Submissions

Hackathon Submissions

Project	Description	Links	Team
Image-Driven Analysis of Bubble Dynamics and Foam Stability ID: H-001	This project presents a fully image-driven and chemistry-agnostic analysis of foam stability using time-resolved microscopy images. Interpretable bubble-scale features—such as population dynamics, coarsening rates, size heterogeneity, and spatial organization—are extracted to quantify and rank foam stability. An explainable machine-learning model demonstrates that early microscopic dynamics can reliably predict long-term macroscopic foam behavior.	Code Video Doc	Hesham Eina Abdalla
Contrastive Micrograph-Metadata Pre-Training ID: H-002	We implement CLIP ( Contrastive Language-Image Pre-Training) but for HAADF-STEM images and paired metadata vectors. We show that it can effectively find a shared embedding space where metadata that fits to the style of of an image is embedded to vectors that have high cosine similarity to the embedded vector of the image. Similarly, metadata combinations that do not fit with the image style are pushed away. We also discuss the usefulness of the learned embedding space for a possible lightweight physics-informed denoiser	Code Video Doc	Henrik Eliasson Angus Lothian
Automated Crystalline Domain Segmentation in Polycrystalline TEM Images via Machine Learning-Based FFT Analysis ID: H-003	This project introduces a fully automated pipeline for crystallographic analysis of HRTEM images, effectively eliminating the subjectivity and inefficiency of manual FFT peak selection. By utilizing Gaussian Mixture Models (GMM) for automated signal classification and integrating them with DBSCAN clustering, we enable consistent, high-throughput segmentation of crystalline domains. This workflow offers a widely accessible and reproducible alternative to resource-intensive 4D-STEM techniques for characterizing complex heterogeneous nanomaterials.	Code Video Doc	Woojin Bae Jinho Rhee Shihyun Kim
Alexa for AFM ID: H-004	We demonstrate how to operate a microscope with voice control, by connecting a speech-to-text module with an instrument application programming interface.	Code Video Doc	Gunstheimer Corte-León
Towards Automated Materials Analysis: Deep LearningDenoising and Phase Identification from 4D-STEM ID: H-005	This project develops a deep learning-assisted 4D-STEM pipeline that denoises low signal-to-noise diffraction patterns and automates phase identification, enabling reliable detection of O1 and O3 phases in NMC811. By making nanoscale phase mapping robust to complex background noise and weak reflections, it supports systematic investigation of structural deviations linked to degradation in Ni-rich layered cathodes.	Code Video Doc	Ethan Lo Fanzhi Su
stem-denoising-hackathon ID: H-006	Deep learning for ultra-low dose STEM denoising	Code Video Doc	Avital Wagner Willem de Kleijne Jay te Beest Akshaya Kumarjaishankar
AFM-SPARK ID: H-007	AFM-SPARK is a system-specific super-resolution framework for Atomic Force Microscopy (AFM) imaging. The method reconstructs a global high-resolution (HR) AFM image from a fast low-resolution (LR) scan and a small number of selectively acquired HR patches, significantly reducing acquisition time while preserving nanoscale fidelity.	Code Video Doc	Youngwoo Choi Seungwhan Ryu Junho Yang Sanggil Park Jihui Won Chaeyul Kang
LLM-Assisted Structural Interpretation of Microscopy Images ID: H-008	Recent advances in machine learning and large language models (LLMs) offer new opportunities to assist scientific interpretation. However, directly applying LLMs to microscopy images risks hallucination and loss of scientific rigor if not carefully constrained by quantitative evidence. This project addresses this gap by designing a lightweight, explainable pipeline that combines classical image analysis with evidence-grounded LLM-ready interpretation, without relying on opaque end-to-end black-box models.	Code Video Doc	Shyam Sundar Debsarkar Iman Chatterjee Ugochukwu Philip Ochieze
MASK-ViT: Mask-Free Nanoscale Feature Detection ID: H-009	MASK-ViT leverages self-supervised Vision Transformer embeddings and saliency-driven patch pooling to automatically detect and characterize nanoscale features in microscopy images without pixel-level annotations. The approach enables scalable, annotation-efficient materials characterization through embedding-based analysis and unsupervised discovery.	Code Video Doc	Gowtham Nimmal Haribabu Gaaurav Lath
Smartscan ID: H-010	SmartScan is an AI-powered adaptive AFM scanning system that uses machine learning to dynamically optimize scan parameters (speed, resolution, force) in real-time, achieving 37% faster scanning while simultaneously improving image quality by balancing thermal drift against tracking error. The system learns from physics-based simulations on real AFM data to discover that faster scanning (10-12 µm/s) actually produces better results than traditional conservative speeds (5 µm/s), a counterintuitive finding that challenges conventional AFM operation practices.	Code Video Doc	Abdulrhman Mohamed Ahmed Mahmoud Syed Ahmed Khaderi
Digital Bubble Stability tracking microscope ID: H-011	Digital foam simulator: Random Forest Regressor (Aiscia Platform) predicts bubble metrics from lab inputs, then nearest-neighbor mapping returns realistic microscope frames/animations—no live microscope needed.	Code Video Doc	Aziza Mohammed Ahyen Mostofa Dr. Usman Chaudhry Yasmin Abdelkarim Souhil Sid Rehaan Hussain Dr. Fadwa El Mellouhi Dr. Harris Rabbani
The Gold Standard for low-dose STEM ID: H-012	This project presents a deep-learning workflow for ultra-low-dose STEM imaging, where low dwell times introduce severe scan artifacts that limit image quality. By combining Fourier-domain pre-processing with a U-Net trained on real paired data and a custom loss that preserves high-frequency detail, the method aims to remove these artifacts while maintaining resolution at low doses.	Code Video Doc	Jay te Beest Avital Wagner Willem de Kleijne Akshaya Kumar Jaishankar
AIScientist4Microscopy ID: H-013	AIScientist4Microscopy is a proof-of-concept “AI Scientist” workflow tailored to microscopy. The goal is to demonstrate an end-to-end pipeline that can (1) generate credible, research-grade hypotheses in AI-for-microscopy, (2) translate each hypothesis into a runnable experimental plan with code, and (3) produce research artifacts (figures, metrics, and paper-style writeups) in a way that is easy to demo and extend.	Code Video Doc	Adib Bazgir Mahule Roy Yuwen Zhang
VLRIMM - Vision-Language Retrieval for Identical Materials Morphology ID: H-014	VLRIMM is a multi-modal RAG pipeline designed to bridge the gap between raw visual data and textual knowledge, transform a single micrograph into an interactive dialogue with global current scientific knowledge by aligning visual morphology with textual research. Pairing state-of-the-art foundation models (Meta's DINOv3 and OpenAI's text embedding) and RAG, VLRIMM is designed for the future of Self-Driving Labs. It provides an autonomous link between observed microstructures and peer-reviewed information like synthesis method, eliminating the knowledge lag and manual research bottlenecks.	Code Video Doc	Kevin Zhang Sartaaj Khan Mohammad Taha Thomas Pruyn
Unsupervised Microstructure Image Analysis and Question Answering Using CLIP and VLMs ID: H-015	We explore automated understanding of materials microscopy images using multimodal AI models. By combining scalable zero-shot classification with CLIP and reasoning-based analysis using vision language models, our approach identifies imaging techniques and material categories from real-world microscopy data, enabling efficient and interpretable microscopy analysis at scale.	Code Video Doc	Subham Ghosh Shubham Tiwari Ayush Kumar Pandey Pranjul Chandra Bhatt
MCP Servers for Theory - Experimental Matching ID: H-016	Used existing FerroSIM simulation and developed an MCP server to control FerroSIM through agents, and then combine this with the digital twin AFM investigating a ferroelectric sample to perform some rudimentary theory-experiment matching through digital-digital agentic AI workflows	Code Video Doc	Vani Nigam Achuth Chandrashekhar
MARVL: Materials-Science Aware Reasoning Dataset for Vision–Language Models ID: H-017	We introduce MARVL, a comprehensive, materials-science-aware reasoning dataset designed to train and benchmark vision–language models on microscopic imaging tasks across AFM, SEM, and TEM modalities. Using automated literature mining and curated experimental data, MARVL captures real imaging conditions, tip or instrument-induced artifacts, and scientifically grounded explanations. This dataset enables next-generation VLMs to perform accurate feature interpretation with transparent reasoning, expanding AI capabilities across nanoscale materials characterization.	Code Video Doc	Mohd Zaki Indrajeet Mandal Prince Sharma Megha Mondal Sudhakar Kumar Vishal Bhaskar Arjun Chand Shivnandi
µStack: An AI-Powered Platform for Atomic Surface Analysis and Microscopy Simulation ID: H-018	µStack is an open-source, AI-driven platform that simplifies atomic surface structure research by combining machine learning potentials, multi-agent workflows, and natural language interfaces into a unified system. The platform allows researchers to generate, relax, and analyze atomic surfaces through conversational queries, automatically producing visualizations, standardized data outputs, and AI-generated summary reports. It supports multiple microscopy simulation modalities (STM, AFM, IETS) and is designed for applications in catalysis, semiconductor research, and 2D materials science.	Code Video Doc	Aritra Roy Kevin Shen Ben Blaiszik Piyush Ranjan Maharana
Latent-Constrained Autoencoder for 4D-STEM Clustering ID: H-019	This project applies unsupervised deep learning to identify crystalline and amorphous domains in 4D-STEM diffraction data from metallographic samples. By training a rotation-invariant autoencoder to compress diffraction patterns into low-dimensional latent representations and clustering them with HDBSCAN, the method robustly detects grain boundaries while reducing spurious cluster fragmentation. Incorporating physically motivated rotational invariance significantly improves boundary localization and accurately captures the finite width of grain boundaries in experimental data.	Code Video Doc	Martin Eriksen Joaquín Otón Mauricio Matta
HAADF-Segment: Bridging Supervised Segmentation and Unsupervised Anomaly Detection for Automated Defect Analysis in 2D Transition Metal Dichalcogenides ID: H-020	This project establishes an automated deep learning pipeline for HAADF-STEM analysis, utilizing a custom U-Net architecture to perform precise multi-class segmentation of atomic species in 2D transition metal dichalcogenides. By coupling this supervised model with unsupervised manifold learning techniques, the framework enables the robust identification and quantification of point defects and structural instabilities without relying on exhaustive manual labeling.	Code Video Doc	Cobi Allen Mostafa A. Mostafa Nannan Zhang nz257@exeter.ac.uk Navid Haghmoradi
4Denoising ID: H-021	We adapted state-of-the-art unsupervised deep-learning model (UDVD) to work on 4DSTEM datacubes. To train the model we created semi-synthetic low-dose data by modifying open-source datacubes. Our approach is generalizable to other types of high-dimensionality data streams, and after training can potentially be deployed on live applications.	Code Video Doc	Leonardo Cancellara William Talbott Ankit Shrivastava Jordi Weingard Dingqiao Ji Ian MacLaren
DeepScan Pro: Intelligent Microscopy Agent ID: H-022	DeepScan Pro is an interactive software pipeline that transforms passive scanning electron microscopes (SEM/STEM) into intelligent discovery agents. Addressing the inefficiency of blind raster scanning, it utilizes a Multi-Stage Active Learning architecture.	Code Video Doc	Md Habibur Rahman
Anomaly Detection and Clustering of Atomic-Resolution STEM Images Using Semantic Segmentation ID: H-023	This project develops a machine-learning framework for automated defect detection, localisation, and clustering in atomic-resolution HAADF-STEM images. It combines supervised image classification, weakly supervised U-Net segmentation, and unsupervised clustering to identify multiple, coexisting defect types without pixel-level annotations. Demonstrated on CdTe and SrTiO₃ datasets, the framework achieves high classification accuracy and produces quantitative, material-independent heatmaps, providing a scalable and reproducible workflow for high-throughput, data-driven electron microscopy analysis.	Code Video Doc	Claudia Sosa Espada Surabhi Gunjur Sathish Thomas Karagiannis
U-Net-based pipeline for high-throughput quantification of crack growth during in situ TEM tensile testing ID: H-024	This project analyzes crack growth and propagation in materials using computer vision and deep learning techniques applied to microscopy images. The system tracks crack tip positions, measures crack geometry parameters (length, area, width, CTOD), and monitors temporal evolution across multiple experimental datasets captured at different magnifications.	Code Video Doc	Vivek Devulapalli
Integrating AFM and STEM Digital Twins with LLMs for Automated Data Interpretation ID: H-025	This project introduces Microscopy-LLM, a framework that bridges Claude AI with AFM and STEM digital twins via specialized MCP servers. It enables researchers to perform automated image segmentation and data interpretation through natural language, transforming complex binary datasets into an intuitive, conversational workflow. This 'AI co-pilot' effectively lowers the technical barrier to advanced materials characterization and provides a scalable blueprint for the future of autonomous laboratories.	Code Video Doc	Josep Cruañes Giner Fanzhi Su
Graph-O-Foam: ActiveScan Copilot ID: H-026	The Graph-O-Foam project forecasts foam stability by converting lab data into synthetic microscopy frames and extracting bubble metrics with computer vision. Energy-based models (EBM) showed top accuracy, physics-informed transformers balanced realism and predictions, and a Streamlit dashboard enables real-time visualization and analysis.	Code Video Doc	Vedashree Yemula Mohamed Bouchekouf Param Trimbake Mohammed Tanvir Muntasir Mahmud
Phase Mapping Ni-based alloys with EBSD Using Deep Learning ID: H-027	We applied convolutional neural networks to classify EBSD patterns into different Ni-based alloys. Both contrastive and cross entropy learning were tested, and their accuracies were compared with each other.	Code Video Doc	Alfred Yan Gabriel Trindade dos Santos Ramandeep Mandia Roberto dos Reis Vinayak Dravid
Proof of Concept for Automated STEM Tilting Using Image Analysis for Quantifoil Grids ID: H-028	Quantifoil grids with circular holes provide an interesting test case for automated tilting in STEM - the holes of the grid will appear elliptical when mistilted, and perfectly circular when perpendicular to the beam direction. The goal of this project was to develop an algorithm to simulate the automated tilting of a Quantifoil grid by examining a simulated micrograph, quantifying the circularity of the hole, and using tilt to optimize for maximum circularity	Code Video Doc	Elizabeth Heon Sai Venkata Gayathri Ayyagari Sita Sirisha Madugula Darel Pates Lynnicia Massenburg Andrew Balog
CAVA: A Causal Analysis and Validation Agent for Microscopy-Informed Materials Discovery ID: H-029	CAVA is an agentic causal reasoning framework designed to support microscopy-informed materials discovery by integrating data-driven causal analysis with literature-based validation. The system first identifies robust causal relationships in microscopy-derived data using established causal discovery algorithms and regime-aware consensus, then employs a large language model agent to contextualize and validate these relationships against prior scientific studies. By combining automated causal discovery with agent-based literature reasoning, CAVA provides a transparent and interpretable foundation for causal understanding and decision-making in experimental materials science.	Code Video	Kamyar Barakati Elyar Tourani Vivek Chawla Mohammad Amin Moradi
eJect ID: H-030	We developed a digital twin STEM framework to simulate atom manipulation experiments, then validated the approach on a WS₂ monolayer using a ThermoFisher Spectra 300. Survival analysis of atoms under point and blast irradiation revealed both ballistic and non-ballistic damage contributions, highlighting differences between simulated and experimental behavior.	Code Video Doc	Austin C. Houston Dominick L. Pelaia Levi M. Dunn
RHEED universal translator ID: H-031	RHEED Universal Translator is a bidirectional deep-learning tool that links in-situ RHEED diffraction patterns with ex-situ XPS stoichiometry for SrTiO₃ thin films. It learns a shared latent representation to predict stoichiometry from a RHEED image or generate a RHEED-like pattern from a target stoichiometry, and it’s packaged in a Streamlit no-code app with pretrained weights for easy demos and extension.	Code Video Doc	Asraful Haque Jawad Chowdhury Sumner Harris
Bio-STEMGPT: An AI Assistant for TEM Data Analysis ID: H-032	We present Bio-STEMGPT, a chatbot that integrates segmentation tools from a user's chosen segmentation model to help users identify features in TEM images and explore the resulting segmented data. We envision this as a tool for scientists integrating TEM data into their projects who may be unfamiliar with analyzing this type of data.	Code Video Doc	Bridget Vincent Arielle Rothman Saurin Savla
Identifying the Spatial Domains of Sm-doped BiFeO3 using STEM and Machine Learning ID: H-033	Spatial Domains of Sm-doped BiFeO3 using various Sm concentrations and window sizes are investigated using STEM, principal component analysis, k-means clustering, and supervised machine learning. The results suggest that the ideal window size is 40, with the domains being well distinguished for both k-means clustering and supervised machine learning using random forests. These methods have the potential to create a more efficient and automated method to determine the best window size to distinguish between spatial domains of ferroelectric materials.	Code Video Doc	Fatima Anwar Hayden Dennison Julie Schlanz Emily Stump
SPM–(S)TEM correlative studies for nanoscale characterisation and data curation for implied transfer learning. ID: H-034	In electron microscopy and Scanning probe microscopy, the scarcity of curated training data and the low throughput of atomic-resolution techniques often limits the adoption of supervised learning methods.Although automation is improving through manufacturer-provided software interfaces, e.g. ‘Autoscript’ from Thermofisher scientific in electron microscopy, and digital twin software, human oversight remains essential during operation. An approach that would enable a combinatorial approach to generating sufficient labelled data could be through transfer learning. Here we present a framework for correlative SPM–(S)TEM studies based on physics-informed digital twins.	Code Video Doc	Timothy Lambden
Angle-Dependent Morphologies of Ferroelectric Domain Walls ID: H-035	The morphology of ferroelectric domain walls with different characteristic angles were analyzed by neural networks and computer vision algorithms. Neural networks could differentiate between high-angle (>130 degree) and lower-angle (<130 degree) domain walls from their morphology. Further, higher-angle domain walls were found to prefer a curved morphology, while lower-angle domain walls prefered a straight morphology, in line with expectations for ferroelectric–ferroelastic domain walls.	Code Video Doc	Laurel Washburn Ehtesham Anwar Sho Watanabe Aidan Cotton Kaurab Gautam Nabin Khanal Ralph Bulanadi
Moiré Fringes in High-Resolution Transmission Electron Microscopy ID: H-036	Our project involved separating HRTEM moiré fringes into two their two lattice layers. This was done using a machine learning model that predicted what the two layers should look like based on the moiré fringe.	Code Video Doc	Yuval Noiman Hank Perry Jackie Cheng
Automated Detection of Ice Artifacts in Cryo-EM Using Machine Learning ID: H-037	Cryogenic electron microscopy (Cryo-EM) enables structural analysis of biological macromolecules by reconstructing three-dimensional structures from two-dimensional images of rapidly frozen samples. However, Cryo-EM micrographs suffer from low signal-to-noise ratio, poor contrast, and contamination from ice artifacts, making it difficult to distinguish true biological signal from noise. To address this challenge, we propose an automated ice artifact detection approach using a deep convolutional neural network.	Code Video Doc	Tharushi Rajaguru Shehani Kahawatte Elif Dursun Diana Oliveira
Automated Region-of-Interest Detection in Microscopy Images Using VGG-16 and Patch-Based Classification ID: H-038	This project presents a patch-based deep-learning framework for automated region-of-interest detection in atomic-resolution STEM images. Using a VGG-16–based model, the approach combines supervised defect classification with unsupervised clustering to identify defect-bearing regions, extract their spatial coordinates, and enable scalable, reproducible, and automated microscopy analysis across different material systems.	Code Video Doc	R. A. W. Ayyubi Shoaib Masood Fatemeh Karimi Zahira El Khalidi Jaeyeon Jo
Automated Defect Detection in STEM Images Using Machine Learning ID: H-039	We developed an ensemble deep learning framework for automated defect detection in high-resolution scanning transmission electron microscopy (STEM) images of CdTe and SrTiO₃ (STO). The approach follows patch-level classification using custom convolutional neural networks (CNNs) trained on frequency-enhanced image representations to identify crystalline and defect regions.	Code Video Doc	Noah Holt Joshua Marvin Ramji Subedi Kamal Khanal Mohd Tauhid Khan
Machine learning–enabled SEM-driven surface morphometrics ID: H-040	Recently, a classical computer-vision approach for quantitative SEM analysis, defining microcracks as a feature class and estimating crack count, length distribution, and area within the field of view, are applied. To streamline analysis, the framework was extended with neural‑network segmentation, improving generalization, reducing preprocessing demands, and enabling additional pixel‑level crack detection alongside classical approach. For this purpose, plasma treated cobalt sputtered samples were imaged at multiple fields of view to demonstrate the implementation.	Code Video Doc	Vineet Kumar Yogesh Paul Himanshu Mishra PhD. Saurabh Sudan
Machine Learning Denoising of Reciprocal Space Maps for realistic center-of-mass evaluation ID: H-041	This project utilizes the self-supervised Noise2Self method to denoise reciprocal space maps from nano-XRD experiments. It enables rapid denoising and more accurate center-of-mass calculations without requiring clean reference data. The approach effectively reduces noise while preserving subtle structural features. Further optimization is needed for very weak diffraction signals, such as those from quantum wells	Code Video Doc	Leonardo Oliveira Soroush Motahari Navyanth Kusampudi Kartik Umate Meizhong Lyu
EM-Caddie ID: H-042	EM-Caddie is a web-based, no-coding-required platform that enables microscopy researchers to analyze and process images using ML tools. Users can describe desired operations in plain English, and the system applies pre-trained models for tasks like super-resolution, segmentation, FFT analysis, and line profile extraction, all through an interactive interface. The platform streamlines workflows, reduces reliance on specialized software or scripts, and supports extension with additional models and tools.	Code Video Doc	Jessica Gerac Kohen Goble Kyle Hollars
Automated particle detection and quantitative analysis from electron microscopy images ID: H-043	We developed NanORange, an AI-assisted workflow that automatically detects nanoparticles in noisy cryo-TEM images by combining denoising/contrast enhancement, adaptive thresholding, and instance-aware boundary separation to prevent merged detections in crowded regions. The segmented particles are then fit with circles/ellipses to extract quantitative metrics and generate particle-size distributions for datasets. This pipeline improves consistency across variable-contrast micrographs, and outputs analysis-ready tables for rapid, reproducible vesicle statistics.	Code Video Doc	Seyed Aref Golsorkhi Mohammad Javad Raei Frances Joan Alvarez Marc Mamak
3-D Membrane Reconstruction from a Single 2-D FIB-SEM Micrograph ID: H-044	This project introduces an open-source computational framework that statistically reconstructs 3D membrane microstructures from a single 2D FIB-SEM slice using Gaussian Random Fields. This tool enables rapid, understandable prediction of critical transport properties like permeability and pore connectivity without the need for expensive 3D imaging.	Code Video Doc	Antonio M. Lancuentra Bill Yinqi Wang Ahmad Ali Thomas Zhang
GridScope: LLM assistant for automated microscopy ID: H-045	GridScope is an AI-powered automation platform for Scanning Transmission Electron Microscopy (STEM) that bridges the gap between experimental design and instrument execution. Researchers describe imaging objectives in natural language—such as 'acquire a 5×5 grid at 3 µm spacing' or 'explore tilt angles from 0° to 60°'—and receive executable Python scripts validated against a physics-based Digital Twin.	Code Video Doc	Shuchi Sanandiya Alexander Pattison Sanchit Bansal
iPotNET ID: H-046	This project introduces a physics-informed deep learning framework to solve the inverse scattering problem in Scanning Transmission Electron Microscopy (STEM). By fusing visual detector data with physical metadata (thickness, rotation) into a SwinUNETR backbone, the model reconstructs quantitative atomic electrostatic potentials with high fidelity (>54 dB PSNR). This approach explicitly mitigates non-linear dynamical scattering artifacts, significantly outperforming traditional iDPC methods which depends on ideal condition and assumptions on optically thick and tilted crystalline samples.	Code Video Doc	Haipei Shao Sridurgesh Ravichandran Zeyu Wang Mattia Lizzano Andrea Cicconardi
Nanoscale Structure Disentanglement-KFUPM ID: H-047	NanoscaleAnalyzer is a React component that provides a UI for uploading microscopy images and running a simulated ML pipeline to extract nanoscale structural features (grains, domains, defects, and optional STM spectroscopy). It is built with React and uses Tailwind CSS for styling and lucide-react for icons.	Code Video Doc	Abbas Adamu Abdullahi Aminu Rabiu Doguwa Mauliady Satria Dr Abdurahman Aliyu Abubakar Dahiru Shuaibu Ahmad Abbas Dalhatu Sara Isabel Gracia Uribe Dr Mariah Batool
RAG for Microscopy Data Analysis ID: H-048	Developed Retrieval Augmented Generation (RAG) for Microscopy data analysis. Here the RAG agent interfaces with process-specific programs and generates code based on user-specified prompts.	Code Video Doc	Ganesh Narasimha Zijie Wu
Learning Atomic Defects Without Real Data: YOLO Trained on Fully Synthetic STEM Imagery ID: H-049	We present a framework for generating fully synthetic STEM images of atomic lattices containing vacancy, interstitial, and grain boundary defects and using this data to train an object detection model. A YOLO-based detector trained exclusively on synthetic data accurately identifies defects in experimentally acquired STEM images, achieving up to 99.5% accuracy on labeled test data and strong qualitative performance on additional datasets. This approach enables real-time, scalable defect detection while reducing reliance on manually labeled experimental data.	Code Video Doc	Victoria Augoustides Jed Doman Mahmoud Hawary Tatiana Proksch Jingyun Yang
Cell Checker ID: H-050	My project develops a basic cell checker app that analyzes a users AFM data and evaluates if the surface indicates a health surface to culture cells of the selected type.	Code Video Doc	Dale Herzog
Interpretable Digital Twins for Autonomous STEM Aberration Correction ID: H-051	We present a machine-learning-assisted framework for automated aberration correction in STEM, addressing the nonlinear and strongly coupled nature of corrector tuning that limits conventional, operator-dependent workflows. By combining LLM-based log parsing, symbolic regression, a corrector digital twin, and reinforcement learning, the framework learns aberration–response relationships to enable faster, more stable, and reproducible correction.	Code Video Doc	Yingheng Tang Kang’an Wang Haozhi Sha Juhyeok Lee Peter Ercius
Multi-Pass Preprocessing for Robust Hysteresis Loop Fitting in Ferroelectric Materials ID: H-052	Hysteresis loops are experimental signatures representing the characteristics of ferroelectric materials. Consequently, extracting useful features from the hysteresis loops proves to be crucial in describing the ferroelectric properties and its switching behavior. In this study, we aim to improve the hysteresis loop fitting using a 50x50 hysteresis loop dataset on a PbTiO3 thin film, acquired by piezoresponse force spectroscopy.	Code Video Doc	Mingxin Zhang Thanakrit Yoongsomporn Hardik Tankaria Sivakorn Kanharattanachai
Combined Approaches for Drift Correction and Domain Dynamics Analysis in AFM Imaging ID: H-053	Atomic Force Microscopy is inherently a time-consuming imaging technique, with data acquisition times on the order of minutes. This makes the instrument sensitive to small positional drifts caused by thermal fluctuations, mechanical perturbations, piezoelectric scanner instabilities or smaller mechanical perturbations of the AFM head or sample. Therefore, this project is aimed at distinguishing between changes due to instrumental error (location) drift and sample dynamics, as well as drift correction.	Code Video Doc	Aadarsh Kumar Yevhen Brych
Score-Based Super-Resolution for Atomic-Scale MoS2 Imaging ID: H-054	We propose Score-Based Super-Resolution (SBSR), a conditional diffusion framework for atomic-scale reconstruction from low-dose HAADF-STEM images. By combining physics-informed degradation modeling with score-based generative learning, our approach restores high-fidelity atomic lattice structures while reducing reliance on high electron doses. This enables more robust imaging of beam-sensitive nanomaterials without introducing artificial structures.	Code Video Doc	Xinyuan Wang Mingli Huang Jingyuan Sun Fanzhi
Uncertainty-Driven AFM Sampling with LLM Guidance ID: H-055	AFM provide amazing detail, but they are incredibly slow. Scanning every single pixel is inefficient, especially when large parts of a sample might be flat or empty. Our motivation was simple: can we scan fewer lines—saving time—but still get a perfect image?	Code Video Doc	Shuting Xie Wenyi Yao
Defect Classification for 2D-Material STEM Datasets ID: H-056	This project presents a physics-guided, machine learning–based framework for automated identification and classification of atomic-scale defects in HAADF-STEM images of Janus MoWSSe. By combining interpretable Z-contrast descriptors with a two-stage convolutional neural network, the approach enables scalable defect localization, classification, and statistical analysis across large experimental datasets.	Code Video Doc	Vikas Reddy Paduri Nirmal Singh Prabhat Prajapati Mehran Yasir Vinayak Srivastava Karishma Begum Abinava Yeshwanth K.J
Mapping 2D-flexibility and large conformational transitions from HS-AFM with parsimonious data-driven models ID: H-057	The problem we tackle in this hackathon is to develop a lightweight and self-contained ML tool that helps dissect, interpret and identify both 2D-flexibility and large conformational transitions of biomacromolecules. In addition, HS-AFM could require external control and/or deep knowledge of the system, which we attempt to substitute with a conceptual CG model that reflects interfacial dynamics with surfaces of different hydropathy. The difference between our simulations and existing fitting tools, is that in our case the effects of the surface are included, and the defined observables are reduced to 2D. Hence, tackling a 2D-flexibility of domains is possible.	Code Video Doc	Assit. Prof. Horacio V. Guzman M.Sc.c. Ian Addison-Smith M.Sc.c. Celica Krigul Ph.D.c. Willy Menacho
VantaScope 5090 Pro ID: H-058	Deep learning ensemble with Fuzzy Logic and Explainability Layers that automatically characterizes graphene samples with linguistic and human-understandable outputs. The tool then provides predictive analytics on the material properties based on the structure.	Code Video Doc	Haidar Bin Hamid Juhyeon Park
MCP for simulator ID: H-059	MCP for simulator	Code Video Doc	Guanlin He Yongwen SUn
Automated Atom-Resolved Defect and Element Classification in 2D MoWSSe HAADF-STEM Dataset ID: H-060	2D material defect classification and identification	Code Video Doc	Cheng-Yu Chen Swarnendu Das George Hollyer Pawan Vedanti
IQ-X: Image Quality Assessment via Cross-Modal Validation ID: H-061	IQ-X is a reference-based, physics-informed framework for assessing AFM image quality using SEM as a stability baseline.	Code Video Doc	Aditya Raghavan Chiranjib Chakrabarti
ANCHOR: Registration by Alignment ID: H-062	We developed an automated, landmark-based workflow for aligning multi-round microscopy datasets. The method uses cellular landmarks to match features across imaging rounds, fits smooth axial mismatches, and assigns local confidence scores to indicate alignment reliability. This approach enables reproducible integration of multiple imaging rounds while avoiding common failures of commonly used manual or global registration methods.	Code Video Doc	Saven Denha Dima Traboulsi Jonah Wilbur
MicroSeg Lab: One-Shot Microscopy Segmentation with LLM-Guided Hybrid Refinement ID: H-063	MicroSeg is a one-shot microscopy segmentation pipeline that produces instance and union masks from a single uploaded image. It runs fast classical segmentation first, then uses review-gated SAM refinement and optional LLM planning to improve results only when needed. Reference images/masks and minimal user hints provide a lightweight way to steer segmentation toward the intended phase on hard cases.	Code Video Doc	Shakti P. Padhy Sushant Sinha Chase Katz
RONIN - Ronchigram based Optical Neural Inference for aberration detectioN ID: H-064	Aberration correction is critical for achieving high spatial resolution in scanning transmission electron microscopy (STEM). Conventional correction relies on expert interpretation of ronchigrams, making the process slow and difficult to automate. In this project, we present RONIN, a physics-informed deep learning framework that predicts dominant electron-optical aberrations directly from ronchigram images. Using synthetically generated ronchigrams and a ResNet-based regression model, RONIN demonstrates accurate inference of several low-order aberrations, highlighting its potential for closed-loop and autonomous microscope alignment.	Code Video Doc	Sriram Sankar Manikandan Sundararaman Aditya Raghavan
Autonomous Identification of Metal Microstructural Features via Latent Space Mapping-Based Microscope Control ID: H-065	We develop an autonomous microscopy framework that uses machine learning to identify metal microstructural features directly during EBSD acquisition, without prior knowledge or human intervention. By encoding Kikuchi diffraction patterns into a latent space and adaptively guiding data acquisition, the approach efficiently resolves multiscale microstructural heterogeneity while minimizing the number of required measurements. This work provides a foundation for autonomous microstructure mapping and data-driven alloy design.	Code Video Doc	Pierre BELAMRI-REGENPIED Martin COURTOIS Mathieu CALVAT Neal BRODNIK JC STINVILLE Henry PROUDHON
MicrosCopilot: An Agentic, Physics‑Aware AI Framework for Confocal Microscopy ID: H-066	The project is a “Microscopilot” that combines classical image processing, particle tracking, and a configurable digital twin of Brownian motion with large language model agents that guide the user through the full workflow: simulating or loading movies, detecting and tracking particles, extracting physical parameters (such as diffusion coefficients), and explaining the results in plain language for materials and tribology researchers. The system is organized as multiple cooperating agents (data loading, detection/tracking, physics analysis, and explanation) behind a simple UI so that experimentalists can quickly go from raw z‑stacks or time‑series images to quantitative, reproducible results and narrative reports during the hackathon.	Code Video Doc	Abhishek Gupta Samuel Hee
Generative Topographical Interpretation of AFM Data ID: H-067	Nano-Constellations is an automated pipeline that reconstructs raw AFM data into high-fidelity 3D topographical maps of TiO2 surfaces. By implementing asymmetric cropping to eliminate edge-noise and utilizing Delaunay Triangulation for structural reconstruction, we transform raw TiO2 intensity data into a 3D geometric web. The results demonstrate a high-fidelity 3D reconstruction that allows for perspective flipping and real-time visualization of interatomic strain.	Code Video Doc	Sinny J. Trivedi
TwinSpec: A Digital Twin Framework for GIWAXS Data, Geometry, and Physics-Aware ML ID: H-068	TwinSpec is a modular digital twin framework for grazing-incidence wide-angle X-ray scattering (GIWAXS) that integrates a data digital twin, an interactive lab console, and a geometry-driven visual twin. By converting literature-derived GIWAXS measurements into reproducible, instrument-aware representations and coupling them to an interactive control surface, TwinSpec enables exploration of how experimental geometry and processing conditions influence scattering outcomes without requiring synchrotron access. A lightweight, physics-aware ML module demonstrates that the resulting data infrastructure is directly usable for downstream machine learning workflows. https://www.twinspec.org	Code Video Doc	Tajah Trapier

Use the resources below to prepare your final hackathon submission.
Each section includes links and instructions to help you navigate the process.

---

📦 Access the Hackathon Dataset

Find all microscopy datasets used in the Digital Twin Microscope and the Hackathon. This includes raw STEM data, metadata, and other related files.

Open Dataset

🧪 Digital Twin Microscope – Demo Notebooks

Explore the Digital Twin Microscope through interactive notebooks. These show how to simulate scans, load data, and work with the digital twin environment.

Open Demo Notebooks

📁 Preparing Your Data For Submission

Use this notebook to properly format your datasets and prepare files for submission. It demonstrates how to create clean, well-structured datasets from raw microscopy data.

Open Preparation Notebook

---

📬 Need Help?

For more details or assistance with the hackathon datasets, please contact:

Rama Vasudevan: vasudevanrk@ornl.gov

Utkarsh Pratiush: upratius@vols.utk.edu