MatNexus

MatNexus: AI hub using Gemini 3 to turn XRD/SEM into physics-grounded models. Automates phase ID, CIF generation, synthesis loops, and literature mining to accelerate discovery from lab.

Comment

Inspiration

As a Master's student in Computational Materials Science, I noticed a persistent "characterization bottleneck" in the laboratory. Experimentalists often spend a lot of time correlating surface-level morphology from Scanning Electron Microscopy (SEM) with atomic-scale data from X-Ray Diffraction (XRD). I have always wanted to build something that could solve this-something beyond the scientific Machine Learning scripts. As a beginner in this area (dev), I wanted to challenge myself to build a interactive web application that could put power of senior scientist into pocket of every researcher. MatNexus was born from desire: to turn specialized training into seconds of AI-driven insights.

What it does

MatNexus is a multimodal Material Informatics Hub the automated the transition from raw lab data to physics-grounded models. It features:

  • The Lab Debugger : A vision-integrated engine that analyzes XRD and SEM images simultaneously to identify material phases, space groups, and habit.
  • Autonomous Discovery Loop : It compares experimental data against the Material Project API to detect lattice discrepancies and provide actionable synthesis advice.
  • 3D Crystal Visualizer : Instantly renders unit cells and simulated synthetic powder diffraction patterns for experimental verification.
  • Literature Miners : Uses high-context indexing to extract structured technical data from dense research PDFs.

How I built it

The project is built on a foundation of Python-based scientific libraries and cutting-edge AI:

  • Gemini 3 Flash : Serves as the reasoning backbone, handling multi modal vision tasks and 1M+ token context for literature mining.
  • Pymatgen : The core physics engine used for crystallographic analysis and diffraction simulation.
  • Material Project API : Integrated as scientific Ground truth for validating AI-generated structural predictions.
  • Streamlit : Powers the UI, and enables seamless deployment.

Challenges we ran into

The most significant challenge was Structural Data Integration. I had to learn how to parse raw text outputs from a LLM into Crystallographic Information Files (.CIF) that a physics engine could read. Additionally, ensuring the Multimodal Vision accurately interpreted the physics within a noisy plot was difficult. I had to refine prompts to ensure the AI understood the relationship in Bragg's law: $$n\lambda = 2d \sin \theta$$ where d is interplanar spacing. Getting AI to "see" the physics across different scales required extensive iterative testing.

Accomplishments that I am proud of

I am proud of building a platform that feels like a professional workstation. Seeing the 3D Unit cell render for the first time-knowing it was generated from a raw image of a diffraction pattern-felt amazing. Furthermore, as someone new to application development, creating a development-ready product that securely handles API secrets.

What we learned

This is first hackathon for me and it has been a masterclass in multimodal prompt engineering and UI/UX design for scientists. I learned that technical sophistication is nothing without human relevance; the way you present data is just as important as the accuracy of prediction itself. I also discovered the power of Gemini 3 Flash, which allowed me to balance processing speed with deep scientific reasoning.

What's next for MatNexus

I think this is just beginning. The roadmap includes:

  • Integration of TEM/STEM data : Expanding the vision engine to include atomic-resolution microscopy images.
  • Real-time Lab Connectivity : Connecting MatNexus directly to lab for on-the-fly characterization during experiments.
  • Advanced DFT Proxies : Moving beyond density predictions to full electronic Density of States estimations.

Built With

Share this project:

Updates