Data-driven candidate structure ensemble generation beyond the convex hull with crystALL approach
Abstract
Real materials consist of multiple stable or metastable states under real-world conditions, existing as statistical ergodic ensembles of configurations rather than ideal crystals. Yet, data-driven computational materials discovery increasingly biases itself toward finding "new stable structures" from ground-state DFT calculations alone. This limits our long-term ability to discover materials whose properties emerge from metastable phases, to explain experimental observations, or to model thermodynamic behavior comprehensively. Predicting ergodic sets of atomic configurations of known and hypothetical solids is a fundamental challenge in theoretical crystallography, approached via many complementary methods. We discuss this conceptually and introduce crystALL, a pragmatic and powerful data-driven approach to discovering ensembles beyond the convex hull, embracing structural diversity over stability alone, enabling robust experiment matching in cases evading traditional methods. crystALL leverages 5 million relaxed structures from OPTIMADE and MPDD to rapidly generate a comprehensive candidate set by: (1) species substitutions; (2) ML feature-space analysis with pySIPFENN; (3) representative structure selection via clustering and energy-minimization or property matching; and (4) construction of the final candidate ensemble for experimental or ab initio validation. We also uniquely address the common challenge of experimental composition uncertainties spanning many integer stoichiometries by new efficient searches across anonymized compositional ranges rather than fixed formulas. Overall, crystALL's ensemble-focused philosophy provides critical advantages over the prevailing convex-hull-searching paradigm, which often fails to account for key experimental observations and increases AI/ML discovery bias over time - akin to model collapse in LLMs retrained on iteratively generated content.
Related
- Machine-learning screening and first-principles lattice dynamics resolve the most stable structure of NaCa(BH4)3 from a data-driven candidate search
- Direct Observation of Dipolar-Driven Anisotropic Quantum Projection Noise in a Solid-State Spin Ensemble
- Floquet analysis of coherence in periodically driven diamond NV ensemble systems
- A data-driven heuristic-ensemble method for behavioural inference in collective-risk dilemma experiments
- Beyond Candidate Generation: A Validation-First Protocol for AI-Assisted Transition-State Discovery
BibTeX
@misc{krajewski2026datadriven,
title = {{Data-driven candidate structure ensemble generation beyond the convex hull with crystALL approach}},
author = {Adam M. Krajewski and Ricardo Amaral and Jonathan W. Siegel and Luke A. Myers and Rushi Gong and Shun-Li Shang and Zi-Kui Liu},
howpublished = {arXiv},
year = {2026},
month = nov,
doi = {10.48550/arxiv.2611.02584},
eprint = {2611.02584},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2611.02584},
note = {CC BY-SA 4.0},
}
RIS
TY - UNPB
TI - Data-driven candidate structure ensemble generation beyond the convex hull with crystALL approach
AU - Adam M. Krajewski
AU - Ricardo Amaral
AU - Jonathan W. Siegel
AU - Luke A. Myers
AU - Rushi Gong
AU - Shun-Li Shang
AU - Zi-Kui Liu
PY - 2026
DA - 2026/11/02
PB - arXiv
DO - 10.48550/arxiv.2611.02584
UR - https://arxiv.org/abs/2611.02584
AB - Real materials consist of multiple stable or metastable states under real-world conditions, existing as statistical ergodic ensembles of configurations rather than ideal crystals. Yet, data-driven computational materials discovery increasingly biases itself toward finding "new stable structures" from ground-state DFT calculations alone. This limits our long-term ability to discover materials whose properties emerge from metastable phases, to explain experimental observations, or to model thermodynamic behavior comprehensively. Predicting ergodic sets of atomic configurations of known and hypothetical solids is a fundamental challenge in theoretical crystallography, approached via many complementary methods. We discuss this conceptually and introduce crystALL, a pragmatic and powerful data-driven approach to discovering ensembles beyond the convex hull, embracing structural diversity over stability alone, enabling robust experiment matching in cases evading traditional methods. crystALL leverages 5 million relaxed structures from OPTIMADE and MPDD to rapidly generate a comprehensive candidate set by: (1) species substitutions; (2) ML feature-space analysis with pySIPFENN; (3) representative structure selection via clustering and energy-minimization or property matching; and (4) construction of the final candidate ensemble for experimental or ab initio validation. We also uniquely address the common challenge of experimental composition uncertainties spanning many integer stoichiometries by new efficient searches across anonymized compositional ranges rather than fixed formulas. Overall, crystALL's ensemble-focused philosophy provides critical advantages over the prevailing convex-hull-searching paradigm, which often fails to account for key experimental observations and increases AI/ML discovery bias over time - akin to model collapse in LLMs retrained on iteratively generated content.
ER -