Challenges in Non-Polymeric Crystal Structure Prediction: Why a Geometric, Permutation-Invariant Loss is Needed

arXiv:2509.00832 · cs.LG, cond-mat.mtrl-sci · Submitted 2025-08-31 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "Challenges in Non-Polymeric Crystal Structure Prediction".

Jane: Here's the gist of the paper. Crystalline structure prediction is an essential prerequisite for designing materials with targeted properties.

Tom: In this work, we focus on the molecular assembly problem, where a set S of identical rigid molecules is packed to form a crystalline structure.

Jane: We propose a better formulation that introduces a loss function capturing key geometric molecular properties while ensuring permutation invariance over S.

Tom: Remarkably, we demonstrate that within this framework, a simple regression model already outperforms prior approaches, including flow matching techniques, on the COD-Cluster17 benchmark.

Paper summary: Tom: So, this paper, "Challenges in Non-Polymeric Crystal Structure Prediction: Why a Geometric, Permutation-Invariant Loss is Needed," it’s looking at how we can predict those three dimensional crystal structures for molecules that aren't polymers. It sets up the molecular assembly problem where you take a set S of identical rigid molecules and pack them to form something that can then be repeated in space into a crystal.

Jane: That sounds like a really specific challenge because predicting those patterns is essential for designing materials with certain properties, which we need to know before we even start making things. What’s the core problem they’re pointing out about current methods?

Tom: They argue that even though we have sophisticated techniques now, the learning objective itself is still ill-posed, which means it's not well defined mathematically for this specific problem <ref:2509.00832#pg1>. They suggest a better formulation with a loss function that captures these key geometric molecular properties while making sure it stays invariant to the order of the molecules in the set S.

Lu: The authors are proposing two rigid body loss functions, one called LRMSD which looks at absolute positions, and another one called LGeom which extends that idea to relative molecular positions <ref:2509.00832#pg3>. Their goal is to have a differentiable soft matching objective that ignores global geometric transformations and the order of the repeated molecular units.

Meng: If it’s invariant to order, that sounds like a big deal for real-world applications because molecules don't come out in a specific sequence when you pack them together <ref:2509.00832#pg3>. But what does this actually mean for the computer trying to learn the crystal structure?

Jane: It means they are minimizing a loss function called L star ML, which involves some parameter P where P has to be positive and sum over i and j, ensuring that P times one equals one <ref:2509.00832#pg1>. Essentially, they’re trying to find a way to guide the prediction process efficiently.

Tom: And the results are pretty compelling because they show that a simple regression model already beats prior approaches, including flow matching techniques, on the COD-Cluster17 benchmark <ref:2509.00832#pg1>. That’s a clean win for simplicity over complexity.

Paper summary: Lu: They also showed that training with LGeom is the solution when looking at the geometric relative packing loss metric called L star Geom, which is in equation number ten <ref:2509.00832#pg1>. And they found that using losses that are permutation invariant actually gave a significant gain, sometimes dividing the evaluation metric by up to three point eight compared to just using absolute positions <ref:2509.00832#pg1>.

Meng: Three point eight is a solid improvement, but I have to ask about the data. They’re using the COD-Cluster17 assembly dataset, which they describe as a simplified and sanitized version of the 507k crystals from the real world Crystallography Open Database <ref:2509.00832#pg1>. How reliable is that simplification for predicting actual complex materials?

Jane: That’s a fair question, Meng, because they did admit a limitation right in the paper. They state that one fundamental limitation of COD-Cluster17 is the absence of periodic boundary conditions, and they say that this hinders the generalisation capability of any model <ref:2509.00832#pg1>.

Tom: So, while it’s a huge step forward on the formulation itself, they are telling us that to make these models truly useful for real crystal structures, we still need to train them end-to-end so they can learn the conformation and the crystal structure prediction jointly <ref:2509.00832#pg1>.

Lu: That makes sense. The paper points out that a model trained this way can take a number of approximations and hypotheses to solve the original problem, which is computing the infinite three dimensional structure given its substituent chemical compounds <ref:2509.00832#pg3>. They’re proposing an efficient and model agnostic way to guide this process.

Meng: Efficiency is key when you're trying to ship models to production, Tom, because if the loss function requires too much complex optimization or a massive number of molecules in that sum over i and j, it just won't run on our hardware <ref:2509.00832#pg1>. What about the runtime cost of this new loss compared to what we’re using now?

Tom: The paper suggests that a simple regression model already outperforms prior methods on both the COD-Cluster17-5K subset and the full dataset by a significant margin <ref:2509.00832#pg1>. It implies that this new framework isn't just theoretically better, it’s practically faster or at least requires less training effort to get good results <ref:2509.00832#pg1>.

Paper summary: Jane: So, what does this mean for someone who is just trying to design a new drug or material? Does this loss function actually translate into something tangible for them beyond just better numbers on a benchmark like COD-Cluster17?

Lu: It translates into a more robust way to learn the relationship between the molecules and their final crystal pattern <ref:2509.00832#pg3>. The paper is really about showing that you need this specific geometric and permutation invariant loss to get past the current hurdles in predicting those three dee structures <ref:2509.00832#pg1>.

Meng: I’m still skeptical about generalizing it beyond the sanitized dataset, though. If a model only learns from that simplified version, its performance on a novel crystal structure might drop off sharply when we introduce real complexity <ref:2509.00832#pg1>. We need to see if this holds up when we run it on actual traffic or real molecular diversity.

Jane: It sounds like the next step is training these models end-to-end, so they learn the structure and the conformation together <ref:2509.00832#pg1>. That would be a big leap for making them useful in practice.

Tom: Exactly. So to wrap up this discussion on "Challenges in Non-Polymeric Crystal Structure Prediction: Why a Geometric, Permutation-Invariant Loss is Needed," the authors are pushing for this specific loss formulation because it addresses the ill-posed nature of the objective function that plagued previous attempts <ref:2509.00832#pg1>. It shows that geometric properties and permutation invariance are necessary components for a solid approach here.

Lu: They’re showing that even within this simplified molecular assembly problem, introducing this specific type of loss function can significantly improve the performance over existing state-of-the-art methods <ref:2509.00832#pg1>.

Meng: It’s a good theoretical foundation, but until we see it run reliably on real data with periodic boundary conditions, it’s just a strong hypothesis for now <ref:2509.00832#pg1>.

Jane: And the implication is that if we can solve this molecular assembly subproblem better, we get closer to predicting those complex infinite three dimensional structures for real materials <ref:2509.00832#pg3>.

Tom: That’s the direction they are pointing us toward, focusing on making the loss function itself more representative of what a crystal structure actually is <ref:2509.00832#pg1>.

Conclusion: Tom: So we’re wrapping up on "Challenges in Non-Polymeric Crystal Structure Prediction: Why a Geometric, Permutation-Invariant Loss is Needed." Essentially, these authors are arguing that predicting those ordered three dee crystal patterns for molecules isn't just a simple math problem anymore; it needs a loss function that understands the actual geometry and ignores how you happen to list the molecules.

Jane: That sounds like they’re trying to fix a fundamental flaw in how we teach these AI models about physical shapes. But what does "permutation invariance" actually mean for someone who isn't deep into math?

Lu: It means the model doesn't care if molecule one comes first or last in the input set S; it just cares about the final arrangement of all those molecules together, which is what matters for the crystal. It’s about learning a pattern, not a sequence.

Meng: From my side, I’m thinking about deployment cost and stability. If this loss function makes training much harder to stabilize at scale, then it might be nice in theory but impractical to ship to production right now.

Tom: Yeah, the paper shows that even with this complex setup, a simple regression model actually beats some of the more complicated flow matching techniques on their test set. That’s a huge practical win for engineers trying to get something working fast.

Jane: So, the authors are pointing toward this specific geometric and permutation invariant loss as what they need to move past those initial hurdles in predicting crystal structures. It seems like a lot of heavy lifting just to define the right way to measure success.

Lu: Exactly. The paper shows that training with these permutation invariant losses gives a big performance boost, sometimes dividing the evaluation metric by up to three point eight compared to just using absolute position losses. That’s concrete evidence for the methodology they proposed.

Meng: A gain of three point eight is significant, but I still see a gap between this sanitized dataset and real-world complexity where things break. We need to see if this holds up when we introduce real molecular diversity that isn't just in the COD-Cluster17 set.

Tom: Well, the authors are being pretty clear about that limitation right there. They state that one big thing is the dataset lacks periodic boundary conditions, which limits how well any model can generalize to structures with infinite repeating patterns.

Jane: So they’re basically saying this approach is a strong foundation for understanding the problem, but we still need to train these models end-to-end so they learn the structure and the conformation together. That’s where we need to go next.

Lu: The implication is that solving this molecular assembly subproblem better is what gets us closer to predicting those complex infinite three dimensional structures for real materials, which is a much bigger goal than just getting a high score on one benchmark.

Tom: It seems like the core message here is that we need to stop treating crystal structure prediction like just another regression problem and start focusing on the actual geometric rules of how molecules pack together.

Jane: Right, so this paper gives us a much clearer idea about what kind of mathematical machinery we need to build if we want to actually tackle predicting real organic crystal structures.

Emmanuel Jehanno, Romain Menegaux, Julien Mairal, Sergei Grudinin

University Grenoble Alpes · inria

cs.LG, cond-mat.mtrl-sci

Submitted: 2025-08-31

Updated: 2026-11-09

Journal ref: Published in Transactions on Machine Learning Research (05/2026)

Code: https://github.com/EmmanuelJhno/SinkFast

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 91/100

The gist: Crystalline structure prediction is an essential prerequisite for designing materials with targeted properties.

Key concepts

Molecular Assembly Problem
This is the core task of packing a finite set of identical rigid molecules into a specific spatial arrangement. The goal is to determine how these individual molecules assemble from an initial state into a final configuration that can then be replicated to form a crystal structure.
Permutation Invariance
This property means the model's output should not change if the order of the input molecules is shuffled. In crystal packing, since all molecules are identical, their specific sequence in the input set does not matter for the final geometric structure.
LGeom Loss Function
This loss function measures distances between relative molecular positions rather than absolute ones. It is designed to be invariant to global transformations and molecule ordering, making it a more robust metric for assessing how well the predicted assembly matches the ground truth crystal structure.

Terminology

Summary

Crystalline structure prediction is an essential prerequisite for designing materials with targeted properties. p. 0

Where the paper says it

grudinin@univ-grenoble-alpes. fr Abstract Crystalline structure prediction is an essential prerequisite for designing materials with targeted properties.

In this work, we focus on the molecular assembly problem, where a set S of identical rigid molecules is packed to form a crystalline structure. p. 0
Where the paper says it

Despite recent advances in computational materials science, accurately predicting three-dimensional non-polymeric crystal structures remains elusive. In this work, we focus on the molecular assembly problem, where a set S of identical rigid molecules is packed to form a crystalline structure.

We propose a better formulation that introduces a loss function capturing key geometric molecular properties while ensuring permutation invariance over S. p. 0
Where the paper says it

However, while recent state-of-the-art methods have increasingly adopted sophisticated techniques, the underlying learning objective remains ill-posed. We propose a better formulation that introduces a loss function capturing key geometric molecular properties while ensuring permutation invariance over S.

Remarkably, we demonstrate that within this framework, a simple regression model already outperforms prior approaches, including flow matching techniques, on the COD-Cluster17 benchmark. p. 0
Where the paper says it

We propose a better formulation that introduces a loss function capturing key geometric molecular properties while ensuring permutation invariance over S. Remarkably, we demonstrate that within this framework, a simple regression model already outperforms prior approaches, including flow matching techniques, on the COD-Cluster17 benchmark, a curated non-polymeric subset of the Crystallography Open Database (COD).

Problem Formulation and Dataset

A non-polymeric crystal is a solid material in which molecules are arranged in a highly ordered pattern repeating in the three spatial dimensions (3D). The molecular assembly subproblem is a simplification of the original problem, where a finite set S of identical rigid molecules is packed together from a state Sinitial into a state Sfinal which forms a pattern that can be then replicated in space into a crystal. p. 2

Where the paper says it

To solve this very challenging task, one can take a number of approximations and hypotheses. The molecular assembly subproblem is a simplification of the original problem, where a finite set S of identical rigid molecules is packed together from a state Sinitial into a state Sfinal which forms a pattern that can be then replicated in space into a crystal.

The dataset used in this work is the COD-Cluster17 assembly dataset introduced by Liu et al. p. 2
Where the paper says it

We propose an efficient and model-agnostic way to guide any machine learning model with physical knowledge of the task. Dataset In this work we use the COD-Cluster17 assembly dataset introduced by Liu et al.

[2024a], specifically constructed for the task of non-polymeric crystal structure prediction. p. 2
Where the paper says it

Dataset In this work we use the COD-Cluster17 assembly dataset introduced by Liu et al. [2024a], specifically constructed for the task of non-polymeric crystal structure prediction.

This dataset contains 111k assemblies and is a simplified, sanitized version of the 507k crystals from the real world Crystallography Open Database (COD) [Gražulis et al. p. 2
Where the paper says it

To the best of our knowledge, this is currently the only available benchmark for this task. This dataset contains 111k assemblies and is a simplified, sanitized version of the 507k crystals from the real world Crystallography Open Database (COD) [Gražulis et al.

, 2009].

Loss Functions and Invariance

We propose two rigid-body loss functions: LRMSD, which operates on absolute positions of molecules, and LGeom, which extends this formulation to relative molecular positions. The geometric loss LGeom is defined as 1/M-1 sum over i=1 to M-1 of LRMSD(T−1j,pred ◦ Ti,pred, T−1j,gt ◦ Ti,gt). p. 5

Where the paper says it

(9) We can then extend the above expression to the comparison of M − 1 molecules to a reference one. Without loss of generality, we can assume it is the Mth molecule and define the geometric loss LGeom as follows: LGeom (Tpred, Tgt) = 1 M − 1 M X−1 i=1 LRMSD(T −1 M,pred ◦ Ti,pred, T −1 M,gt ◦ Ti,gt).

We propose a differentiable soft matching objective that is invariant to global geometric transformations and to the order permutation of repeated molecular units. p. 1
Where the paper says it

Permutation-invariant loss. We propose an effective differentiable soft matching objective that is invariant to global geometric transformations and to the order permutation of repeated molecular units.

This is achieved by minimizing L∗ ML = min P sum ij C Lij. p. 4
Where the paper says it

Formally it is defined by the linear sum assignment problem. Let P be a boolean pairing matrix in which Pij = 1 if and only if molecule i from ground truth assembly is mapped with molecule j in the predicted assembly: L ∗ := min P X ij C L ij .

Pij with Pij ∈ [0, 1] s. p. 4
Where the paper says it

Formally it is defined by the linear sum assignment problem. Let P be a boolean pairing matrix in which Pij = 1 if and only if molecule i from ground truth assembly is mapped with molecule j in the predicted assembly: L ∗ := min P X ij C L ij .

t. P · 1 = P⊤ · 1 = 1 and P ≥ 0.

Experimental Results and Comparison

Our method SinkFast outperforms prior approaches on both the COD-Cluster17-5K subset and the full dataset by a significant margin. Notably, our method is the only deep learning method that outperforms PackMol on both datasets. p. 7

Where the paper says it

Then, we show that our method outperforms all other baselines on both 5k subset and the full dataset by a significant margin. Notably our method is the only deep learning method that outperforms PackMol on both datasets.

Ablation studies show that training with LGeom is the solution for the geometric relative packing loss metric L∗Geom (Eq. p. 8
Where the paper says it

02 (2) both previous absolute losses fail to perform on the geometric relative packing loss metric L ∗ Geom (Eq. 10) while training with it is the solution.

10). Furthermore, training with S-permutation invariant losses shows a significant gain, dividing by up to 3. p. 8
Where the paper says it

Moreover, these results along with ablation in Appendix E show a significant gain, dividing by up to 3. 8 the evaluation metric, when training with the S-permutation invariant losses, while not being useful in flow matching.

8 the evaluation metric compared to absolute losses. p. 8
Where the paper says it

Moreover, these results along with ablation in Appendix E show a significant gain, dividing by up to 3. 8 the evaluation metric, when training with the S-permutation invariant losses, while not being useful in flow matching.

Limitations and Future Directions

A fundamental limitation of COD-Cluster17 is the absence of periodic boundary conditions, which hinders the generalisation capability of any model. The current approximation is valid, however future models should be trained end-to-end, jointly learning conformation and crystal structure prediction. p. 9

Where the paper says it

2 the rigid body approximation and our model’s performance on a different real-world test set, in which we generate new molecular conformations through RDKit [Landrum et al. We conclude that the current approximation is valid, however future models should be trained end-to-end, jointly learning conformation and crystal structure prediction.

We believe our study helps to identify key challenges and can serve as a foundation for future work in organic crystal structure prediction. p. 9
Where the paper says it

We conclude that the current approximation is valid, however future models should be trained end-to-end, jointly learning conformation and crystal structure prediction. 6 Conclusion In this paper we have focused on a simpler subtask of the complex organic crystal structure prediction.

Code Availability

Our code is available at https://github.com/EmmanuelJhno/SinkFast. p. 0

Where the paper says it

Our code is available at https://github. com/EmmanuelJhno/SinkFast 1 Introduction Generative modeling and deep learning have enabled rapid progress in the understanding and design of materials, molecules, and drugs.

The common environment packages are released with the code through a conda environment. p. 0
Where the paper says it

Remarkably, we demonstrate that within this framework, a simple regression model already outperforms prior approaches, including flow matching techniques, on the COD-Cluster17 benchmark, a curated non-polymeric subset of the Crystallography Open Database (COD). Our code is available at https://github.

p. 0
Where the paper says it

Remarkably, we demonstrate that within this framework, a simple regression model already outperforms prior approaches, including flow matching techniques, on the COD-Cluster17 benchmark, a curated non-polymeric subset of the Crystallography Open Database (COD). Our code is available at https://github.

The common environment packages are released with the code through a conda environment.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Physics grounded loss introduction allows for improved prediction of crystalline structures by introducing a domain-specific rigid-body, model agnostic loss, grounded in physical principles. This enables models to learn representations that are physically meaningful for molecular assembly.

  2. Permutation invariant loss ensures that the learning objective is invariant to global geometric transformations and to the order permutation of repeated molecular units. This addresses the issue where all ASUs in a crystal are geometrically, physically and chemically equivalent, leading to more robust predictions.

  3. Rigid RMSD loss (LRMSD) provides a more suitable rigid-body transformation loss that is strictly equivalent to the RMSD atom metric in eq. 2 by incorporating an inertia tensor, naturally providing a weight between rotation and translation contributions without additional hyperparameters.

  4. Geometric loss (LGeom) is proposed as an assembly transformation-invariant PM metric for these molecules, which reflects the relative packing of molecules rather than memorizing their absolute positions, and this invariance is achieved through the differentiable optimal assignment problem.

  5. The system can perform a linear sum assignment problem during training using a differentiable version via the Sinkhorn algorithm, which provides feedback to the model with multiple possible assignments weighted by P, behaving like a probability map.

Abstract

Crystalline structure prediction is an essential prerequisite for designing materials with targeted properties. Yet, it is still an open challenge in materials design and drug discovery. Despite recent advances in computational materials science, accurately predicting three-dimensional non-polymeric crystal structures remains elusive. In this work, we focus on the molecular assembly problem, where a set S of identical rigid molecules is packed to form a crystalline structure. Such a simplified formulation provides a useful approximation to the actual problem. However, while recent state-of-the-art methods have increasingly adopted sophisticated techniques, the underlying learning objective remains ill-posed. We propose a better formulation that introduces a loss function capturing key geometric molecular properties while ensuring permutation invariance over S. Remarkably, we demonstrate that within this framework, a simple regression model already outperforms prior approaches, including flow matching techniques, on the COD-Cluster17 benchmark, a curated non-polymeric subset of the Crystallography Open Database (COD). Our code is available at https://github.com/inria-thoth/SinkFast.

Sources

Related papers