EZGA transition kernels bridge structure discovery and thermodynamic sampling in one engine
- 1Fritz Haber Institute of the Max Planck Society, Faradayweg 4–6, Berlin, Germany, 14195
Abstract
We introduce a transition-kernel layer for EZGA, a modular evolutionary framework for atomistic structure exploration. The layer extends EZGA’s population-based optimization machinery into a statistically controlled sampling engine. Moves generate trial structures, while the selected kernel accepts, rejects, or reweights them. Different statistical methods are obtained by combining the required transition kernel and moves. The engine, archive, move pool, data model, and energy backend remain common across compatible methods. For methods with a stationary target distribution, the central abstraction is a move-level detailedbalance contract. Each move reports the logarithms of the probabilities or probability densities for proposing the forward and reverse transitions. The kernel combines their difference with the ensemble-specific target weight. This treatment supports asymmetric and particle-numberchanging moves without embedding ensemble corrections in individual moves. We implement and validate fifteen statistical methods spanning equilibrium sampling, density-of-states and partitionfunction reconstruction, generalized ensembles, and free-energy calculations. Each implementation is compared with reference solutions obtained by quadrature for a non-interacting gas and/or by exact enumeration for an interacting lattice gas. By placing structure exploration and thermodynamic sampling in the same engine, EZGA can follow candidate structures from discovery to statistical characterization. The same infrastructure can quantify their equilibrium populations, reconstruct free-energy landscapes, and compare competing states. Structure discovery thus becomes the starting point for thermodynamic characterization rather than the endpoint of a calculation.
Related
- Reactive Flux Matching: Mechanism Discovery and Adaptive Sampling of Rare Events
- Building the Data Foundations for the Design and Discovery of Transition Metal Complexes
- CatWalk: Reaction Pathway Generation and Transition State Sampling via Score-based Diffusion Models
- Thermodynamic stability and electronic structure of thorium boron carbides
- Chemistry-Informed Multimodal Model for Structure Elucidation in Automated Reaction Discovery
BibTeX
@misc{lombardi2026ezga,
title = {{EZGA transition kernels bridge structure discovery and thermodynamic sampling in one engine}},
author = {Juan Manuel Lombardi and Gianmarco Ducci and Chiara Panosetti and Hendrik H. Heenen and Julian Holland and Konstantin Jakob and Felix Riccius and Karsten Reuter and Christoph Scheurer},
howpublished = {ChemRxiv},
year = {2026},
month = aug,
doi = {10.26434/chemrxiv.15008068/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008068/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - EZGA transition kernels bridge structure discovery and thermodynamic sampling in one engine
AU - Juan Manuel Lombardi
AU - Gianmarco Ducci
AU - Chiara Panosetti
AU - Hendrik H. Heenen
AU - Julian Holland
AU - Konstantin Jakob
AU - Felix Riccius
AU - Karsten Reuter
AU - Christoph Scheurer
PY - 2026
DA - 2026/08/31
PB - ChemRxiv
DO - 10.26434/chemrxiv.15008068/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008068/v1
AB - We introduce a transition-kernel layer for EZGA, a modular evolutionary framework for atomistic structure exploration. The layer extends EZGA’s population-based optimization machinery into a statistically controlled sampling engine. Moves generate trial structures, while the selected kernel accepts, rejects, or reweights them. Different statistical methods are obtained by combining the required transition kernel and moves. The engine, archive, move pool, data model, and energy backend remain common across compatible methods. For methods with a stationary target distribution, the central abstraction is a move-level detailedbalance contract. Each move reports the logarithms of the probabilities or probability densities for proposing the forward and reverse transitions. The kernel combines their difference with the ensemble-specific target weight. This treatment supports asymmetric and particle-numberchanging moves without embedding ensemble corrections in individual moves. We implement and validate fifteen statistical methods spanning equilibrium sampling, density-of-states and partitionfunction reconstruction, generalized ensembles, and free-energy calculations. Each implementation is compared with reference solutions obtained by quadrature for a non-interacting gas and/or by exact enumeration for an interacting lattice gas. By placing structure exploration and thermodynamic sampling in the same engine, EZGA can follow candidate structures from discovery to statistical characterization. The same infrastructure can quantify their equilibrium populations, reconstruct free-energy landscapes, and compare competing states. Structure discovery thus becomes the starting point for thermodynamic characterization rather than the endpoint of a calculation.
ER -