Concat-tree genetic programming for graph neural architecture search: Efficient materials property classification toward high-throughput screening
Computer Methods in Materials Science · 2026 · 26 (4)
DOI: 10.7494/cmms.2026.4.1050 ↗Kurz gefasst
Fachartikel in der Zeitschrift Computer Methods in Materials Science, 2026. Verfasst von Maciej Krzywda* und 3 weiteren Autor:innen. Die Arbeit nennt 34 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
We propose Concat-Tree Genetic Programming (CTGP) to automatically discover graph neural network (GNN) architectures for binary classification of DFT-derived materials properties in the JARVIS-DFT database. CTGP is inspired by Cartesian genetic programming (CGP) and classical tree-based genetic programming (GP): genotypes are binary concat trees that decode into explicit, heterogeneous sequential GNN stacks, evolved by mutation-only elitist search. Seven tasks span electronic structure (metal vs. insulator), magnetism, photovoltaic efficiency proxies, topological band character, and thermoelectric transport. Using fixed JARVIS train/validation/test splits and published binarization thresholds, we compare four search configurations – (1 + 1) and (1 + 4) regimes, each with and without a learned surrogate filter – with MLP, RF, DT, and SVM baselines and fair GNN references across four representation levels (L0–L3).L0 uses atomic number only, L1 adds learned elemental descriptors, L2 adds bond-distance summaries, and L3 adds bond-angle statistics. Each configuration uses ten independent random seeds per task. The search yields compact architectures with tens to hundreds of thousands of parameters and millisecond-scale model-forward inference. Accuracy is reported with model complexity, search cost, test AUC, macro-F1, and property-aware analysis of discovered operator sequences. Fixed GNN baselines and random search often match or exceed CTGP on test accuracy, while CTGP remains competitive on selected electronic-structure tasks and provides architecturally transparent stacks for exploratory, property-related interpretation, without claiming causal explainability. On benchmarks with a stable signal above majority-class performance, these surrogates may support rapid screening of large preprocessed graph libraries at low inference cost.
Zitieren
APA
Krzywda*, M., Wermiński, M., Łukasik, S., & Gandomi, A. H. (2026). Concat-tree genetic programming for graph neural architecture search: Efficient materials property classification toward high-throughput screening. Computer Methods in Materials Science, 26(4). https://doi.org/10.7494/cmms.2026.4.1050
Vancouver
Krzywda* M, Wermiński M, Łukasik S, Gandomi AH. Concat-tree genetic programming for graph neural architecture search: Efficient materials property classification toward high-throughput screening. Computer Methods in Materials Science. 2026;26(4). doi:10.7494/cmms.2026.4.1050
BibTeX
@article{Krzywda2026,
author = {Krzywda*, Maciej and Wermiński, Mariusz and Łukasik, Szymon and Gandomi, Amir H.},
title = {Concat-tree genetic programming for graph neural architecture search: Efficient materials property classification toward high-throughput screening},
journal = {Computer Methods in Materials Science},
year = {2026},
volume = {26},
number = {4},
doi = {10.7494/cmms.2026.4.1050}
}
EUVIMED