20-30% error rates endemic to characterization analyses. Autonomy, integrity, and economic barriers remain unresolved. Research stage maintained.

• 2026-Feb: Capability and vendor ecosystem maturation accelerate: unsupervised ML methods (HDBSCAN) applied to STEM-EDX for multi-component high-entropy materials characterization; GrainBot AI toolkit published in Matter for automated microstructure feature extraction from AFM images with perovskite solar cell validation; SAM2 foundation models adapted for SEM industrial metrology (OPC calibration) with few-shot learning on 60 production images. JPhys Photonics roadmap synthesizes field maturity across detection, segmentation, classification, tracking. Institutional adoption breadth confirmed: 200+ peer-reviewed publications across 15+ disciplines using desktop SEM systems. Yet adoption barriers intensify: consultancy analysis documents 80%+ enterprise AI project failure rates and 95% GenAI pilot failure-to-production scaling rate, reflecting broader economic and organizational challenges constraining materials microscopy deployment. Field maintains research stage with narrowly targeted production deployments amid persistent autonomy, integrity, tool complexity, and economic barriers.

• 2026-Mar to Apr: Agentic AI breakthrough: Cornell University publishes EMSeek (Science Advances, April 2026), a modular multiagent LLM-orchestrated platform for autonomous electron microscopy analysis, achieving ~50-fold speedup (2-5 minutes vs. weeks) and ~90% structural similarity on STEM2Mat benchmark across 20 materials and 5 tasks. Vendor ecosystem expansion: Thermo Fisher launches Metrios 6 automated STEM and Scios 3 FIB-SEM with AI-enabled workflows; Molecular Devices CellXpress.ai adopted at Emory/UCLA. Market growth confirmed: AI microscopy market reached $1.5B (2025), 15.4% CAGR to $6.3B (2035); electron microscope market $3.17B (CAGR 9.9% to $6.13B by 2033) with AI identified as the primary transformation driver. Industrial deployment examples: PNNL generalizable deep learning grain-boundary segmentation validates cross-condition accuracy (0.34 µm grain size MAE); KIBi random-forest system transitions to production for building-materials aggregate characterisation across 27 datasets spanning fine sands to recycled materials. Research publications advance automated analysis further: YOLOv9-based STM molecular image analysis automates counting and morphological measurement; SEMDI-Net deep learning denoising for SEM images of fiber materials achieves 75% grain boundary detection accuracy. Barriers remain: integrity crisis unresolved (researchers distinguish AI-fabricated images at chance rates); enterprise AI failure rates persist (80%+); autonomy limited to controlled conditions. Field advances to bleeding-edge tier with capability and vendor maturity progressing, but adoption scope and production deployment scale remain narrowly constrained.

• 2026-May: Production deployments multiplied across institutions and vendors. NIMS Japan reduced aerospace alloy characterisation from 7+ years to 13 days using ML-enabled automated SEM on 1 million precipitate features. Theia Scientific's YOLO-based production TEM analysis achieved a 43× speedup over U-Net with 3% accuracy on grain size. Lawrence Livermore released the LIST open-source toolkit for high-throughput SEM nanomaterial analysis, and Bruker launched Python-based AI control for AFM with real-time defect recognition and closed-loop automation. MiViA released an AI solution for metallographic grain-size and layer measurement handling colour-etched and twin microstructures. Nature Computational Science published a perspective on agentic AI in electron microscopy addressing autonomous decision-making, and a practitioner survey confirmed AI integration as standard across commercial platforms from ZEISS, Leica, and Olympus—consolidating the field's advance into the leading-edge tier. Research capability continued advancing: MIT and ORNL demonstrated domain-adaptive ML applied to X-ray photon correlation spectroscopy (XPCS) to extract quantitative grain boundary kinetic parameters (diffusivity, stiffness, GB concentration) from nanocrystalline materials; an MIT team published a multihead-attention model (trained on 2,000 semiconductor materials) capable of detecting up to six simultaneous point defect types at 0.2% concentration sensitivity, exceeding conventional detection limits. CNN-based deep learning was applied to full-chip CMP nanotopography prediction from white-light interferometry and AFM data at nanometer-scale accuracy for semiconductor quality control, extending AI microscopy into production metrology.

• 2026-Jun: The field's most prominent new deployment was Aeye-1, the Chinese Academy of Sciences' fully autonomous AI-TEM at Dalian, which passed formal scientific evaluation as "internationally leading": 300× faster image analysis, 168 samples per day, unmanned end-to-end operation without human intervention. Market signals confirmed commercial momentum: Bruker reported >20% organic bookings growth in AI-driven semiconductor metrology for Q1 2026, and the AFM market is tracking $569M (2025) to $1.022B (2032) with AI-based automatic image analysis named as the primary driver. Research capability advanced across multiple methods: Paul Scherrer Institute's XtalPaint diffusion model reconstructed hydrogen positions in crystal structures with 97% overall success rate; Tsinghua's electron ptychography achieved an 18-picometer information limit on silicon samples 3× thicker than conventional methods allow; a multimodal framework integrating image contrast with beam metadata reached 98% accuracy on simulated STEM defect classification with near-human agreement on experimental images; YOLOv5+SegFormer demonstrated superior automated precipitate detection in Cr-based superalloy electron microscopy versus established tools (Weka, ilastik); and a semi-supervised pre-annotation workflow for steel microstructure achieved 78% annotation time reduction (170h→37h), a concrete industrial productivity signal. A Vision Transformer applied to 8,493 ceramic orthopedic implant SEM images achieved 90.7% fracture-origin classification accuracy; LLM-assisted control system engineering for a Scanning Helium Microscope was demonstrated on physical hardware (RSC peer-reviewed), advancing autonomous instrumentation. A multi-institution collaboration (LLNL, UNLV, Stony Brook, UC Davis) deployed ML grain boundary prediction in LLNL's fusion materials programme. Against this deployment momentum, enterprise AI failure rate meta-analysis (RAND, BCG, KPMG, McKinsey, Gartner) confirmed 88–95% of POCs fail to scale and only 5% of custom AI tools reach production—a persistent structural barrier independent of model quality. The 2017 Nobel laureate in cryo-EM publicly warned that AI contributions to microscopy-dependent sciences are overestimated and strip scientific creativity—a credible counterweight to autonomous-systems momentum from Aeye-1 and Georgia Tech's agentic work.

• 2026-Jul: Vendor and infrastructure maturity continued: Bruker shipped production automated AFM/CIPT metrology systems for high-volume semiconductor manufacturing, and Nikon launched an AI-powered industrial microscope (ECLIPSE LV100AMS) with automated grain-size and cleanliness inspection compliant to ISO/ASTM standards. Research advanced on two fronts—agentic AI orchestrating autonomous multi-instrument labs for MOF synthesis, and a 391,606-pair multimodal dataset (MatSciFig) built for materials-microscopy image understanding—while a CAS analysis of 310,000+ publications and a JPS review mapped AI adoption patterns and persistent barriers (data standardization, explainability) across the field. A practitioner newsletter flagged a validation reality-check: fewer than 5% of GNoME materials-discovery predictions have been independently validated. Vendor consolidation continued later in the month with Tescan's acquisition by Shimadzu (4,000+ installed electron-microscope systems across 80+ countries), and advanced-node fabs (ASML, Applied Materials, Lam Research) scaled AI defect detection on electron-microscope images at 3nm/2nm; KAIST used AI-assisted optical/AFM screening across 120,000+ semiconductor flakes to fabricate and analyze 1,615 transistors. A UW-Madison researcher warned that AI-generated microscopy images are increasingly fooling peer review, reinforcing the field's scientific-integrity barrier alongside market forecasts projecting the electron-microscope market to reach $10.80B by 2035.

• 2026-Aug: Autonomous microscopy advances continued alongside sharper limits on generalization: an arXiv benchmark of agentic LLM-controlled microscopy (53 tests, 1,949 runs) found surrogate models fail to predict agent performance on unseen tasks, while MIT/ORNL algorithm-guided electron microscopy achieved picometre-scale atom repositioning, creating 40,000+ stable quantum defects in 40 minutes, and Stanford/SLAC's agentic "AI X-ray Scientist" (Claude Opus via MCP) achieved sub-5° orientation error on a live synchrotron beamline. Vendor and commercial-service maturity advanced: Oxford Instruments added AI-based shadow-masking to its AZtecHKL EBSD software, DigiM's FIB-SEM/X-ray ML segmentation service scaled across 50+ pharmaceutical labs with FDA/UConn validation, and Northwestern published a new AI method combining machine learning with quantum mechanics to predict atomic structures of grain boundaries and interfaces for battery and fuel-cell applications.