MC3D: The Materials Cloud computational database of experimentally known stoichiometric inorganics

This paper introduces MC3D, a comprehensive online database on the Materials Cloud portal containing experimentally known stoichiometric inorganic crystal structures with fully reproducible, DFT-optimized geometries derived from automated workflows and curated input protocols.

The Gist

Imagine you are a chef trying to invent the perfect new dish. You have a massive library of recipes from around the world, but they are written in different languages, some have missing ingredients, and some are just theoretical ideas that no one has ever actually cooked.

MC3D is like a super-organized, high-tech kitchen that takes all those messy, scattered recipes and turns them into a single, reliable "Master Cookbook" of real, edible dishes.

Here is the story of how they built it, explained simply:

1. The Great Recipe Hunt (Data Collection)

The scientists started by gathering nearly one million recipes (crystal structures) from three giant libraries: the COD, ICSD, and MPDS.

• The Problem: These libraries were messy. Some recipes had typos, some were for dishes that don't exist in reality (like a cake made of pure air), and some were just theoretical sketches.
• The Cleanup: They ran these recipes through a strict "quality control" machine. They threw out the ones with bad syntax, the ones that weren't balanced (stoichiometric), and the ones that were just molecular clouds (like water vapor) rather than solid materials.
• The Result: From that million, they kept only 72,589 unique, solid, real-world recipes. This is their "Source Collection."

2. The Simulation Kitchen (DFT Optimization)

Now, they had the recipes, but they needed to test them. In the real world, you have to bake a cake to see if it rises. In the computer world, they use a method called Density-Functional Theory (DFT). Think of this as a "virtual oven."

• The Challenge: Running a virtual oven is expensive and slow. If you put a complex cake (a structure with many atoms) in, it might take days to bake, or the oven might break (crash).
• The Solution: They built an automated robot chef (an automated workflow).
  • The robot puts the ingredients in the oven.
  • If the oven starts smoking (an error), the robot doesn't panic. It checks the smoke, adjusts the temperature, and tries again.
  • If it fails too many times, it gives up and moves to the next recipe.
• The Success Rate: This robot was incredibly good. It successfully "baked" 85% of the recipes it tried, even when the original recipes were tricky.

3. The Master Cookbook (The MC3D Database)

The result is the MC3D (Materials Cloud 3D Structure Database).

• It contains 32,013 perfectly "baked" structures.
• Why is this special? Most other databases are like a pile of unsorted notes. MC3D is a curated, consistent collection. Every single entry was cooked using the exact same "recipe" (computational protocol). This makes it perfect for training AI. If you want to teach a computer to predict new materials, you need a clean, consistent dataset, not a messy pile of notes.

4. The Open Kitchen (Accessibility)

The best part? They didn't lock the cookbook in a vault.

• They put it on the Materials Cloud, a public website.
• The Interface: Imagine a website where you can type in "I want a material with Iron and Oxygen," and it instantly shows you a list of 3D models you can spin around, rotate, and inspect.
• The "Receipt": For every single calculation, they kept a full "receipt" (provenance). If you want to know exactly how they got a result, you can click a button and see the entire history: which computer they used, what settings they tweaked, and how they fixed errors. This makes the science completely transparent and reproducible.

Why Does This Matter?

Think of Machine Learning (AI) in materials science as a student trying to learn chemistry.

• If you give the student a messy pile of contradictory notes, they will learn the wrong things.
• If you give them the MC3D, you are giving them a clean, verified textbook.

This database helps scientists:

1. Find new materials faster: Instead of guessing, they can search this database for materials that might be super-strong, super-conductive, or great for batteries.
2. Train better AI: The consistency of the data helps computers learn the rules of chemistry more accurately.
3. Save time: Researchers don't have to waste time cleaning up messy data; they can just start their experiments.

In short: MC3D is a massive, clean, and open library of "digital crystals" that has been rigorously tested by a robot chef, making it the perfect foundation for discovering the next generation of materials that will power our future.

Technical Summary

1. Problem Statement

While Density Functional Theory (DFT) is a cornerstone of computational materials discovery, existing high-throughput databases often suffer from several limitations:

• Inconsistency: Different databases may use varying computational setups (functionals, pseudopotentials, protocols), leading to small but significant inconsistencies across the periodic table. This hinders the training of accurate machine learning (ML) models.
• Data Provenance: Many databases lack full transparency regarding the computational history of their data, making reproducibility difficult.
• Source Bias: Major databases (e.g., Materials Project, OQMD) rely heavily on the Inorganic Crystal Structure Database (ICSD) and often miss structures found in other repositories like the Crystallographic Open Database (COD) or the Materials Platform for Data Science (MPDS).
• Experimental vs. Theoretical: Distinguishing between experimentally verified structures and purely theoretical predictions is often unclear in large datasets, complicating the selection of starting points for experimental validation.

2. Methodology

The authors developed a fully automated, reproducible pipeline to create the Materials Cloud Three-Dimensional Structure Database (MC3D). The workflow consists of three main stages:

A. Data Import and Curation (MC3D-source)

• Sources: Data was imported from three major repositories: COD (revision 213553), ICSD (version 2017.2), and MPDS (version 1.0.2), totaling ~901,210 structures.
• Filtering Pipeline:
  1. Parsing & Validation: CIF files were cleaned and validated. Invalid syntax or inconsistent information led to the discard of ~178,000 files.
  2. Stoichiometry: Non-stoichiometric structures (containing partial occupancies) were removed (~245,000 structures) as they require complex supercell treatments outside the study's scope.
  3. Deduplication: Structures were normalized to primitive cells and compared using chemical formulas and space groups. ~197,000 duplicates were removed.
  4. Molecular Crystal Exclusion: Structures containing hydrogen that were unique to the COD were removed (~208,000 structures) to exclude molecular crystals, focusing the database on inorganic bulk materials.
• Result: A curated set of 72,589 unique stoichiometric inorganic crystal structures (MC3D-source). Of these, ~69,284 are flagged as experimentally known.

B. Geometry Optimization Workflows

• Software Stack: The workflows are implemented in AiiDA (for provenance tracking) and use Quantum ESPRESSO (QE) accelerated by the SIRIUS library (enabling GPU usage).
• Scope: Structures containing lanthanides or actinides were excluded due to the complexity of $f$-electron treatment and pseudopotential limitations. The analysis focused on structures with up to 64 atoms in the primitive cell.
• Protocols: Three versions of the database were generated using different functionals and protocols:
  • PBE-v1 and PBEsol-v1: Earlier versions.
  • PBEsol-v2: The most accurate version, using the PBEsol functional, SSSP v1.3 efficiency pseudopotentials, and aiida-quantumespresso v5 workflows.
• Automation & Error Handling: The workflow includes an automated error handling mechanism (based on BaseRestartWorkChain). It detects convergence failures (ionic/electronic), adjusts numerical parameters (e.g., mixing parameters, diagonalization algorithms), and restarts calculations up to 5 times.
• Magnetic Initialization: All structures are initially treated as magnetic and metallic (high-spin ferromagnetic) to ensure robust convergence, with magnetic moments re-initialized before the final optimization step.

C. Data Release and Interface

• Provenance: Full provenance graphs (inputs, outputs, parameters) are preserved for every calculation.
• Access: Data is available via the Materials Cloud Archive (AiiDA archives) and a web interface with an OPTIMADE-compliant API.

3. Key Results

• Database Size: The final, most precise version (MC3D PBEsol-v2) contains 32,013 unique optimized structures.
• Success Rate: The automated workflow successfully completed geometry relaxations for 85.5% of the processed structures (33,142 out of 38,739).
  • 67.4% of workflows succeeded on the first attempt.
  • The error handling mechanism recovered ~50% of the cases that initially failed.
  • Common failures included convergence issues (ionic/electronic) and walltime limits.
• Volume Validation: Comparing initial experimental volumes with DFT-optimized volumes showed that 78.1% of structures had a volume change within ±5%. Outliers were largely attributed to high-pressure/temperature experimental conditions or layered structures lacking van der Waals corrections in the DFT.
• Novelty: Compared to the Materials Project and OQMD, MC3D contributes 3,328 novel unique structures (Subtype C in Table 1: same composition/spacegroup but non-duplicate geometry, plus new compositions and spacegroups).
• Reproducibility: The database includes the full provenance graph, allowing users to reconstruct the exact computational steps taken to generate any specific geometry.

4. Significance and Contributions

1. Consistency for Machine Learning: By applying a single, rigorous, and consistent protocol (PBEsol-v2) to a large set of experimentally known structures, MC3D provides a high-quality, homogeneous dataset ideal for training ML models, addressing the "inconsistency" problem in existing databases.
2. Expanded Coverage: MC3D integrates data from COD, ICSD, and MPDS, capturing thousands of structures not present in other major computational databases, particularly those from the open COD.
3. Experimental Focus: The strict filtering ensures the database is dominated by experimentally known materials, making it a superior starting point for "materials by design" studies where experimental validation is the goal.
4. Open Science and Reproducibility:
   • The entire pipeline is open-source (AiiDA, QE, SIRIUS).
   • Full provenance is preserved, enabling "more-than-FAIR" data practices.
   • Users can access raw inputs/outputs and reconstruct calculations via the web interface.
5. Robust Workflow Engineering: The paper demonstrates a highly scalable automated workflow capable of handling tens of thousands of calculations with an 85.5% success rate through sophisticated, automated error recovery.

Conclusion

The MC3D database represents a significant step forward in computational materials science by providing a curated, reproducible, and experimentally grounded collection of inorganic crystal structures. It bridges the gap between experimental crystallography and high-throughput DFT calculations, offering a reliable foundation for future materials discovery and machine learning applications.