An Investigation in the Kinetic Persistence of TiO$_2$ Polymorphs using Machine Learning Driven Pathfinding in Crystal Configuration Space

This study introduces a machine learning-driven pathfinding algorithm based on the Crystal Normal Form to map diffusionless transformation pathways and analyze the kinetic persistence of metastable TiO$_2$ polymorphs by correlating potential energy landscape topography with experimental phase stability.

The Gist

The Big Problem: The "Synthesis Bottleneck"

Imagine you are an architect who can design thousands of amazing, futuristic houses on a computer. You have a super-fast AI that can dream up new structures that are theoretically perfect. But there's a catch: just because you can draw a house doesn't mean you can build it.

In the world of materials science, scientists have used computers to predict thousands of new crystal structures (tiny, repeating patterns of atoms). However, we are stuck in a "bottleneck." We have a massive list of theoretical materials, but we don't know which ones can actually be built in a real lab, and which ones will instantly collapse or turn into something else the moment we try to make them.

The big question is: Why do some materials stay stable, while others disappear?

The Old Way vs. The New Way

The Old Way (Thermodynamics):
Traditionally, scientists looked at the "energy" of a material. Think of this like a ball on a hill. If the ball is at the very bottom of the valley, it's stable. If it's on a hill, it wants to roll down. Scientists assumed that if a material was "low energy" (at the bottom of the hill), it would be easy to make.

The Problem:
Sometimes, a ball is sitting in a small, shallow dip on the side of a mountain. It's not at the very bottom (the most stable state), but it's stuck there. It won't roll down because there's a big wall or a deep valley blocking its path. This is called kinetic persistence. The material is "metastable"—it's not the best, but it's stuck in a good spot.

The old methods couldn't see these "walls." They only saw the bottom of the valley.

The New Tool: The Crystal Map

This paper introduces a new way to look at the landscape. The authors (Gallant, Mrdjenovich, and Persson) developed a method to map out the "terrain" between different crystal shapes.

Think of the Crystal Normal Form (CNF) as a unique fingerprint for every crystal structure.

• Imagine every crystal is a specific arrangement of Lego bricks.
• The CNF turns that Lego arrangement into a unique string of numbers.
• If two structures are slightly different, their number strings are slightly different.

By connecting these number strings, the scientists created a giant graph (a map of dots and lines). Each dot is a crystal structure, and the lines are tiny steps you can take to turn one structure into another.

The "Ceiling" Game: Finding the Path

The hardest part of this map is finding the path from Point A (a theoretical crystal) to Point B (a known, stable crystal) without climbing over a mountain that is too high.

The authors invented a clever game called the "Ceiling Lowering Algorithm":

1. The Ceiling: Imagine a giant glass ceiling hovering over the map. Any path that goes above this ceiling is blocked.
2. The Search: The computer tries to find a path from A to B that stays under the ceiling.
3. The Drop: If it finds a path, it lowers the ceiling a tiny bit (like lowering a floodgate).
4. Repeat: It tries again. It keeps lowering the ceiling until it can't find a path anymore.

The point where the path finally breaks is the energy barrier.

• Low Barrier: The ceiling didn't have to go very low. The path is easy. The material will likely change shape quickly (it's not stable).
• High Barrier: The ceiling had to be lowered almost to the ground. The path is hard to cross. The material is "kinetically trapped" and will likely stay as it is.

The Case Study: Titanium Dioxide (TiO₂)

To test this, they looked at TiO₂ (Titanium Dioxide), a material used in white paint, sunscreen, and solar cells.

• The Knowns: We know three main forms: Rutile, Anatase, and Brookite.
• The Unknowns: There are many other theoretical forms that scientists have never seen in nature.

What they found:

• The "Ghost" Materials: They found that several theoretical forms of TiO₂ have very low barriers to turning into the known forms. It's like they are standing on a slippery slope with no wall. This explains why we never find them in nature—they instantly transform into the stable versions.
• The "Stuck" Materials: They confirmed that the known forms (like Anatase) have high walls protecting them. Even though Rutile is technically more stable, Anatase has a high energy wall preventing it from turning into Rutile too quickly. This is why we can still buy Anatase-based sunscreen.
• The "Secret" Paths: They discovered specific routes (like a secret tunnel) where one crystal turns into another. For example, they found a very easy path between two high-pressure forms, explaining why they are often seen together in experiments.

The "AI Teacher" Trick

Doing this math is incredibly hard. It's like trying to solve a maze with a billion turns. Calculating the energy of every single step using standard physics (DFT) would take a supercomputer years to finish.

To speed this up, they used Machine Learning:

1. They ran a few calculations using the slow, accurate method (DFT).
2. They taught a fast AI (a Machine Learning Potential) to mimic the slow method.
3. They used the fast AI to explore the maze.
4. They checked the AI's work occasionally to make sure it wasn't lying.

This is like hiring a student to do the homework, but you only check their answers every few pages to make sure they are still on the right track.

The Bottom Line

This paper gives us a new pair of glasses. Instead of just asking "Is this material the lowest energy?", we can now ask, "Is there a wall blocking this material from changing?"

By mapping these invisible walls, scientists can finally predict which of the thousands of theoretical materials are actually worth trying to build in the lab, and which ones are just mathematical ghosts that will vanish the moment we try to create them. It's a massive step toward solving the "synthesis bottleneck" and building the materials of the future.

Technical Summary

1. Problem Statement

The field of computational materials science has successfully generated vast libraries of hypothetical crystal structures using high-throughput simulations and generative deep learning. However, a "synthesis bottleneck" exists: many theoretically stable structures have never been experimentally realized.

• The Gap: Traditional thermodynamic metrics (e.g., formation energy, convex hull distance) and dynamic stability checks (phonon calculations) are insufficient to predict kinetic persistence. A structure may be thermodynamically metastable and dynamically stable (no imaginary phonon modes) yet remain unobserved because the energy barrier to transform into a more stable phase is too low to prevent decomposition, or conversely, the barrier to form it is too high.
• The Core Question: Why do certain metastable polymorphs persist experimentally while others do not? The authors hypothesize that the topography of the Potential Energy Landscape (PES)—specifically the height of energy barriers separating metastable phases from ground states via diffusionless transformations—determines kinetic viability.
• The Challenge: Identifying these transformation pathways is computationally expensive. Existing methods like Nudged Elastic Band (NEB) require an initial guess of the pathway, while exhaustive search methods (like the original "water-filling" algorithm on the Crystal Normal Form graph) suffer from combinatorial explosion as system size increases, making them infeasible for unit cells with more than ~6 atoms.

2. Methodology

The authors developed a novel, iterative pathfinding framework that combines the Crystal Normal Form (CNF) graph representation with Machine Learning Interatomic Potentials (MLIPs) and an adaptive A search algorithm*.

A. Crystal Normal Form (CNF) Graph

• Representation: The CNF provides a unique, unambiguous fingerprint for crystal structures by representing the lattice via seven Voronoi vectors and the motif via relative atomic coordinates. This resolves unit cell ambiguity.
• Graph Structure: Neighbors in the CNF graph represent structures related by small lattice strains or atomic displacements (phonon amplitudes). Paths through this graph correspond to diffusionless transformation pathways.

B. The Iterative Ceiling-Lowering Algorithm

To overcome the computational cost of exhaustive searches for larger systems, the authors introduced a hybrid approach:

1. A Pathfinding:* Uses a Manhattan distance heuristic on the CNF coordinate list to find the shortest path between two polymorphs.
2. Energy Ceiling Filter: The search is constrained by an energy ceiling ($C$). Only nodes with energies $E \le C$ are explored.
3. Iterative Refinement:
   • Start with a high energy ceiling to find a feasible path quickly.
   • Once a path is found with a maximum energy $E_{max}$, lower the ceiling to $E_{max} - \delta C$.
   • Repeat the search on the restricted subgraph.
   • This process iteratively lowers the energy barrier until no path exists or the algorithm terminates, providing an upper bound on the true transition barrier.

C. Machine Learning Potential (MLIP) Workflow

To enable the millions of energy evaluations required for this search, Density Functional Theory (DFT) was replaced by a fast MLIP.

• Distillation Strategy:
  1. Initial pathways were generated using a pre-trained GrACE (linear graph atomic cluster expansion) model.
  2. Frames from the central regions of these paths were calculated using DFT (GGA) to create a high-fidelity dataset.
  3. A large MACE (Message-Passing Atomic Cluster Expansion) model was fine-tuned on this DFT data.
  4. The fine-tuned MACE model was used to generate a massive training set (thousands of frames) to distill a new, highly accurate GrACE model.
• Performance: The distilled GrACE model achieved an accuracy of 19.1 meV/atom against DFT, a significant improvement over the pre-trained version (184 meV/atom), while remaining orders of magnitude faster than DFT.

3. Key Contributions

1. Algorithmic Innovation: Introduced an iterative "ceiling-lowering" A* algorithm that enables low-energy pathway discovery for unit cells with up to 24 atoms, a scale previously inaccessible to exhaustive CNF searches.
2. MLIP Distillation Pipeline: Demonstrated a robust workflow for distilling high-accuracy, linear MLIPs from complex message-passing models, specifically optimized for the high-throughput demands of pathfinding.
3. Systematic Application to TiO2: Applied the framework to the TiO2 system, analyzing 10 polymorphs (3 experimental, 3 high-pressure experimental, 4 theoretical) to map the kinetic landscape.

4. Results

The study mapped energy barriers (in meV/atom) between TiO2 polymorphs, correlating barrier heights with experimental observability.

• Explained Absence of Phases:
  • $\beta$-TiO2, Baddeleyite, and Pnma-I: Found to have low-energy barriers (39–53 meV/atom) connecting them to Anatase. This explains why these phases are not observed at ambient conditions; they kinetically decompose rapidly into Anatase.
  • Pnma-II (Theoretical): Connected to other phases only via extremely high barriers (>1500 meV/atom) or no path found. This suggests it is kinetically trapped or the search space was too complex to find a low-energy route, explaining its absence.
• Validated Known Transformations:
  • Columbite $\leftrightarrow$ Baddeleyite: Found an extremely facile pathway (8 meV/atom), consistent with literature.
  • Baddeleyite $\leftrightarrow$ Rutile: Found a low barrier (20 meV/atom), confirming previous findings.
  • Brookite $\leftrightarrow$ Rutile: Confirmed the two-step mechanism (Brookite $\to$ Columbite $\to$ Rutile) with barriers of 106 and 96 meV/atom respectively, while the direct path was high energy (460 meV/atom).
• Pathway Character:
  • Low-energy pathways tend to involve a balanced mix of lattice strain and atomic shuffling (diagonal trajectory in step-space).
  • High-energy pathways often exhibit a "two-stage" character (lattice distortion followed by motif distortion), which the authors attribute to heuristic biases in the A* algorithm for larger unit cells (24 atoms).
• Limitations Identified: The algorithm struggled to find low-energy paths for Brookite $\leftrightarrow$ Anatase (finding a 1316 meV/atom barrier vs. literature values of 73–320 meV/atom). The authors attribute this to the combinatorial complexity of 24-atom unit cells and the heuristic bias of the A* algorithm favoring lattice alignment over motif alignment in large systems.

5. Significance

• Beyond Thermodynamics: The study demonstrates that kinetic persistence is a critical filter for materials synthesis. A structure's "synthesizability" is not just about how stable it is, but how difficult it is to transform away from it.
• Scalability: By combining CNF graph theory with iterative A* search and distilled MLIPs, the method bridges the gap between small-system exhaustive searches and large-system heuristic searches, opening the door to studying diffusionless transformations in complex oxides and other functional materials.
• Design Guidance: The framework provides a quantitative metric (barrier height) to screen hypothetical materials. If a metastable phase has a low barrier to a ground state, it is unlikely to be synthesized under standard conditions unless specific kinetic traps (e.g., rapid quenching) are employed.
• Future Directions: The authors highlight that improvements in search heuristics (unbiased searches) and adaptive resolution strategies are necessary to fully resolve the pathways for larger, more complex unit cells.

In summary, this paper provides a powerful computational tool to navigate the "synthesis bottleneck" by quantifying the kinetic barriers that dictate whether a theoretically predicted material can actually exist in the real world.