Beyond the static picture: a machine learning and molecular dynamics insight on singlet fission

Abstract

Singlet fission (SF) is a promising mechanism to overcome the current efficiency limit in solar cells. Theoretical studies have focused extensively on static pairs of molecules, the minimum system where SF can occur. Our work presents a complementary two-step approach. First, we developed a neural network model to investigate correlations between selected descriptors and the SF driving force across a library of organic molecules. SHAP analysis suggests that ionization potential (IP) and the second-to-lowest triplet (T₂) are the most influential features. Notably, SF-active and SF-inactive molecules exhibit distinct energy ranges: 2.0–3.0 eV vs. 3.7–4.5 eV for T₂ and 5.0–6.5 eV vs. 8.0–9.5 eV for IP. Second, we performed a molecular dynamics simulation on the α-polymorph of 1,3-diphenylbenzofurane, which is SF-active. We followed the evolution of the electronic states and calculated electronic couplings within a diabatic framework. Values of electronic couplings suggest a charged-transfer mediated mechanism, with the largest electronic couplings (20 meV) observed in inter-stack pairs, and intra-stack pairs exhibiting lower values. This work attempts to illustrate how machine learning can uncover relationships that may be relevant in the design of SF materials, and highlight the role of structural changes in modulating electronic couplings.