Impact of Dataset Diversity on Machine Learning Prediction of Reorganisation Energies in Organic Semiconductors - Slide presentation

Aug 3, 2026

This presentation investigates how dataset input characteristics, especially chemical diversity, influence machine learning performance for predicting hole and electron reorganisation energies in organic semiconductors. It explains the role and molecular computability of reorganisation energy using DFT and motivates machine learning as a cost-effective alternative to exhaustive quantum-chemical evaluation. Two databases with markedly different diversity are built and used to compare model and feature choices, showing that greater dataset diversity can reduce prediction accuracy while generalisable prediction remains difficult.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages1

Reading Time3 min

AccessFree to download

Research aim

Charge Transport and Organic Semiconductors

Why reorganisation energy matters

Methodology

Datasets

Modeling and feature selection

Results

Electron predictions vs hole predictions

Conclusion

Impact of Dataset Diversity on Machine Learning Prediction of Reorganisation energies in Organic Semiconductors Malin Zollner, Tahereh Nematiaram and Yashar Moshfeghi University of Strathclyde This work investigates how input characteristics, particularly dataset diversity, affect the performance of machine learning algorithms in predicting hole and electron reorganisation energies, ultimately aiding the identification of promising organic semiconductors candidates. Research aim Charge Transport Organic Semiconductors ❑ To improve efficiency in organic semiconductors, higher charge ❑ Organic semiconductors offer a sustainable alternative to inorganic mobility is essential, which depends on: [8,9] semiconductors across a range of applications. [1] [2] [3] [4] ❑ Compared to inorganic semiconductors, organic semiconductors are easier and cheaper to fabricate [5], flexible [5], easy to fine tune [6] and the larger band gap (2-4 eV vs 1-2 eV) [7] enables higher energy light absorption and emission and reduces undesirable side reactions. ❑ Of these, reorganisation energy (􀣅) is the only one computable and controllable at the molecular level [10] using DFT and: 􀣅 = 􀡱􀢔􀢉 − 􀡱􀢔􀢔 + 􀡱􀢉􀢔 − 􀡱 􀢉􀢉 / 􀫛 ❑􀡱􀢔􀢉 = Neutral molecule’s energy in charged geometry, 􀡱􀢔􀢔 = Neutral molecule’s energy in its optimised neutral geometry , 􀡱􀢉􀢔 = Charged molecule’s energy in neutral geometry , 􀡱 􀢉􀢉 = Charged molecule’s energy in its optimised charged geometry Why is it important? ❑ To identify promising new organic semiconductors, molecules with low reorganisation energy must be found. ❑ However, computing reorganisation energy across the entire search space is computationally unfeasible. Machine learning provides a solution by enabling rapid prediction of reorganisation energy at a fraction of the computational cost. Dataset Methodology Database 1 : 1500 thiophene derivatives Database 2: 4500 diverse small organic molecules Research Top performing electron 􀟣 predictions Top performing hole 􀟣 predictions Results Conclusion ❑ Two molecular databases of differing chemical diversities were constructed. ❑ Hole and electron 􀟣 was computed using DFT and predicted using machine learning. ❑ Careful feature and model selection is crucial. The optimal model and feature set was identified (GBRT + Descriptors + All_fp) . ❑ Electron 􀟣 can be predicted with higher accuracy than hole 􀟣 . ❑ Increased chemical diversity in the dataset significantly reduces prediction accuracy. ❑ Accurate and generalisable prediction of 􀟣 remains challenging. References [1] [https://www.scmp.com/business/companies/](https://www.scmp.com/business/companies/)[ ](https://www.scmp.com/business/companies/)[article/2113079/ultra-thin-flexible-display-maker-royole-](article/2113079/ultra-thin-flexible-display-maker-royole-)[raises-another-us800m](raises-another-us800m) [2] [https://sistinesolar.com/solar-panel-standards-and](https://sistinesolar.com/solar-panel-standards-and)certification/ [3] [https://weartechdesign.com/smart-health-sensor](https://weartechdesign.com/smart-health-sensor)provides-premium-cardiovascular-check/ [4] [https://www.greenlanemarketing.com/resources/ articles](https://www.greenlanemarketing.com/resources/ articles) [5] J. Zaumseil and H. Sirringhaus, Chem. Rev., 2007, 107 , 1296-1323 [6] S. Huang, B. Feng, X. Cheng, X. Huang, J. Ding, K. Yu, J. Dong and W. Zeng, Chemical Engineering Journal, 2023, 476 , 146436 [7] O. Ostroverkhova, Chem. Rev, 2016, 116 , 13279-13412 [8] S. Hutsch, M. Panhans and F. Ortmann, npj Computational Materials, 2022, 8, 228 [9] S. Fratini, M. Nikolka, A. Salleo, G. Schweicher and H. Sirringhaus, Nature materials, 2020, 19, 491–502 [10] S. F. Nelsen, S. C. Blackstock and Y. Kim, Journal of the American Chemical Society, 1987, 109, 677–682 Acknowledgements Results were obtained using the ARCHIE-WeSt HighPerformance Computer ([www.archie-west.ac.uk](www.archie-west.ac

FAQs

Why is dataset diversity important for predicting reorganisation energies with machine learning?

The work shows that chemical diversity in the training set can significantly affect model accuracy, with increased diversity reducing prediction performance.

How are hole and electron reorganisation energies obtained for training and evaluation?

Reorganisation energies are computed using DFT, then predicted using machine learning to compare accuracy and generalisation.

What is the key takeaway about electron vs hole prediction accuracy?

Electron reorganisation energy is predicted with higher accuracy than hole reorganisation energy, though achieving fully accurate and generalisable hole prediction remains challenging.

Aug 3, 2026

Upload by Valentina