Learning the limits - how data, diversity, and representation control machine-learning predictions of reorganisation energy

Aug 5, 2026

Accurate and scalable prediction of hole and electron reorganisation energies remains a bottleneck for data-driven design of organic semiconductors because routine ab initio calculations are impractical for large molecular libraries. The study systematically and interpretably evaluates how molecular representation, chemical diversity, and dataset size limit both accuracy and transferability of machine-learning models for predicting lh and le. Fifteen descriptor schemes and twelve learning algorithms are benchmarked across multiple curated datasets, with SHAP-based interpretation linking physical design trends to performance limits.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages13

Reading Time33 min

AccessFree to download

Introduction

Charge transport and the role of reorganisation energy

Reorganisation energy evaluation methods

Learning limits for prediction of reorganisation energies

Open Access Article . Published on 11 February 2026. Downloaded on 2/ 19/2026 10:01:02 AM . hicle is licensed under a Creative C mmons A 3 0 U d[ttr .](ttr .)ibution-NonCommercial npor e nce. Journal of Materials Chemistry C | | | | | | --- | --- | --- | --- | | P | APER | View Article Online | | | View Journal | | | | Cite this: DOI: 10. 1039/d5tc04408a Received 16th December 2025, Accepted 10th February 2026 DOI: 10.1039/d5tc04408a[rsc.li/materials-c](rsc.li/materials-c) Learning the limits: how data, diversity, and representation control machine-learning predictions of reorganisation energy Malin Zollner,a Yashar Moshfeghi b and Tahereh Nematiaram *a Accurate and scalable prediction of hole and electron reorganisation energies (lh and le) is a persistent bottleneck in the data-driven design of organic semiconductors, as routine ab initio calculations remain impractical for large molecular libraries. This work presents a systematic and interpretable evaluation of how molecular representation, chemical diversity, and dataset size constrain the accuracy and transferability of machine-learning models for predicting lh and le . Three complementary datasets are analysed: a chemically diverse benchmark of approximately 5000 molecules with paired lh and le values, a thiophene-focused dataset comprising 1486 molecules, and a sequence of progressively augmented datasets extending to nearly 13000 structures. Fifteen molecular descriptor schemes and twelve learning algorithms, spanning linear, kernel-based, ensemble, and graph-based models, are benchmarked under consistent training and validation protocols. Across broad chemical space, predictive performance is primarily governed by molecular representation, with hybrid descriptors that combine RDKit features and multiple molecular fingerprints consistently outperforming single-source encodings, while graph neural networks underperform in highly diverse regimes. Constraining chemical diversity leads to substantial accuracy gains, particularly for electron reorganisation energies, whereas increasing dataset size improves robustness and generalisation with rapidly diminishing returns beyond modest augmentation. Model interpretation using SHAP analysis reveals stable and physically meaningful design trends across all datasets, showing that rigid, extended p-conjugation, low conformational flexibility, and balanced charge distribution systematically reduce reorganisation energies. These results define realistic performance limits for machine-learning prediction of reorganisation energy and provide concrete guidance on representation choice, dataset design, and molecular optimisation strategies for high-mobility organic electronic materials. 1 Introduction Molecular semiconductors have attracted considerable interest owing to their advantages over conventional inorganic counterparts, including mechanical flexibility, low cost, chemical tunability, biocompatibility, and sustainability.1–3 These attributes enable their integration into a wide range of emerging technologies, such as organic photovoltaics,4,5 neuromorphic devices,6,7 and lightemitting diodes.8,9 The performance of these devices is strongly governed by charge-transport properties; therefore, the discovery and rational design of high-mobility materials remain central to the advancement of organic electronics.10–14 a Department of Pure and Applied Chemistry, University of Strathclyde, 295 Cathedral Street, Glasgow G1 1XL, UK. E-mail: [tahereh.nematiaram@strath.ac.uk](tahereh.nematiaram@strath.ac.uk) b Department of Computer and Information Sciences, University of Strathclyde, 26 Richmond Street, Glasgow G1 1XH, UK Charge transport in molecular semiconductors is inherently complex, and a variety of theoretical frameworks have been developed to elucidate its underlying mechanisms, as reviewed in several recent articles.15–19 A notable example is a large-scale screening study of the Cambridge S

FAQs

Why are predictions of hole and electron reorganisation energies challenging for data-driven organic semiconductor design?

Routine ab initio calculations are impractical for large molecular libraries, making accurate and scalable prediction of reorganisation energies a persistent bottleneck.

Which factors were tested to determine their impact on machine-learning model accuracy and transferability?

The work evaluates how molecular representation, chemical diversity, and dataset size constrain the accuracy and transferability of models predicting lh and le.

What insights does the interpretation analysis provide about physical design trends?

SHAP analysis reveals stable, physically meaningful trends: rigid extended p-conjugation, low conformational flexibility, and balanced charge distribution systematically reduce reorganisation energies.

Aug 5, 2026

Upload by Valentina