Lifetime Sample Tracking (LiST): A Data Platform for Materials Science

The Lifetime Sample Tracking (LiST) platform is a comprehensive data management system developed by the 2DCC-MIP to automate the curation, analysis, and dissemination of diverse materials science data, thereby enabling closed-loop synthesis-design iterations and supporting machine learning research for approximately twenty thousand samples.

The Gist

Imagine a massive, bustling kitchen where thousands of chefs are trying to invent the perfect new recipe for a dish called "2D Materials." These aren't just any dishes; they are ultra-thin layers of atoms used in advanced electronics and science.

The problem? Every chef (researcher) has their own way of cooking. Some use a high-tech oven (MOCVD), others use a vacuum chamber (MBE), and some bake them on different types of pans (substrates). In the past, if a chef made a great dish, they wrote down the recipe in a fancy notebook and published it. But if they burned the toast or made a bland soup, they threw it in the trash and never mentioned it. This made it hard for other chefs to learn what not to do, and it was impossible to know exactly how a specific dish was made years later.

Enter LiST: The Ultimate Recipe Book and Kitchen Manager

The paper introduces LiST (Lifetime Sample Tracking), a digital platform built by the 2D Crystal Consortium (2DCC-MIP) at Penn State. Think of LiST as a super-smart, automated kitchen manager that solves the chaos of the experimental kitchen.

Here is how it works, broken down into simple parts:

1. The "Black Box" Recorder (Automated Data Capture)

In a normal kitchen, a chef might forget to write down that they added a pinch of salt at 2:00 PM. In the LiST kitchen, the ovens and mixers are connected to a computer.

• How it works: When a scientist starts growing a material, a "Gatekeeper" computer automatically grabs the recipe, the temperature, the timing, and the specific ingredients from the machine. It doesn't rely on human memory.
• The Analogy: It's like a smart fridge that automatically logs every time you open the door, what you put inside, and exactly how long it stayed there, so you never have to guess later.

2. The "Full Life Story" (Tracking the Sample)

LiST doesn't just care about the final dish; it cares about the entire life of the ingredient.

• The Analogy: Imagine tracking a tomato from the moment it was a seed, through the soil it grew in, the truck that carried it, the washing it received, and finally how it was chopped. LiST does this for atoms. It tracks the "substrate" (the base material) just as carefully as the final product.
• Why it matters: If a material has a weird property, scientists can look back in LiST to see if it was caused by the raw ingredients, the oven temperature, or a scratch on the pan.

3. The "Good and Bad" Library (FAIR Data)

Most science books only show the "Five-Star Reviews" (successful experiments). LiST is different. It stores everything, including the "burnt toast" (failed experiments).

• The Analogy: If you are trying to learn to bake, knowing that "adding too much flour at 300 degrees causes a fire" is just as valuable as knowing the perfect recipe. LiST saves these negative results so computers can learn from them.
• The Result: This creates a massive library of "Findable, Accessible, Interoperable, and Reusable" (FAIR) data.

4. The "Digital Passport" (DOIs and Data Packages)

When a scientist publishes a paper, they usually just say, "We made a great sample." LiST gives that sample a Digital Passport (DOI).

• The Analogy: Instead of just saying "I baked a cake," the scientist hands you a QR code. When you scan it, you see the exact recipe, the brand of flour used, the oven model, and a photo of the cake.
• The Benefit: This makes science transparent. Anyone can verify the work, and if they want to try it themselves, they have the exact instructions.

5. The "Crystal Ball" (Machine Learning)

Because LiST has so much data (about 20,000 samples and thousands of failed attempts), it is perfect for teaching Artificial Intelligence (AI).

• The Analogy: Imagine showing a robot chef 10,000 photos of cakes, including the ones that collapsed. The robot learns to predict exactly how to bake a perfect cake by spotting patterns humans might miss.
• What the paper found: The researchers used LiST data to train AI to:
  • Look at a photo of a material and guess what temperature it was baked at.
  • Predict how much of a surface is covered by crystals just by looking at a picture.
  • Even guess what a material sounds like (spectroscopy) just by looking at its texture (microscopy).

Summary

The paper describes LiST not as a magic wand that invents new materials instantly, but as a powerful organizational tool. It turns the chaotic, messy process of experimental science into a clean, digital, and searchable history. By recording every step, saving every failure, and connecting the data to the final publication, LiST helps scientists move from "guessing and hoping" to "designing and knowing."

It is currently being used by the 2DCC and is expanding to help other research centers, acting as a bridge between the messy reality of the lab and the clean, organized world of data science.

Technical Summary

Technical Summary: Lifetime Sample Tracking (LiST): A Data Platform for Materials Science

Problem Statement

The experimental materials science community faces significant challenges in data management, hindering the transition from trial-and-error research to "materials-by-design" as envisioned by the Materials Genome Initiative (MGI). While computational repositories (e.g., AFLOW, Materials Project) and narrowly defined experimental databases (e.g., ICSD) exist, the broader experimental landscape lacks a unified organizational hierarchy. Key obstacles include:

• Complexity and Heterogeneity: The vast array of material compounds, physical forms, and synthesis techniques (ranging from bulk growth to MOCVD and MBE) defies standardization.
• Data Provenance and Defects: Unlike theoretical models that assume ideal crystal structures, real material properties are intimately tied to defects and specific synthesis histories. Capturing these details is difficult due to inconsistent human record-keeping and proprietary instrument software.
• Fragmentation: Research often occurs in small, short-term collaborations where record-keeping does not prioritize long-term community needs. Failed or sub-optimal results are frequently discarded, depriving machine learning (ML) and AI algorithms of the "negative data" necessary for robust training.
• FAIR Principles Gap: There is a critical need to make experimental data Findable, Accessible, Interoperable, and Reusable (FAIR) to support future re-use and autonomous synthesis.

Methodology and System Architecture

The 2D Crystal Consortium Materials Innovation Platform (2DCC-MIP) developed the Lifetime Sample Tracking (LiST) platform to address these challenges. LiST is a modular, extensible data engine designed to capture the full lifecycle of a material sample, from substrate preparation to final publication.

Core Architecture Components:

1. LiST Application and API:

   • Front-end: A React-based web application using Next.js for fast data binding.
   • Back-end: A .NET framework server with an SQL database.
   • Authentication: Supports Microsoft Entra ID and Shibboleth InCommon Federation, issuing JSON Web Tokens (JWT) for secure access. It supports anonymous read-only access to public data and authenticated access to private/collaborative data.
   • API: A RESTful API allows programmatic access to metadata and datasets, supporting metadata-based queries (composition, synthesis parameters) and integration with external ML workflows.

2. Automated Data Collection:

   • Gatekeeper System: To accommodate Windows-based instrument software and maintain network isolation, labs utilize a local "Gatekeeper" computer. Instruments transfer logs and data to this local node, which then syncs to Microsoft SharePoint cloud storage via the Microsoft Graph API.
   • Ingestion: The LiST server ingests files and metadata from SharePoint, indexing them in the SQL database.
   • Standardization: The system enforces standardized naming conventions and sample IDs. Synthesis software integration prompts users for metadata (material, substrate, user) at the start of a run, automatically assigning unique IDs. Characterization data is organized into folders labeled with these IDs.

3. Data Processing and Simulation:

   • Integration: The platform integrates theoretical modeling tools, including ReaxFF reactive force field parameter sets and first-principles DFT simulation results.
   • Workflow: Work is ongoing to integrate Globus pipelines to pull high-performance computing (HPC) simulation data, ensuring configuration files are preserved for reproducibility.

Key Functional Features:

• Data Packages: Users can group data used in publications into "Data Packages" that are archived in the Penn State ScholarSphere repository and assigned a Digital Object Identifier (DOI). This allows for the inclusion of raw data and growth recipes in the methods sections of papers.
• Instrument Identification: LiST integrates with DMR-FIRST, a tool that assigns persistent identifiers to scientific instruments, enabling the tracking of instrument productivity and impact similar to ORCID for researchers.
• Sample Provenance: The system tracks the full history of a sample, including substrate origin, vendor details, polishing/cleaving activities, growth parameters, and all subsequent characterization.

Key Results and Data Scale

As of the time of writing, the LiST platform has captured data on:

• Sample Volume: Approximately 20,000 unique material pieces derived from ~15,000 growth events.
• Material Diversity: Over 1,000 unique materials and heterostructures, including Transition Metal Dichalcogenides (TMDs like MoS2, WSe2, WS2), topological insulators (Bi2Se3), iron-based superconductors, and 2D magnets.
• Characterization Data: The system hosts diverse datasets, including ~5,700 RHEED measurements, ~5,200 AFM images, ~2,350 Raman/PL spectra, and ~1,900 electrical transport measurements.
• Inclusivity: The system explicitly records "bad" or non-ideal growths, providing the necessary negative data for training predictive models.

Applications in Data-Driven Discovery

The paper demonstrates the utility of LiST through several machine learning applications:

• Feature Detection (Transfer Learning): A Convolutional Neural Network (CNN) trained on AFM images of MOCVD-grown MoS2 successfully classified growth temperatures (900°C, 950°C, 1000°C). Surprisingly, weights pretrained on ImageNet (macroscopic objects) outperformed those from MicroNet, suggesting classification relied on texture rather than specific microscopic features.
• Crystal Growth Quantification: Regression models applied to WSe2 micrographs outperformed semantic segmentation models in determining crystal coverage percentages.
• Sequential Transfer Learning: A sequential fine-tuning approach (e.g., training on 2 classes, then adding a 3rd) demonstrated better generalization for classifying multiple TMDs compared to single-step training, regardless of dataset volume.
• Cross-Modal Learning: Unsupervised representation learning (ResNet + UMAP) and regression models successfully predicted AFM features from Raman spectra and generated complete Raman/PL spectra from AFM data, revealing mutual information between morphological and spectral data.
• Anomaly Detection: Multimodal analysis of GaSe samples using RHEED embeddings and regression models successfully identified growth anomalies consistent with traditional rocking curve analysis.
• Counterfactual Analysis: Using synthetic counterfactuals generated via PCA and ResNet, the system quantified how deposition order and flux ratios in MBE-grown SnSe affected isotropy and grain density.

Significance and Claims

The paper positions LiST not merely as a database, but as a foundational infrastructure for a "closed-loop" materials development process. Its significance lies in:

1. Enabling FAIR Data: By automating capture and enforcing standards, LiST makes complex experimental data reusable and interoperable, addressing the "black box" nature of experimental synthesis.
2. Supporting ML/AI: By preserving both successful and failed experiments, LiST provides the comprehensive datasets required to train robust ML models for materials design and autonomous synthesis.
3. Bridging the Loop: The integration of synthesis, characterization, theory, and publication (via Data Packages and DOIs) creates a continuous feedback loop, moving the field toward "materials-by-design."
4. Community Utility: The platform is designed to be modular and extensible, allowing it to adapt to the specific needs of diverse research groups and facilities beyond the 2DCC, as evidenced by its adoption by the 3DFeM2 and Q-NEXT centers.

The authors conclude that while LiST is currently a core tool for the 2DCC, its architecture and philosophy offer a scalable model for the broader materials science community to organize the "chaotic landscape" of experimental data.