Chemical-space completeness: a new strategy for crystalline materials exploration
Fengyu Xie, Ruoyu Wang, Taoyuze Lv, Yuxiang Gao, Hongyu Wu, Zhicheng Zhong · 2025
Read Chemical-space completeness: a new strategy for crystalline materials exploration by Fengyu Xie, Ruoyu Wang, Taoyuze Lv, Yuxiang Gao, Hongyu Wu, Zhicheng Zhong (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-12420.txt.
Introduction
Fengyu Xie and five co-authors propose a loop to explore the chemical space of one bounded materials system until it is effectively complete. General atomistic models trade accuracy and speed for breadth across all chemistry. Within a single system, that trade can be eased because the chemistry is limited.
Their strategy couples a crystal-structure generator with machine-learned force fields in repeated rounds of generation, evaluation and refinement. Testing on a lithium-phosphorus-sulfur system, they produce tens of thousands of candidate structures, including many stable, unique and novel ones. The variety of local environments saturates after a few rounds and model errors converge, giving a practical measure of when exploration is complete.
Catalog record
| CREATOR | Fengyu Xie, Ruoyu Wang, Taoyuze Lv, Yuxiang Gao, Hongyu Wu, Zhicheng Zhong |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 40 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2511-12420
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Chemical-space completeness: a new strategy for crystalline materials exploration Fengyu Xie, Ruoyu Wang, Taoyuze Lv, Yuxiang Gao, Hongyu Wu, Zhicheng Zhong Abstract The emergence of deep learning has brought the long-standing goal of comprehensively understanding and exploring crystalline materials closer to reality. Yet, universal exploration across all elements remains hindered by the combinatorial explosion of possible chemical environments, making it difficult to balance accuracy and efficiency. Crucially, within any finite set of elements, the diversity of short-range bonding types and local geometric motifs is inherently limited. Guided by this chemical intuition, we propose a chemical-system–centric strategy for crystalline materials exploration. In this framework, generative models are coupled with machine-learned force fields as fast energy evaluators, and both are iteratively refined in a closed-loop cycle of generation, evaluation, and fine-tuning. Using the Li–P–S ternary system as a case study, we show that this approach captures the diversity of local environments with minimal additional first-principles data while maintaining structural creativity, achieving closed-loop convergence toward chemical completeness within a bounded chemical space. We further demonstrate downstream applications, including phase-diagram construction, ionic-diffusivity screening, and electronic-structure prediction. Together, this strategy provides a systematic and data-efficient framework for modeling both atomistic and electronic structures within defined chemical spaces, bridging accuracy and efficiency, and paving the way toward scalable, AI-driven discovery of crystalline materials with human-level creativity and first-principles fidelity. keywords Chemical system, Materials discovery, Crystal structure generation, Machine-learning force-field, Crystalline materials † † equal-contributors: These authors contributed equally to this work. † † equal-contributors: These authors contributed equally to this work. 1 Introduction Exploring and understanding materials across the periodic table remains a grand challenge in modern materials science. The ultimate goal is to achieve universal, accurate, and efficient modeling of atomic and electronic structures for all compositions under realistic temperature and pressure conditions, spanning stable and metastable phases Curtarolo et al. 2013 ; Jain et al. 2013 , defects Freysoldt 2014 , and interfaces Harmer 2011 . Such capability would transform not only materials simulation workflows but also the next-generation AI-native databases for materials design. Recent advances in deep generative models for crystals (e.g., MatterGen Zeni et al. 2025 , DiffCSP Jiao et al. 2023 , ADiT Joshi et al. 2025 , Uni3DAR Lu et al. 2025 ) and machine-learned force fields (MLFFs) (e.g., DPA-3 Zhang et al. 2025 , MatterSim Yang et al. 2024 , MACE Batatia et al. 2022 , CHGNet Deng et al. 2023 ) have brought this vision closer to reality. Generative models explore latent structural spaces beyond template filling Schmidt et al. 2022 or elemental substitution approaches Hautier et al. 2011 ; Merchant et al. 2023 , and offer composition variability compared to traditional crystal-structure prediction methods Pickard and Needs 2011 ; Oganov and Glass 2006 ; Glass et al. 2006 ; Wang et al. 2010 ; Wang et al. 2012 . Meanwhile, MLFFs provide near-DFT accuracy in energy and force evaluation at orders-of-magnitude lower computational cost. In principle, integrating these capabilities could yield a universal autonomous engine for structure generation, evaluation, and property prediction. However, a fundamental challenge remains. In all-element exploration, a difficult trade-off between accuracy and efficiency arises from the vast combinatorial diversity of element combinations and bonding environments. Large foundational models such as UMA Wood et al. 2025 exhibit impressive cross-chemistry generalizability. Yet, their universal validity across arbitrary element combinations and bonding environments, particularly in defected systems, has not been systematically established. Given their enormous training and inference cost, such large models are inefficient for most real-world research scenarios, where rapid iteration, domain-specific fine-tuning and multi-fidelity adaptation are often essential. In contrast, smaller models such as MatterSim, DPA-3, or MACE enable flexible fine-tuning and efficient inference, but often fall short of universal accuracy. Without simultaneously ensuring both efficiency and accuracy, high-throughput exploration remains fundamentally restricted. Therefore, a new strategy that enables sufficiently accurate atomistic and electronic-structure modeling at manageable computational cost is of pivotal importance. The nature of chemical bonds itself provides key intuition. As suggested by the Materials Project Horton et al. 2025 , the majority of realistic crystalline materials can be grouped into chemical systems comprising a small ( <7 ) and finite set of elements. Within such a chemical system, the diversity of short-range bonding types and local geometries is inherently limited by valence and coordination constraints. For instance, carbon atoms predominantly adopt sp^{3} , sp^{2} , or sp^{1} hybridization even in amorphous networks Robertson 1986 ; Robertson and O’Reilly 1987 , while silicon in silicates typically forms SiO 4 tetrahedra Liebau 1985 . Once these local building blocks are enumerated, the remaining structural diversity largely arises from their spatial arrangements. Together, these observations suggest that chemical-system-level completeness may be attainable without pursuing all-element universality, but rather through a divide-and-conquer strategy applied to individual chemical spaces. Achieving completeness in a target chemical space should also be sufficient in most real-world applications. Grounded in this intuition, we introduce a chemical-system–centric strategy for crystalline materials exploration. Rather than seeking all-element universality, our framework targets chemical-space completeness, i.e., achieving structural and property coverage within a defined chemical system. This is realized through an iterative fine-tuning and active-learning loop that couples generative models and MLFFs: the generative model proposes candidate structures, the MLFF evaluates their energetics, and selected DFT data are used to refine both models (see Figure 1 and Supplementary Figure 1). By focusing on finite bonding environments and compositional boundaries, the workflow rapidly converges toward meV-level accuracy and saturates the diversity of local motifs, achieving closed-loop convergence toward chemical completeness at manageable computational cost. We demonstrate this strategy on the Li–P–S ternary system, a prototypical yet chemically intricate solid-state electrolyte chemistry Sakuda et al. 2013 ; Zhang et al. 2019 . Phosphorus and sulfur form diverse bonding networks, including tetrahedral PS 4 units, P–P–linked trigonal motifs, and bridging or polymeric sulfide species, generating complex anionic frameworks coupled to lithium coordination. This system provides a stringent test case to examine whether, despite its rich anionic diversity and glass-forming tendency, the underlying short-range chemical environments remain finite and learnable. With only \sim 1500 CPU hours (32-core Xeon Gold) and \sim 800 GPU hours (NVIDIA V100), our framework completed seven exploration cycles, generating 70,000 candidate structures, more than 10,000 of which are stable–unique–novel. The MLFF achieved meV-level accuracy in energy and force within just two to three iterations, and structural analysis revealed rapid saturation of local environments, suggesting a quantifiable notion of chemical completeness within the system. Notably, although the generative model was trained exclusively on existing structures from the Materials Project Jain et al. 2013 and Alexandria databases Schmidt et al. 2022 ; Schmidt et al. 2023 –datasets often regarded as chemically comprehensive–our framework autonomously rediscovered a broad range of chemically plausible P–S anionic motifs and metastable phases beyond those databases. These include the (PS 3 ) {}_{3}^{3-} trimer, the P 3 S {}_{8}^{3-} ring, two isomers of the P 2 S {}_{8}^{2-} ring, and polymeric (PS 4 ) {}_{n}^{n-} chains. Their existence and internal bonding patterns, such as P–P and S–S linkages, are consistent with historical experiments, underscoring their chemical credibility. Together, these results highlight our framework’s ability to creatively and rationally infer bonding rules from minimal priors and to reconstruct a more complete representation in