GEWUM: General Exploration Workflow for the Utopia of Materials: A Unified Platform for Automated Structure Generation, Selection, and Validation

GEWUM is a unified, open-source platform that automates materials discovery by integrating Selective Random Structure Search with universal Machine Learning Interatomic Potentials to enable efficient, scalable structure generation, selection, and validation on HPC clusters.

The Gist

Imagine you are a master chef trying to invent the perfect new dish. You have a pantry full of 100 different ingredients (atoms), and you want to find the specific combination that tastes amazing (is stable) and has the perfect texture (has useful properties like conducting heat or electricity).

In the past, trying to find this "perfect dish" was like searching for a needle in a haystack, but the haystack was the size of a galaxy. Scientists had to write different recipes for every step: one script to mix the ingredients, another to cook them, a third to taste-test them, and a fourth to check if the dish would explode in the oven. It was messy, slow, and required a lot of manual labor.

Enter GEWUM.

Think of GEWUM (General Exploration Workflow for the Utopia of Materials) as a fully automated, super-smart robotic kitchen that does all of this for you. It's a new software platform designed to help scientists discover new materials faster and easier than ever before.

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

1. The "Magic Taste-Testers" (uMLIPs)

Traditionally, checking if a new material works required a supercomputer to run a very slow, heavy calculation (like a slow-motion cooking simulation).
GEWUM uses Universal Machine Learning Interatomic Potentials (uMLIPs). Think of these as AI sous-chefs that have tasted millions of dishes. They haven't actually cooked every single dish, but they have learned the rules of flavor so well that they can predict how a new combination will taste instantly and with near-perfect accuracy. This makes the process thousands of times faster.

2. The "Smart Randomizer" (SRSS)

Instead of guessing randomly and hoping for the best, GEWUM uses a strategy called Selective Random Structure Search (SRSS).
Imagine you are looking for a new outfit. Instead of trying on every single shirt in the world, you ask a stylist to pull out 1,000 random shirts, but the stylist is smart enough to throw away the ones that are clearly ugly or don't fit, keeping only the most unique and promising ones. GEWUM does this with atoms: it generates millions of random crystal structures, filters out the bad ones, and keeps the most interesting ones to study.

3. The "Conveyor Belt" (Unified Workflow)

The biggest problem scientists faced was that their tools were scattered. They had to move files from one computer program to another, manually fixing errors along the way.
GEWUM is a unified conveyor belt. Once you tell it what ingredients you want (e.g., Aluminum, Scandium, and Nitrogen), it automatically:

• Generates the random structures.
• Filters them to find the best candidates.
• Relaxes them (lets them settle into their most comfortable shape).
• Tests them for stability (will they fall apart?).
• Calculates their properties (how hard are they? how well do they conduct heat?).

All of this happens automatically, without the scientist needing to write a single line of code in between steps.

4. The "Super-Server" (HPC & SLURM)

Doing this for thousands of materials requires massive computing power. GEWUM is built to talk directly to HPC (High-Performance Computing) clusters, which are like massive server farms with thousands of processors.
Think of GEWUM as the conductor of a massive orchestra. It tells thousands of computers exactly what to do at the same time. If one computer crashes, the system notices and just moves that task to another computer, so the work never stops. You don't need to be a computer expert to use it; you just press "start."

What Did They Discover? (The Menu)

To prove their robotic kitchen works, the team used GEWUM to cook up three new "dishes":

1. Al-Sc-N (A new piezoelectric material): They found new, stable shapes for a material used in sensors and speakers that were previously unknown.
2. U3Si5 (A nuclear fuel discovery): Everyone thought Uranium Silicide only had one shape. GEWUM found a second shape (a new crystal structure) that is stable and different from what we thought we knew.
3. ThH10 (High-pressure hydrogen): They looked for a material that might become a superconductor (conducts electricity with zero resistance) under extreme pressure, finding the most stable version at 150 GPa (a pressure higher than the center of the Earth).

Why Does This Matter?

GEWUM is like giving scientists a time machine. Instead of spending years manually testing materials, they can now screen thousands of possibilities in days or weeks. This accelerates the discovery of materials needed for:

• Clean Energy: Better batteries and solar panels.
• Sustainability: Materials that last longer and use fewer resources.
• Technology: Faster electronics and stronger construction materials.

In short, GEWUM takes the messy, fragmented, and slow process of material discovery and turns it into a streamlined, automated, and powerful engine for innovation. It's not just a tool; it's a new way to explore the "Utopia of Materials."

Technical Summary

1. Problem Statement

The discovery of materials with tailored properties increasingly relies on computational methods, particularly the integration of Universal Machine Learning Interatomic Potentials (uMLIPs) and Crystal Structure Prediction (CSP). While uMLIPs offer near-Density Functional Theory (DFT) accuracy at a fraction of the computational cost, the current software ecosystem suffers from critical fragmentation:

• Workflow Discontinuity: Researchers must manually stitch together disparate scripts for structure sampling, diversity selection, and rigorous stability validation (e.g., convex hull analysis, phonon calculations).
• Lack of Scalability: Most existing frameworks lack native integration with High-Performance Computing (HPC) workload managers (like SLURM), making it difficult to scale workflows from testing a few candidates to screening thousands of structures across complex chemical spaces.
• Accessibility Barrier: The complexity of managing job submission scripts, resource allocation, and dependency management hinders researchers without specialized HPC expertise.

2. Methodology: The GEWUM Platform

To address these gaps, the authors developed GEWUM (General Exploration Workflow for the Utopia of Materials), a unified, open-source Python platform designed to automate and accelerate materials discovery.

Core Architecture & Design

• Modular Design: The system is built on a four-tier hierarchy:
  1. Command Line Interface (CLI): A unified syntax (gewum <command> --mode <mode>) for user interaction.
  2. Workflow Command Layer: Object-oriented dispatching (e.g., RDCommand, PTCommand) handling logic.
  3. Script Repository: Organized directories for specific workflows and shared functional modules.
  4. Computational Backend: Integration with external libraries (ASE, PyXtal, Phonopy, Phono3py, Materials Project API) and uMLIPs.
• SLURM Integration: A template-driven configuration system (slurm_config.yaml) allows users to deploy workflows on HPC clusters by modifying a single file. It supports two-level parallelism:
  • Outer Level: Job submission via SLURM templates.
  • Inner Level: Multi-core parallelization within nodes using Python's multiprocessing or GNU Parallel.
• Fault Tolerance: Implements idempotent task filtering to skip already processed structures, enabling safe resubmission after node failures without redundant computation.

Key Functional Modules

1. Random Design (RD): Implements the Selective Random Structure Search (SRSS) strategy.
   • Generates structures across 230 space groups, 80 layer groups, and 32 point groups.
   • Uses K-means or HDBSCAN clustering for diversity-preserving selection.
   • Performs high-throughput relaxation using uMLIPs, followed by thermodynamic (convex hull) and dynamic (phonon) stability screening.
2. Perturbation (PT): Enables systematic exploration from known configurations via supercell generation, element substitution, structural mutation, and doping engineering.
3. Property Calculations:
   • ELA: Elastic constants and moduli via energy-strain methods.
   • TC: Lattice thermal conductivity using Boltzmann transport equations (Phonopy/Phono3py).
   • QHA: Quasi-harmonic approximation for thermal expansion and heat capacity.
   • MD: NVT ensemble simulations for kinetic stability.
4. Fine-Tuning (FT): Uncertainty quantification to identify high-uncertainty structures for targeted DFT validation.

3. Key Contributions

• Unified End-to-End Workflow: GEWUM is the first platform to seamlessly integrate unbiased structure generation (SRSS), uMLIP-based relaxation, diversity selection, and rigorous stability validation into a single, automated pipeline.
• Native HPC Scalability: By abstracting SLURM complexities, GEWUM enables massive parallelization (thousands of cores) without requiring users to write complex job scripts.
• Open-Source Accessibility: The platform is distributed under the MIT License, supporting backend-agnostic uMLIPs (MACE, DPA3, SevenNet, etc.) via the ASE interface, lowering the barrier to entry for the scientific community.
• Comprehensive Property Suite: It extends beyond structure prediction to include automated calculation of thermophysical properties (thermal conductivity, expansion, elasticity) using uMLIPs.

4. Results & Case Studies

The authors validated GEWUM through four distinct case studies:

• Case 1: Al-Sc-N System (Complex Nitrides)

  • Task: Predict low-energy polymorphs in the ternary Al-Sc-N system.
  • Process: Generated ~13,300 candidates (from 45,800 attempts), filtered via uMLIP, and refined with DFT.
  • Outcome: Identified the ground-state Cm-1 structure and several low-energy metastable phases (P1-3, P1-12, P1-11), revealing a rich configurational landscape.

• Case 2: U3Si5 (Uranium Silicide)

  • Task: Search for unknown allotropes of U3Si5.
  • Outcome: Predicted a novel P-62c phase distinct from the known AlB2-type.
  • Validation: DFT+U refinement showed lattice parameters closer to experimental values than the AlB2 type. XRD simulations and AIMD confirmed the stability of a unique distorted 12-membered silicon ring structure.

• Case 3: ThH10 at 150 GPa (High-Pressure Hydrides)

  • Task: High-pressure structure prediction.
  • Process: SRSS generated 29,400 valid structures; clustering reduced this to 5,880 for optimization.
  • Outcome: Successfully identified the Fm-3m phase as dynamically stable, while ruling out other candidates (F-43m, P42212, etc.) due to imaginary phonon modes. The entire screening process was highly efficient (~6 hours for optimization on a single node).

• Case 4: Thermophysical Property Validation

  • Task: Benchmark TC and QHA modules.
  • Outcome: Calculated lattice thermal conductivity for cubic SiC and U3Si2 matched experimental trends. The QHA module accurately predicted thermal expansion coefficients. The computational cost was reduced to 3–30 minutes per structure, significantly faster than traditional DFT methods.

5. Significance

GEWUM represents a paradigm shift in computational materials science by bridging the gap between advanced machine learning potentials and practical, scalable engineering workflows.

• Efficiency: It drastically reduces the time and expertise required to screen vast chemical spaces, transforming CSP from a manual, script-heavy process into an automated "push-button" workflow.
• Scalability: Its native SLURM integration allows researchers to leverage the full power of modern HPC clusters, making large-scale discovery feasible.
• Impact: By providing a robust framework for predicting both structure and properties, GEWUM accelerates the discovery of functional materials for critical applications in energy, electronics, and sustainability.

The software is available at https://github.com/JesseOOPP/GEWUM.