arxiv: 2604.08972 · v2 · submitted 2026-04-10 · ❄️ cond-mat.mtrl-sci

Recognition: unknown

Selective Random Structure Search (SRSS): Unbiased Exploration of Polymorphs in Crystals

Jiexi Song , Diwei Shi , Aixian She , Chongde Cao , Fengyuan Xuan

Authors on Pith no claims yet

Pith reviewed 2026-05-10 17:24 UTC · model grok-4.3

classification ❄️ cond-mat.mtrl-sci

keywords crystal structure predictionpolymorph discoverymachine learning potentialsunbiased samplingSiCNbSe2GaN nanotubes

0 comments

The pith

SRSS recovers known crystal ground states and uncovers new dynamically stable polymorphs in bulk, 2D, and 1D materials using only CPU resources.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents Selective Random Structure Search as a method for crystal structure prediction that avoids the bias of using known prototypes for seeding. It generates structures randomly under symmetry constraints, selects diverse candidates based on features, and evaluates them quickly with universal machine learning interatomic potentials for energy and stability. This framework is tested on materials like silicon carbide, barium platinum arsenide, niobium diselenide layers, and gallium nitride nanotubes, where it finds the expected lowest energy forms plus many new stable ones that had not been reported. A reader might care because traditional methods often miss unconventional structures, and this approach makes comprehensive exploration of possible crystal forms more practical and less computationally demanding.

Core claim

SRSS is a high-throughput unbiased framework that combines symmetry-constrained random generation with feature-based diversity selection and rapid relaxation and stability evaluation via universal machine-learning interatomic potentials. When applied to bulk SiC and BaPtAs, 2D NbSe2, and 1D GaN nanotubes, it recovers known ground states while revealing numerous previously unreported, dynamically stable polymorphs such as complex cage-like SiC polytypes, low-energy BaPtAs polymorphs, a semiconducting orthorhombic phase of 2D-NbSe2, and distinct armchair and zigzag GaN nanotubes. The workflow operates efficiently on standard CPU resources without GPU acceleration.

What carries the argument

Selective Random Structure Search (SRSS), which uses symmetry-constrained random structure generation, feature-based diversity selection, and universal machine-learning interatomic potentials for relaxation and dynamical stability checks.

If this is right

SRSS enables mapping of the full crystal stability landscape for diverse systems.
New polymorphs can be identified in resource-limited settings without GPU acceleration.
The method bridges exhaustive search and computational feasibility for hypothesis-free discovery.
It successfully recovers ground states while revealing unreported stable phases in the tested materials.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Applying SRSS to a wider range of chemical compositions could uncover polymorphs with novel properties for applications in electronics or energy storage.
Cross-validating the uMLIP predictions with density functional theory calculations for the new structures would strengthen confidence in the discoveries.
The diversity selection step might be adapted to other search algorithms to reduce bias in materials discovery.

Load-bearing premise

That the universal machine-learning interatomic potentials accurately predict the energies and dynamical stability of both known and newly discovered polymorphs without significant systematic errors.

What would settle it

A direct comparison with higher-accuracy calculations, such as density functional theory, showing that one of the newly predicted polymorphs is actually unstable or that a known stable phase is missed would falsify the reliability of the SRSS workflow.

read the original abstract

Crystal structure prediction has traditionally relied on prototype-based seeding, approaches that often bias sampling toward known low-energy basins and overlook metastable polymorphs with unconventional symmetries. Here, we introduce Selective Random Structure Search (SRSS), a high-throughput, unbiased framework designed to explore the configurational space of crystalline materials across all dimensions. SRSS combines symmetry-constrained random generation with feature-based diversity selection and rapid relaxation and stability evaluation via universal machine-learning interatomic potentials (uMLIPs). Applied to diverse systems, including bulk system SiC and BaPtAs, 2D layered compounds NbSe2, and 1D nanotubes GaN, SRSS successfully recovers known ground states while revealing numerous previously unreported, dynamically stable polymorphs. Notable discoveries include complex cage-like SiC polytypes, low-energy BaPtAs polymorphs beyond experimental records, a semiconducting orthorhombic phase of 2D-NbSe2, and distinct armchair/zigzag GaN nanotubes. Crucially, the entire workflow operates efficiently on standard CPU resources without GPU acceleration, demonstrating that rigorous, hypothesis-free polymorph discovery is accessible even in resource-limited settings. SRSS thus establishes a robust, scalable platform for mapping the full landscape of crystal stability, bridging the gap between exhaustive search and computational feasibility.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

SRSS gives a practical CPU-only workflow for unbiased polymorph searches that recovers known states, but the new structures rest on uMLIP stability alone without DFT checks.

read the letter

The core contribution is a clean combination of symmetry-constrained random generation, feature-based selection for diversity, and fast uMLIP relaxation plus phonon evaluation. This lets the method run entirely on standard CPUs and still claim to recover ground states while turning up new candidates in SiC, BaPtAs, NbSe2, and GaN nanotubes. The workflow description is coherent and addresses a real bias problem in older prototype-seeded searches.

Referee Report

3 major / 2 minor

Summary. The manuscript introduces Selective Random Structure Search (SRSS), an unbiased high-throughput framework for crystal polymorph exploration. SRSS combines symmetry-constrained random structure generation, feature-based diversity selection, and rapid relaxation plus dynamical stability assessment using universal machine-learning interatomic potentials (uMLIPs). Demonstrated on bulk SiC and BaPtAs, 2D NbSe2, and 1D GaN nanotubes, the method recovers known ground states while identifying multiple previously unreported dynamically stable polymorphs (e.g., cage-like SiC polytypes, low-energy BaPtAs variants, orthorhombic 2D-NbSe2, and distinct GaN nanotube configurations), all executed efficiently on standard CPU hardware without GPU acceleration.

Significance. If the uMLIP-based energy and phonon predictions prove reliable, SRSS would provide a practical, resource-efficient route to hypothesis-free polymorph mapping that avoids prototype bias and scales to diverse dimensionalities. The CPU-only workflow and explicit recovery of known states plus new candidates represent a concrete advance in accessibility for crystal structure prediction.

major comments (3)

[Results] Results section (SiC and NbSe2 examples): Dynamical stability of all newly reported polymorphs is declared solely from uMLIP phonon spectra after relaxation; no DFT single-point energies, force convergence checks, or DFT phonon calculations are described for the novel candidates. This leaves the central claim of 'numerous previously unreported, dynamically stable polymorphs' vulnerable to uMLIP extrapolation errors outside the training manifold.
[Methods and Results] Methods and Results sections: No quantitative metrics (e.g., success rate of ground-state recovery, energy ranking errors relative to DFT, or fraction of generated structures that survive stability filters) or baseline comparisons (standard RSS without diversity selection, or prototype-seeded searches) are supplied. The abstract's assertion of successful recovery therefore lacks the numerical support needed to evaluate performance.
[Results] Application to BaPtAs and GaN nanotubes: The reported low-energy polymorphs and distinct nanotube chiralities are presented as discoveries, yet the workflow description indicates stability evaluation occurs exclusively via uMLIP; absence of any higher-level validation for these unconventional structures constitutes a load-bearing gap for the discovery claims.

minor comments (2)

[Abstract] Abstract: Inclusion of at least one concrete metric (e.g., energy difference to known ground state or number of new stable structures per system) would strengthen the summary of results.
[Methods] Notation: The term 'feature-based diversity selection' is used without an explicit definition or reference to the feature vector construction in the methods; a short equation or pseudocode would improve reproducibility.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed report. We address each major comment below and will revise the manuscript to incorporate additional validation and quantitative analysis as outlined.

read point-by-point responses

Referee: [Results] Results section (SiC and NbSe2 examples): Dynamical stability of all newly reported polymorphs is declared solely from uMLIP phonon spectra after relaxation; no DFT single-point energies, force convergence checks, or DFT phonon calculations are described for the novel candidates. This leaves the central claim of 'numerous previously unreported, dynamically stable polymorphs' vulnerable to uMLIP extrapolation errors outside the training manifold.

Authors: We agree that reliance on uMLIP alone for dynamical stability leaves the discovery claims open to potential extrapolation concerns. In the revised manuscript we will add DFT single-point energy calculations and force convergence checks for all reported low-energy novel polymorphs in the SiC and NbSe2 systems. We will also perform DFT phonon calculations for a representative subset of the new structures (selected by uMLIP energy) and include a dedicated discussion of uMLIP accuracy for these chemistries, drawing on published benchmarks. These additions will directly address the risk of extrapolation errors while preserving the CPU-efficient screening workflow. revision: yes
Referee: [Methods and Results] Methods and Results sections: No quantitative metrics (e.g., success rate of ground-state recovery, energy ranking errors relative to DFT, or fraction of generated structures that survive stability filters) or baseline comparisons (standard RSS without diversity selection, or prototype-seeded searches) are supplied. The abstract's assertion of successful recovery therefore lacks the numerical support needed to evaluate performance.

Authors: We acknowledge that the current manuscript lacks explicit quantitative performance metrics and baseline comparisons. We will add a new subsection to the Methods section that reports (i) the success rate of ground-state recovery across independent SRSS runs for each material, (ii) energy ranking errors relative to available DFT data, and (iii) the fraction of generated structures that pass the stability filters. We will also include direct comparisons to standard random structure search without the feature-based diversity selection step, quantifying differences in recovered structures and computational efficiency. These metrics will provide the numerical support required to evaluate the method's performance. revision: yes
Referee: [Results] Application to BaPtAs and GaN nanotubes: The reported low-energy polymorphs and distinct nanotube chiralities are presented as discoveries, yet the workflow description indicates stability evaluation occurs exclusively via uMLIP; absence of any higher-level validation for these unconventional structures constitutes a load-bearing gap for the discovery claims.

Authors: We recognize that the absence of higher-level validation for the BaPtAs polymorphs and GaN nanotube configurations weakens the discovery claims. In revision we will perform DFT single-point energy calculations on the reported low-energy BaPtAs structures and on the distinct GaN nanotube chiralities. We will add a discussion of uMLIP reliability for these specific systems, supported by literature benchmarks on related compounds, and will include DFT phonon verification for the lowest-energy nanotube candidates. These steps will supply the requested higher-level validation for the unconventional structures. revision: yes

Circularity Check

0 steps flagged

No significant circularity in SRSS algorithmic framework

full rationale

The paper introduces SRSS as an algorithmic procedure combining symmetry-constrained random structure generation, feature-based diversity selection, uMLIP relaxation, and phonon-based stability evaluation. Its claims consist of empirical results obtained by applying this workflow to external benchmark systems (SiC, BaPtAs, NbSe2, GaN nanotubes), where known ground states are recovered and new polymorphs are identified. No derivations, equations, or first-principles predictions are presented that reduce to the method's own inputs by construction. There are no self-definitional steps, fitted inputs relabeled as predictions, or load-bearing self-citations that render the central results tautological. The approach is self-contained as a computational search tool whose outputs are independently testable against known structures and external validation.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the accuracy of existing universal ML interatomic potentials for stability assessment; no new physical entities or fitted parameters are introduced in the abstract.

axioms (1)

domain assumption Universal machine-learning interatomic potentials can reliably predict relative energies and dynamical stability of crystal polymorphs across the tested material classes.
The workflow depends on uMLIPs for rapid relaxation and stability evaluation without detailing accuracy benchmarks or error bounds for the discovered structures.

pith-pipeline@v0.9.0 · 5541 in / 1251 out tokens · 88309 ms · 2026-05-10T17:24:29.924850+00:00 · methodology

discussion (0)

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

GEWUM: General Exploration Workflow for the Utopia of Materials: A Unified Platform for Automated Structure Generation, Selection, and Validation
cond-mat.mtrl-sci 2026-04 unverdicted novelty 5.0

GEWUM is a unified open-source platform that combines selective random structure search with universal machine learning interatomic potentials to automate materials structure prediction, selection, and validation.

Reference graph

Works this paper leans on

3 extracted references · 3 canonical work pages · cited by 1 Pith paper

[1]

The SRSS proceeds in six stages

Methods 2.1 Overall workflow The goal of selective random structure search (SRSS) workflow is to navigate the vast configuration space by combining high - throughput symmetry constrained random crystal generation with diversity selection and physics - driven filtering, ultimately yielding a compact set of thermodynamically and dynamically favorable struct...

work page
[2]

These systems were selected to validate the versatility and unbiased nature of the SRSS framework across diverse dimensionalities and chemi cal complexities

Results & Discussions For computational demonstration, we apply our SRSS to the following ca ses: 3D - SiC, 3D - BaPtAs, 2D - NbSe 2 , and 1D - GaN. These systems were selected to validate the versatility and unbiased nature of the SRSS framework across diverse dimensionalities and chemi cal complexities. Specifically, SiC serves as a benchmark for polyty...

work page
[3]

Figure 1 illustrates the Uniform Manifold Approximation and Projection (UMAP) dimensionality reduction plots of structures selected from initial pool based on various combinations

or HDBSCAN clustering . Figure 1 illustrates the Uniform Manifold Approximation and Projection (UMAP) dimensionality reduction plots of structures selected from initial pool based on various combinations. 41 It is evident that simple descriptors achieve a relatively even coverage of the initial structural characteristics compared to the more complex S O A...

work page doi:10.5281/zenodo.19446966 2091