Rational design of organic molecules with inverted gaps between the first excited singlet and triplet

Molecular Geometry Impact on Deep Learning Predictions of Inverted Singlet–Triplet Gaps

“…and V), for a total of ≈1650 entries with ≈20% of IST candidates (i.e., molecules showing Δ E ST ref < 0) . It is worth noticing that in the reference data provided with DS1, azulene itself occurs twice in the data set, and in both instances, it is associated with a negative gap (Δ E ST ref ≈ −3 meV and Δ E ST ref ≈ −66 meV), in contrast to experiments: azulene is known to have Δ E ST > 0. ,, In our case, only one candidate presents a nitrogen-substituted azulene scaffold (showing Δ E ST > 0 for both vertical and adiabatic gaps computed at ADC(2)/cc-pVDZ level on M06-2X/def2-TZVP geometries); 6 out of 15 azulenes are predicted with small negative gaps (Δ E ST ML ≥ −10 meV), while 6 more show −10 meV > Δ E ST ML ≥ −40 meV. On the basis of these considerations, we believe that azulene scaffolds reported in Figure are indeed FPs, although it has been recently shown that different patterns of fusion between azulene scaffolds can yield both positive and negative Δ E ST gaps. , Suboptimal quality of reference data on (at least some) azulenes might be one of the possible reasons for this result: the model has learnt that several molecules containing the azulene core should have Δ E ST < 0, which points toward an obvious way to improve performance, i.e., obtaining training data that are more accurate from the quantitative point of view.…”

“…and V), for a total of ≈1650 entries with ≈20% of IST candidates (i.e., molecules showing ΔE ST ref < 0). 17 It is worth noticing that in the reference data provided with DS1, azulene itself occurs twice in the data set, and in both instances, it is associated with a negative gap (ΔE ST ref ≈ −3 meV and ΔE ST ref ≈ −66 meV), in contrast to experiments: azulene is known to have ΔE ST > 0. 17,69,70 In our case, only one candidate presents a nitrogen-substituted azulene scaffold (showing ΔE ST > 0 for both vertical and adiabatic gaps computed at ADC(2)/cc-pVDZ level on M06-2X/def2-TZVP geometries); 6 out of 15 azulenes are predicted with small negative gaps (ΔE ST ML ≥ −10 meV), while 6 more show −10 meV > ΔE ST ML ≥ −40 meV.…”

“…The first data set we utilized, denoted as DS1, was introduced by Pollice et al and recently integrated by Jorner et al with other literature data, in recent preprints, and is publicly accessible on GitHub (https://github.com/kjelljorner/pppinvest). 13,17,19,21 This heterogeneous data set comprises…”

“…Additionally, it includes excited state properties, such as S 1 and T 1 vertical excitation energies as well as S 1 oscillator strengths, computed using methods capable of predicting singlet–triplet inversion, such as ADC(2) and CC2. Furthermore, DS1 encompasses 18 diverse families of compounds, all previously detailed in a series of publications by the preprint’s authors, where comprehensive analyses of this data set were presented. ,, We note that, of the data provided at the GitHub repository, we did not use those obtained from ref , because we were unable to identify a 1-to-1 correspondence between geometries and electronic data. We highlight that, although structural variability within DS1 is somewhat modest due to its genesis, electronic properties across each single family vary significantly, as can be deduced by the large interval of spanned Δ E ST gaps in each family (see ref ).…”

“…Numerous research groups are actively engaged in unraveling the intricacies governing singlet–triplet inversion. Their efforts include formulating design principles, ,, expanding the inventory of molecules exhibiting this exceedingly rare feature, − and developing sophisticated computational methods and strategies to tackle this problem. ,, Since the resurgence of interest in this field, it has become evident that electronic correlation plays a pivotal role, necessitating computational scientists to employ resource-intensive wave function quantum chemical methods to obtain reliable results. Notwithstanding this substantial challenge, Pollice, Aizawa, and Corminboeuf conducted extensive large-scale computational studies aimed at identifying molecules with emission features suitable for application in devices and discovering novel molecules to broaden case studies, thereby extracting valuable design principles despite the inherent computational cost. − …”