🔬 physics

Navigating Order-(Dis) Order Family Trees via Group-Subgroup Transitions

This paper introduces a symmetry-based framework called order-(dis)order family trees that utilizes group-subgroup relations to identify when predicted "novel" ordered crystal structures are actually derived from known disordered phases, thereby providing a critical tool for ensuring genuine novelty in high-throughput materials discovery.

Original authors: Shuya Yamazaki, Yuyao Huang, Martin Hoffmann Petersen, Wei Nong, Kedar Hippalgaonkar

Published 2026-07-14

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Shuya Yamazaki, Yuyao Huang, Martin Hoffmann Petersen, Wei Nong, Kedar Hippalgaonkar

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to find a brand-new treasure map. In the world of materials science, scientists use powerful computers to predict new crystal structures—think of them as unique, microscopic Lego castles made of atoms. For years, the rule for "discovery" was simple: if your computer-designed castle doesn't look exactly like any castle already in the big library of known maps (called a database), it's a new find!

But here's the twist: the authors of this paper, Shuya Yamazaki and their team at Nanyang Technological University, found that this rule is missing a huge piece of the puzzle. They argue that many "new" castles aren't actually new at all. Instead, they are just specific, ordered versions of a messy, disordered castle that scientists have already seen.

The "Messy Room" vs. The "Tidy Room" Analogy

Think of a disordered crystal like a messy teenager's bedroom. You know the room exists, and you know the general layout (a bed, a desk, a closet), but the clothes are thrown everywhere, and you can't tell exactly which sock is on which foot. In science, we call this occupational disorder. It's a "statistical average" where different types of atoms (like red socks and blue socks) share the same spot in a jumbled mix.

Now, imagine a computer predicts a "new" crystal. It draws a picture of that same bedroom, but this time, every single sock is perfectly folded and placed in a specific spot. The computer says, "Look! A brand-new, perfectly organized room!"

The authors say: Wait a minute. That's not a new room. It's just the messy room you already knew about, but now you've imagined a specific way to tidy it up. In their language, the messy room is the "disordered parent," and the perfectly organized room is the "ordered child."

If you only look for new rooms by checking if the "tidy" version exists in the library, you'll keep thinking you've found a new house every time you just tidy up an old one. You're rediscovering the same family, just in a different state of order.

The Family Tree Detective Work

To fix this, the team built a new tool called Order-(Dis)Order Family Trees. Instead of looking at crystals as isolated islands, they map them like a family tree.

The Root: The messy, disordered parent phase (the average structure).
The Branches: The various ways that mess can be organized into tidy, ordered structures.
The Siblings: Different ordered versions that share the same messy parent.

They used a special "translation" tool called SWORD to read the crystal structures and see if a new, tidy prediction is actually just a child of a known messy parent.

What They Found (The Evidence)

The team didn't just guess; they tested this idea on real data. Here is what they discovered:

1. The "A-Lab" Robot Failures
They looked at a project called A-Lab, where robots try to synthesize materials predicted by AI. In a study of 35 "successful" new materials, they found that 60% of them (21 out of 35) were actually just ordered versions of known messy parents.

Example: A robot predicted a new, tidy version of a material called MgTi2NiO6. But when the scientists checked the family tree, they realized it was just a specific ordering of a known messy ilmenite-type phase. The robot thought it found a new species; the family tree showed it just tidied up an old one.

2. The "New" Databases
They checked huge databases of known crystals (ICSD) and computer-generated lists (like MP-20 and GNoME).

In the MP-20 database (a collection of 36,183 structures), 23.27% of the "ordered" crystals were actually children of known disordered parents.
In the GNoME database (a massive list of 50,000 stable structures), only 0.73% were found to have known disordered parents in the current database. However, the authors suggest this low number isn't because these crystals are truly unique. It's likely because the messy parents for these new areas haven't been discovered or recorded in the library yet. The "messy room" might exist in reality, but it's just not in the book yet.

3. The AI Model Mistake
The team tested 10 different AI models that generate new crystal structures. They found a big difference between two types of AI:

The "Symmetry-Agnostic" Models: These models (like MiAD and ADiT) ignore the rules of symmetry and just throw atoms around. They produced a lot of "new" structures, but 14.24% of MiAD's output and 12.61% of ADiT's output were just ordered children of known messy parents.
The "Symmetry-Constrained" Models: These models (like WyFormer) are forced to follow the rules of crystal symmetry. They were much better! WyFormer only produced 3.40% of structures that were just rediscoveries of known messy families.

4. The "P1" Problem
One specific type of crystal symmetry, called P1, is the simplest kind (it has almost no rules). In nature, true P1 crystals are super rare (less than 1% of all crystals). But the "Symmetry-Agnostic" AI models love to make them. The authors found that for these models, P1 structures dominated the list of "rediscovered" messy children.

They ran extra computer simulations (DFT) to check if these P1 structures were actually stable. They found that about 80–90% of them were "metastable," meaning they were just barely hanging on to existence. This suggests the AI is generating these "super simple" structures because they are easy mathematical ways to tidy up a messy parent, not because they are truly new, stable materials.

What They Are NOT Saying

It is important to know what this paper doesn't claim:

They are not saying that all new materials are fake. They are saying that many that look new are actually just re-ordered versions of old ones.
They are not saying that disordered crystals are "bad." In fact, they argue that the messy parent is often the real thing that nature makes, and the "tidy" version is just a theoretical possibility.
They are not claiming that the low numbers in the GNoME database mean those crystals are definitely unique. They explicitly state that the low number might just mean we haven't found the messy parents for those specific chemical combinations yet.

The Big Takeaway

The main finding is that to find truly new materials, we need to stop looking at crystals as single, isolated points. Instead, we need to look at the whole family tree.

If a computer predicts a new, tidy crystal, we shouldn't just ask, "Has this exact tidy version been seen before?" We need to ask, "Is this just a tidy version of a messy room we already know?"

The authors suggest that for the next generation of materials discovery, we need to use these "family trees" to check for novelty. If we don't, we might waste time and money trying to synthesize "new" materials that are actually just old friends wearing a different outfit. By using symmetry-based tools (like the WyFormer model), scientists can avoid these traps and find the truly unique, undiscovered families of crystals.

Technical Summary: Navigating Order-(Dis)Order Family Trees via Group-Subgroup Transitions

Problem Statement
As closed-loop materials discovery systems scale to generate millions of candidate compounds, the definition of "novelty" has become a critical bottleneck. Current novelty assessment relies on comparing predicted ordered crystal structures against databases of ordered entries. However, this approach introduces a systematic blind spot: many predicted ordered structures are not genuinely new compounds but rather specific atomic orderings of known disordered phases (where sites are statistically occupied by multiple species). When these "ordered children" of "disordered parents" are compared only against ordered references, they are erroneously flagged as novel. This leads to the overestimation of discovery rates and the misallocation of experimental resources, as seen in recent A-Lab studies where synthesized products were found to be known disordered parents rather than the predicted ordered phases.

Methodology
The authors introduce a symmetry-based framework called order-(dis)order family trees to organize crystal structures through group-subgroup relations. The methodology relies on three core components:

SWORD Representation: The framework utilizes the Symmetry and Wyckoff-sequence of Ordered and Disordered crystals (SWORD) representation. This unified encoding describes both ordered and disordered structures using space group numbers, Wyckoff positions, and atomic species/occupancies, enabling consistent comparison across different states of order.
Group-Subgroup Transitions: The framework models the relationship between a disordered parent (high symmetry) and its ordered descendants (lower symmetry) as a transition where Wyckoff positions split. This is governed by crystallographic group-subgroup relations, specifically translationengleiche (t-type) transitions that preserve the translation group while lowering point symmetry.
High-Throughput Matching (SWORDFamilyMatcher): A procedure is developed to systematically generate candidate parent descriptions for a query ordered structure and match them against experimental databases (ICSD). This allows the identification of:
- Disordered Parents: Existing experimental phases that serve as the root of the family tree.
- Ordered Siblings: Other ordered structures derived from the same parent.
- Family Metrics: Two new metrics are defined to quantify novelty at the family level:
  - $FT_{disorder}$ : The fraction of ordered structures connected to existing disordered parents.
  - $FT_{order}$ : The fraction of ordered structures connected to existing ordered structures via group-subgroup relations (sharing a common parent).

Key Contributions and Results

Validation on A-Lab Cases: The framework was tested on 35 GNoME compounds synthesized by A-Lab. It successfully recovered experimentally reported disordered parents for 22 of these cases. For three specific targets ( $K_2TiCr(PO_4)_3$ , $KPr_9(Si_3O_{13})_2$ , and $Mn_2VPO_7$ ), the framework identified disordered parents that previous manual analyses had missed, demonstrating its ability to navigate complex order-disorder lineages systematically.
Benchmarking Databases:
- ICSD: 6.13% of ordered structures in the Inorganic Crystal Structure Database (ICSD) were found to belong to families rooted in existing disordered parents, suggesting a non-trivial regime of order-disorder transitions in experimental records.
- Computational Datasets: MP-20 showed a high $FT_{disorder}$ of 23.27%, while Alex-MP-20 and GNoME showed lower values (3.75% and 0.73%, respectively). The authors note that lower values in computational datasets likely reflect incomplete coverage of disordered parents in experimental references rather than a true absence of disorder propensity.
Generative Model Analysis: The framework was applied to 10 state-of-the-art crystal generative models.
- Symmetry-Agnostic Models: All-atom models (e.g., MiAD, ADiT, MatterGen) exhibited significantly higher $FT_{disorder}$ values (up to 14.24%), indicating they frequently generate ordered structures that are merely children of known disordered parents.
- Symmetry-Constrained Models: Models operating in the Wyckoff subspace (e.g., WyFormer, SymmCD) showed $FT_{disorder}$ values 2–4 times lower. These models are less prone to generating "rediscoveries" of known disorder-rooted families.
The P1 Phenomenon: The study analyzed the overproduction of space group P1 structures by all-atom models. It was found that a dominant portion of these P1 structures are ordered children of high-symmetry disordered parents (e.g., $Fm\bar{3}m$ ). DFT validation suggests many of these are metastable, implying that the "stability" of these P1 structures is an artifact of them being ordered configurations of experimentally realized disordered phases that are unlikely to form in that specific ordered state.
Group-Subgroup Polymorphs: The framework identifies a subset of polymorphs connected by explicit symmetry-descent pathways, revealing that roughly 40% of polymorphic systems in ICSD contain pairs related by group-subgroup transitions.

Significance and Claims
The paper argues that the unit of analysis for materials discovery must shift from isolated structures to crystallographic families. The authors claim that order-(dis)order family trees are a key requirement for achieving genuine novelty in data-driven discovery. By explicitly mapping the lineage between disordered parents and ordered children, the framework:

Prevents the mislabeling of known disorder-rooted families as new discoveries.
Provides a more realistic assessment of synthesis outcomes, distinguishing between a new ordered phase and a rediscovery of a known disordered parent.
Suggests that symmetry-constrained generative models are better positioned for genuine family-level discovery than symmetry-agnostic all-atom models.

The work concludes that novelty should not be assessed solely by structural distinctness at the child level but must consider crystallographic lineage. The disorder-aware family matching framework serves as a practical foundation for future discovery pipelines to identify materials that are not only structurally unique but also represent genuinely new families relative to experimental reality.

The "Messy Room" vs. The "Tidy Room" Analogy

The Family Tree Detective Work

What They Found (The Evidence)

What They Are NOT Saying

The Big Takeaway

More like this