Computed materials proposals depart from the structural memory of experimental discovery

2026

Read Computed materials proposals depart from the structural memory of experimental discovery (2026) 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-2606-30967.txt.

Introduction

Dan Nguyen, Karen Cao and four colleagues ask how far computed crystal proposals stray from how real materials were found. Computer methods now propose millions of inorganic structures. Yet no one had a calibrated sense of how experimental chemistry explores. The team placed 167,500 entries from a major crystal database in a similarity space. They split the space into groups and replayed discovery over time. Real discovery hugged what already existed: 82.9% of new formulas landed in existing groups.

The share of new groups fell from 40.2% in the 1930s to 2.6% in the 2010s. The groups also picked out nine golden ages, from cuprates to battery cathodes. The team projected computed sets onto the frozen past maps. Every computed set departed from the real basins, not just the AI-generated ones. They propose a score, built from closeness plus precedent, to triage proposals.

Catalog record

YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE43 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2606.30967

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Computed materials proposals depart from the structural memory of experimental discovery Abstract Authors: Dan Nguyen 1,† , Karen Cao 2,† , Brian Chu 1 , Nick Lemoff 1 , Paul Kienzle 3 , William Ratcliff II 3,4,5,∗ 1 University of California, Berkeley · 2 California Institute of Technology · 3 NIST Center for Neutron Research, National Institute of Standards and Technology, Gaithersburg, MD 20899, USA · 4 Department of Materials Science and Engineering, University of Maryland, College Park, MD 20742, USA · 5 Department of Physics, University of Maryland, College Park, MD 20742, USA † These authors contributed equally to this work. Corresponding author: [email protected] Abstract Generative AI and high-throughput DFT pipelines propose millions of inorganic crystal structures, but lack a calibrated reference frame against experimentally realized chemistry. Here we embed 167,500 Inorganic Crystal Structure Database entries in a continuous structural-similarity space, partition it into graph communities, and replay them in time. Experimental discovery shows strong structural memory: 82.9% of new formulas enter pre-existing communities; new-community formation falls from 40.2% (1930s) to 2.6% (2010s). The communities are chemically meaningful, positively identifying nine textbook field-defining renaissances, including cuprates, colossal-magnetoresistance manganites, MAX phases, and Li-ion battery cathodes. Projecting GNoME, MatterGen-public, Materials Project, JARVIS-DFT, and Alexandria-PBE into frozen historical maps yields a cutoff-robust ordering: held-out ICSD > MatterGen > {GNoME ≈ MP-theoretical} > JARVIS > Alexandria. Structural departure from experimental basins is not specific to generative AI but general across the tested computed sets. Combining structural proximity with reduced-formula precedent defines a historical synthesizability prior for triaging computed materials. Main The current wave of computational materials science has been reshaped by large-scale claims. GNoME reported the discovery of 2.2 million new theoretically-stable crystal structures using a graph-network approach 1 ; MatterGen released a generative diffusion model for crystals with property-targeted conditioning 2 ; the Materials Project (MP) 3 , the Open Quantum Materials Database (OQMD) 4 , JARVIS-DFT 5 , and Alexandria-PBE 6 have together computed millions of theoretical structures via data-mined ionic substitution 7 and enthalpy screening. These claims have prompted equally pointed skepticism. Cheetham and Seshadri 8 argued that many GNoME proposals are better described as trivial dopant variants, symmetry-broken realizations, or chemically implausible high-element-count compositions rather than synthesizable new materials. The debate has been substantive but scattered: each system has been compared against partial references (often just MP, sometimes just exact-prototype matches), and there has been no single calibrated reference frame in which all of these proposals can be evaluated against the chemistry that experimentalists have actually realized. We provide that calibration for the inorganic crystal landscape. We embed all 167,500 Inorganic Crystal Structure Database (ICSD) entries 9 into a continuous structural-similarity space (matminer descriptors 10 augmented with three rounds of message passing on a Voronoi neighbor graph), partition the result into Louvain graph communities, replay communities through publication time, and project five publicly-released external structure samples into the same frozen historical geometry under multiple held-out training cutoffs (Extended Data Fig. 1; Extended Data Table 1). We use frontier throughout to mean “distant from historically occupied structural basins in the learned representation” — a structural-distance claim, not a claim of thermodynamic stability or chemical novelty. The framework distinguishes the structure-level frontier rate at a given cutoff from the formula-level stepping-stone rate; both are facets of path dependence but they refer to different units of analysis. Experimental discovery is heavily path-dependent Two complementary measurements characterize experimental discovery (Fig. 1). At the formula level, restricting to the 16,582 newly reported ICSD reduced formulas with corresponding entries in the Toyota Research Institute (TRI) thermodynamic-stability network 11 (88.1% of the 18,821 TRI-shared formulas, the remainder lacking complete year or community-birth metadata), 13,739 (82.9%) enter pre-existing structural communities; only 2,682 (16.2%) coincide with community birth, and just 161 (1.0%) precede their associated community in publication time. At the decade level, the share of ICSD entries that open new structural communities falls from 40.2% in the 1930s to 2.6% in the 2010s, while same-community attachment rises from 60.9% to 88.0%. A timestamp-shuffle null model (200 random permutations of publication years over the same 163,339 community-assigned entries, holding the final community labels fixed) rules out the partition-artifact reading: the observed birth ratio is below the null 5–95% band in the 1910s–1960s and above it in the 1970s–2010s, indicating two distinct path-dependence regimes — densification of established basins in the early record and frontier-pushing-above-chance in the late record — rather than mechanical front-loading (Extended Data Fig. 2). Continuous structural basins also resist the discrete-prototype-counting trap that inflates novelty in many computational claims. Among the ten largest graph communities, the mean number of distinct space groups per community is 44.2 (range 16–86): a single continuous basin absorbs many crystallographic relabellings of the same underlying topology. Trivial symmetry-breaking and supercell choice are inside the same basin in our metric, not outside it. Communities track field-defining scientific events A direct test of whether the framework’s communities are chemically meaningful objects rather than statistical groupings: do they show abrupt growth aligned with documented field-defining scientific events? Two well-known renaissances confirm this with quantitative clarity (Fig. 2a–b). The cuprate community (community 6425) first appears after Bednorz & Müller’s December 1986 paper 12 : zero of its 378 members were published before 1986, the earliest entry is from 1987, and the post-1986 rate is 22.3 entries per year. The colossal-magnetoresistance (CMR) manganite community (community 160) shows a 32× fold-change in member rate after Jin et al.’s 1994 thin-film report 13 , with only four pre-1994 entries against 22.3 entries per year post-1994. Both communities specifically capture the doped variants — La₁₋ₓSrₓMnO₃, Ba₂₋ₓSrₓYCu₃O₇₋δ — that became active research targets after the renaissance events; the pure parent compounds (LaMnO₃ from Goodenough’s foundational work, La₂CuO₄, YBa₂Cu₃O₇) sit in different communities or HDBSCAN noise as crystallographic outliers. The framework therefore discriminates between Goodenough-era foundational LaMnO₃ and the post-1994 doping campaigns motivated specifically by colossal magnetoresistance, a fine-grained-validity test that the communities track research focus rather than nominal structure type. Converting this into an unbiased survey, we scored every production community of size ≥ 50 for its strongest pre/post fold-change over symmetric ±10-year windows scanning candidate event years 1970–2010 (Fig. 2c). The step-change scores themselves are not produced by the underlying year distribution alone: under 200 year-shuffle permutations of the same 148,425 community-assigned ICSD entries (community labels held fixed, year-to-entry assignments permuted), all 20 observed top-20 scores exceed the shuffled 99th-percentile envelope at the same rank — the top observed score exceeds the corresponding shuffled p99 by approximately 37× and the rank-20 score by approximately 9× (Supporting Information §S8.4; Extended Data Table 2). On top of this null-rejection, all 20 top-ranked step-change communities have an identifiable chemistry interpretation, of which nine are textbook field-defining renaissances (Supporting Information §S8.4): Sm₂Fe₁₇N permanent magnets (Coey ~1990) 14 , CMR manganites (Jin 1994), SOFC perovskite cathodes (mid-1990s), MAX phases (Barsoum 1996+) 15 , NaCoO₂ thermoelectrics (Terasaki 1997) 16 , ordered double perovskites for spintronics (Kobayashi Sr₂FeMoO₆ 1998) 17 , Li-ion battery cathodes (Mizushima/Goodenough 1980 LiCoO₂ 18 , Sony commercialization 1991), dilute magnetic semiconductors (Dietl 2000) 19 , and post-1986 layered Ruddlesden-Popper cuprates. The remaining ranks reflect program-driven systematic surveys and applied-chemistry surges of equivalent statistical strength — not missed renaissances — including the

Read the complete text