The triplet-triplet fusion process occurs when two T ₁ excitations transform into one excited singlet state. To ensure high internal conversion efficiency, it is important to suppress the formation of higher triplet states. Therefore, a criterion for a molecule to be an appropriate candidate for the design of a TTF material is usually expressed in terms of the lowest singlet (S ₁) and two lowest triplet (T ₁ and T ₂) energy levels as follows: 2 T 1 > S 1 , 2 T 1 < T 2 , (1) where the first inequality ensures that there is enough energy in two triplet states to form a singlet excitation, and the second inequality prohibits the formation of higher excited triplet states, thus favoring only singlet excitation formation. Large splitting between S ₁ and T ₁ implies that both states originate from ππ* excitations. Under this assumption, it makes sense to focus the methodology on accurate prediction of ππ* states, and tolerate lower performance for nπ* states. For example, ππ* excitations are relatively unaffected by the solvent polarity, so vacuum calculations should be sufficient and allow for faster computations.

The present work is focused on blue OLED light-emitting materials that require a certain threshold for S ₁ energy. In particular, we can define three numerical criteria to filter compounds appropriate for TTF applications as follows: S 1 > δ a , T 2 − 2 T 1 > δ b , 2 T 1 − S 1 > δ c , (2) where numerical values of the threshold parameters δ _a , δ _b , and δ _c can be adjusted to find a better tradeoff between the number and quality of final candidates. In the ideal case, we should set δ _a = 2.8 eV, δ _b = δ _c = 0 eV, but in practice we use less restrictive values to allow for intrinsic inaccuracies of simulation approaches and finite width of excitation energy levels of the molecules in the OLED emission layer.

In this section, we present our algorithm for the generation of molecular structures. It includes several consecutive steps, illustrated in Figure 2 with Roman numerals. The procedure can be subdivided into two parts. First, we generate a skeleton frame, that is, a graph of connected points, that does not yet specify the atomic types or bond orders (steps I-III in Figure 2). Second, the frame needs to be populated with heteroatoms, double bonds, and side groups that correspond to the correct Kekulé structure of a specific molecule (steps IV-VI in Figure 2).

Figure 2 presents a branch of the structure generation tree that starts from a single frame made of three 6-membered rings (step I). On step II we need to mark the places where a new ring can be attached. All possible pairs of connected atoms located on the perimeter of the frame should be considered. On step III, 5 and 6-membered rings are attached to the marked places. Duplicates already stored in the database are removed, and some simple heuristics are applied to filter out structures with steric hindrances. After processing all frames consisting of N-membered rings, we obtain the next “generation” of frames consisting of (N + 1)-membered rings.

On step IV, heteroatoms are placed in the frame according to the following rules: at most 4 heteroatoms in the frame; only nitrogen and oxygen are considered; even number of π electrons is required; no pair of heteroatoms can be connected by a covalent bond except for two nitrogens. The last rule is introduced to exclude extremely exotic compounds with peculiar distributions of heteroatoms in the molecule that a priori have a very little chance to be synthesizable and photostable.

If the resulting labeled graph can be successfully kekulized with openbabel (O’Boyle et al., 2011), it is considered to be a valid molecule, and its SMILES string is stored in the database. Again, all duplicates are removed. The resulting molecules comprise the set of “cores”, compact polycyclic fragments without side groups that can serve as building blocks to make more complex TTF materials. The chemical space of cores made of 4 or less rings consists of 472, 505 structures, and all of them can be enumerated with the above algorithm.

To make the resulting materials more likely to be applicable in real-life OLEDs, the cores should be modified further. First of all, the N-H bond in secondary amines should be capped with some residue because of low photochemical stability of the N-H bond. We replaced hydrogens with phenyls to simulate adjacent (presumably aromatic) parts of the complex TTF material (see step V in Figure 2). On step VI, the required number of side groups can be placed at the given positions, leading to the final compound.

We have collected data from the literature on well-resolved 0–0 transitions of 55 polycyclic molecules with measured absorption or luminescence from the first singlet or triplet excited states (Halverson and Hirt, 1949; Sponer and Rush, 1949; Halverson and Hirt, 1951; Hirt et al., 1954; Evans, 1957; Ito et al., 1957; Goodman and Kasha, 1958; Goodman, 1961; Shimada, 1961; Dorr and Gropper, 1963; Gropper and Dorr, 1963; Burgos et al., 1977; Schmidt, 1977; Schiedt and Weinkauf, 1997; Reineke and Baldo, 2014; Padula et al., 2019). The structures are presented in Supplementary Figure S2. The compounds can be clearly divided into three groups: pure aromatic hydrocarbons (composition HC), nitrogen-containing compounds (composition HCN) and oxygen-containing compounds (composition HCNO). We made no distinction between absorption and luminescence, since we used data on 0–0 transitions. We also used experiments performed in different media: gas phase, non-polar solvents, rigid matrix or alcohols. Water and other highly polar solvents were not present. We compare this experimental data with calculations performed in the vacuum; this is a reasonable simplification due to the weak dependence of ππ* transitions on the solvent polarity. In any case, data points obtained in different solvents follow the same trend, and the number of experiments performed in every particular medium is too low for reliable statistical analysis. Under these approximations, we were able to collect a dataset suitable for the validation of the utilized computational approach.

Meeting growing needs of computational chemistry, various benchmark molecular datasets are being continuously created nowadays (Wu et al., 2018). A typical dataset contains molecular structures and properties calculated using density functional theory (DFT). One of the most famous developments of this kind in the area of material science has been the Harvard Clean Energy Project (Hachmann et al., 2011), spanning 2.3 million candidate organic photovoltaic materials. However, most datasets do not provide spectral properties since the calculation of excited states using time-dependent DFT (TDDFT) is more time-consuming and often less reliable than the calculation of the ground state. On the other hand, datasets containing spectral properties are either not large enough (Abreha et al., 2019) or have small overlap with compounds relevant for TTF applications (Wu et al., 2018). This makes it necessary to prepare our own training dataset in order to search for candidate TTF compounds.

Since our generated structures amount to four hundred thousands compounds with more than 10 heavy atoms in average, the use of TDDFT to assess spectral properties is extremely computationally expensive. We estimate that TDDFT computations for a dataset of 0.5 million TTF molecules would require more than 100 CPU-years. Moreover, the validity of TDDFT as the correct ab initio method is questionable. One well-known issue is, for instance, the uneven treatment of excitations of different nature or spin multiplicity (Parac and Grimme, 2003). Even valence ππ* excitations of polycyclic compounds can pose substantial challenges (Grimme and Parac, 2003; Prlj et al., 2016). In order to combine computational efficiency with accurate prediction of spectral properties, we used semiempirical methods of quantum chemistry. Despite them not being ab initio approaches, many semiempirical methods, including “spectral” modifications, were initially parametrized on small aromatic and other flat conjugated organic molecules. The accuracy of semiempirical methods for the prediction of the lowest excitation energies is expected to be on par with TDDFT, while greatly speeding up calculations. One can compare different approaches and estimate their typical errors by validating computational approaches against experimental data. Reference data for a small validation dataset can also be obtained with high-level ab initio methods. For molecules of moderate size, such as the TTF cores we consider in this work, even multiconfiguration calculations can be theoretically feasible. We have attempted to apply the complete active space self-consisting field (CASSCF) method supplemented with multiconfiguration second-order perturbation theory (MCQDPT). The maximum reasonable size of active space was (12, 12), which was found to be sufficient for triplet excitation to converge in almost all cases. Unfortunately, first singlet excitations converged much more slowly. Even after some admixture of the ground state, convergence was not achieved. It appears that the CASSCF/MCQDPT approach cannot be used for blind screening without manual inspection of every particular case, so we limited the validation to experimental references.

The first step in the calculation of excitation energies is the optimization of molecular geometry. For this purpose, we have used the SE method PM3 as implemented in the Gaussian 16 software package (Frisch et al., 2016). We used the configuration interaction singles (CIS) approach to compute excitation energies. We have tested three semiempirical methods: AM1 (Dewar et al., 1985), PM3 (Stewart, 1989), and ZINDO/S (Ridley and Zerner, 1973) as implemented in the Gaussian 16 software package. For comparison, we also calculated excitation energies using DFT with the settings optimized for mixed-valence organic compounds (Renz et al., 2009): BLYP35/def2-TZVP (Weigend and Ahlrichs, 2005) geometry optimization followed by TDDFT for S ₁ and T ₂ states, or by ΔSCF for T ₁ state using M062X exchange-correlation functional (Zhao and Truhlar, 2008) and the same basis set. It important to note here that our final goal is to develop a method to predict transition energies within a series of polycyclic organic molecules, so we are not interested in the absolute accuracy of the method but rather in its high precision. Bias can be corrected with a linear transformation applied after the calculation. The performances of different methods are compared in Table 1. From the perspective of these results, we can suggest the SE method PM3 as the optimal choice for all further calculations on large molecular datasets. After proper linear correction, it outperforms other SE methods for triplets and is almost on par with ZINDO/S for singlets. PM3 is also on par with corrected DFT and significantly improves over DFT without correction. Plots of calculated versus experimental transition energies for PM3 and DFT can be found in Supplementary Figure S3.

Although exhaustive enumeration of chemical compounds is possible for certain restricted areas of the chemical space, it is always desirable to “soften” the constraints and search for promising compounds within less restricted regions. Moreover, it is often not easy to formulate clear and complete rules on the chemical diversity of all possible candidate compounds and implement the corresponding deterministic algorithms for library generation. In such cases, approaches based on machine learning and, in particular, deep neural networks can be of great help. The general idea of ML-assisted design proceeds as follows: first, we construct computationally (or extract from experiments) a database with a certain set of molecules that we assume to be relevant for the considered problem. Then, we train a generative machine learning model (usually a deep neural network) on that database, in the hope that the model will capture fundamental structural and chemical features of the dataset and will be capable of suggesting new molecules beyond the training set. If the architecture of the generative model and learning procedures are organized well, one can expect that a large fraction of generated molecules will be relevant for the problem of interest, thus greatly reducing the search space for subsequent validation. In the particular case of TTF compounds, we expect that the model will generate chemical structures featuring excitation energies applicable for the TTF process (see formula (2) and discussion below for details).

To test the applicability of generative models for computational discovery of TTF materials, we have utilized and compared several deep generative models, including character-level recurrent neural networks, adversarial autoencoders, and variational autoencoders (see Section 3.2). The best model that we recommend for practical use is the junction tree variational autoencoder (JT-VAE) architecture introduced by Jin et al. (2018). The decoder of JT-VAE consists of two parts: a graph convolutional neural network (CNN) and a junction tree convolutional neural network. The choice of the architecture was motivated by JT-VAE’s superior ability to encode and decode cyclic fragments of molecules. The latter is often challenging for conventional molecular graph CNNs but is of primary importance for TTF molecules that feature distributed π—electronic orbitals.

Apart from the task of molecular generation, neural networks can also be used to predict excitation energy levels of the molecule. We have trained the neural network to predict energies of the singlet and first two triplet states of TTF molecules. Our neural network for energy prediction is also based on junction tree convolutions; the high-level architecture of this neural network, which we call JT-E (Junction Tree Energies), is presented in Figure 3. JT-E is constructed as follows: the layer of Junction-Tree encoder preceding the latent space is connected to a network with several fully connected layers of decreasing dimensionality. The last layer of that network has three heads that correspond to S ₁, T ₁, and T ₂ excitation energies. The neural network is trained to minimize the sum of root mean square errors between predicted and calculated values of excitation energies. The JT-E model has allowed us to predict with good accuracy if a molecule might be suitable for TTF knowing only its SMILES notation (see Supplementary Table S1 for numerical results on existing benchmarks).

The structure generation algorithm shown in Figure 2 has provided us with 472, 505 non-equivalent compounds with at most 4 rings, which constitutes an exhaustive sampling of the chemical space defined by the constraints listed in Section. 2.2. Applying formula (2), we have selected 5,690 candidates from the set of 472, 505 compounds that are most promising for deep blue TTF applications. These candidates should be subjected to more detailed analysis. An important additional target here is potential synthesizability. The most robust way to ensure synthesizability is to search for already known compounds. Among 5,690 candidates, we found 107 compounds contained in the PubChem (Kim et al., 2021) database. This estimate gives hope that the exploration of considered regions of chemical space can provide a sufficient amount of TTF candidates, both core structures and their derivatives.

Among 107 PubChem hits, several distinct groups of compounds can be identified. The first group of 16 compounds consists of anthracene and its nitrogen-containing structural and isoelectronic analogues. Then, one can found 19 analogues of tetracene, 11 analogues of isobenzofurane, as well as pyrene and two of its analogues. A large and diverse group of 28 compounds contains furane moiety as part of the system of fused rings. The remaining 30 compounds are not so closely related to existing TTF materials. The major part of molecules from the set with 107 elements are analogues of molecules with registered TTF activity (Wang et al., 2020). Introduction of additional nitrogens does not change the electronic configuration, but modulates nuclear charges of chosen atomic sites, which sometimes can make S ₁ levels higher (note the cases of tetracene and isobenzofurane), so the conditions for blue TTF materials are satisfied. Review of the core compounds produced by the screening procedure supports the conclusion about the adequacy of applied methodology and underlying SE approach. This success can be partially explained by the loose criteria used in the screening: more than 1% of the original dataset has passed the filters. This is in line with the general strategy of filtering out definitely bad compounds and allowing all that have a chance to prove useful. We believe that the list of PubChem hits contains some indications useful in the search for novel cores with TTF activity. In the subsequent sections, we apply the same computational procedure to prepare training datasets for targeted design of TTF materials based on substituted compounds. The list of PubChem hits, as well as specific details of the screening procedure, can be found in the Supplementary Material (see Note S2.2 and Supplementary Table S2).

We have investigated the performance of various generative models on a dataset of cores defined in the previous subsection. Well-defined chemical composition of the subspace allows us to measure consistently if generative models are capable of suggesting molecules predominantly from the chemical subspace of interest and whether it is possible to tune generators to suggest novel molecules from the subspace with energies satisfying TTF criteria. We have excluded molecules with low and negative excitation energies from the dataset using the following criteria: S 1 > 1.0 e V , T 1 > 0.5 e V , T 2 > 1.0 e V . (3)

Negative excitation energies are nonphysical and correspond to situations where methods of quantum chemistry fail for low lying excitations. Since we focus on the discovery of TTF molecules suitable for blue OLED applications, we are not interested in those molecules because they will only introduce additional noise to the models. The total number of molecules in the truncated dataset is 341, 433.

Distributions of excitation energies in the dataset are shown in Figure 4. The figure shows that all energies in the dataset feature unimodal distributions with the following means and standard deviations: S ₁ = 1.98 ± 0.66 eV, T ₁ = 1.52 ± 0.51 eV, and T ₂ = 2.19 ± 0.47 eV. The dataset was split randomly into two parts of the same size that were used as training and validation sets (see SI for details).

We have investigated the performance of our implementation of the JT-VAE model (modified from https://github.com/wengong-jin/icml18-jtnn) and three well-known baseline models, namely:

• character-level recurrent neural network (CharRNN) (Preuer et al., 2018; Segler et al., 2018) that models the distribution of the next token in a sequence (SMILES string) with a recurrent neural network; • variational autoencoder (VAE) (Kingma and Welling, 2014; Kadurin et al., 2017; Gómez-Bombarelli et al., 2018)it consists of two networks, encoder and decoder, that learn a mapping of the input into a low-dimensional latent space by minimizing the reconstruction loss and regularization in the form of the Kullback–Leibler divergence between the approximation and the posterior distribution; • adversarial autoencoder (AAE) (Makhzani et al., 2016) that replaces the Kullback-Leibler divergence from VAE with an adversarial objective, training a discriminator network to distinguish samples from the latent space and a prior distribution that the model will sample from to generate new instances.

In all generative models, we use SMILES strings as the input and output representations. We have used the implementations of CharRNN, VAE, and AAE models available athttps://github.com/molecularsets/moses, the benchmarking platform called MOSES (Polykovskiy et al., 2020). All models were trained on the training dataset using hyperparameters and protocols as suggested by Polykovskiy et al. (2020).

We have implemented two different regimes for sampling from the latent space of autoencoder models (VAE, AAE, and JT-VAE): random and seeded. The random regime corresponds to “conventional” sampling of the latent space from the normal distribution N ( 0,1 ) that was used as prior in our models. Seeded sampling was carried out as follows. After training the encoder on the training dataset, a subset of molecules most promising for deep blue applications was selected using the criteria from formula (2): δ _a > 2.7 eV, δ _b = δ _c >—0.1 eV. Only 58 out of 171, 716 molecules in the training set satisfy these criteria; we will further refer to molecules satisfying them as leads. Latent representation vectors ν _i were calculated for each of the leads. Then, three lead vectors ν ₁, ν ₂, and ν ₃ were selected randomly and multiplied by random positive factors α ₁, α ₁ and α ₃ that satisfy the following relation: ∑ i = 1 3 α i = 1 . The sample vector ν in the latent space was constructed as ν = ∑ i = 1 3 α i v i . The molecular structure was obtained by applying the model’s decoder to the resulting latent vector ν. This approach has allowed us to sample the latent space not randomly but mostly in the vicinity of known leads. This should be beneficial if the latent space clusters favorably, separating promising TTF molecules from the rest (see discussion in Section 3.3). We have modified the implementations of VAE and AAE models by Polykovskiy et al. (2020) to run sampling in seeded mode. Note that since CharRNN is not an autoencoder model, it cannot be “seeded” with leads. Therefore, we do not present any results for CharRNN in the seeded mode.

Results for random and seeded sampling of the latent space for different models are presented in Table 2; we have obtained 10, 000 samples from each model. Table 2 clearly shows that all models were capable of suggesting valid SMILES, as checked with RDKit (Landrum, 2012).

One of the most important properties of a generative model is their ability to sample from novel yet not arbitrary regions of chemical space. Since in the present work we are interested in regions of chemical space consisting solely of π-conjugated systems with 5 or 6-membered rings, we measure the extrapolating ability of generative models by the number of suggested molecules from the validation set. Note that all generated molecules should not necessarily satisfy the constraints applied during dataset creation, since no formal restrictions on molecular composition were implemented for the generators. Some generated molecules do not belong to the above-mentioned chemical space of π-conjugated molecules. Therefore, the number of unique molecules is larger than the sum of molecules from the training and validation sets. Table 2 shows that the AAE model was able to suggest the largest fraction of novel molecules (1992 and 2,155 for random and seeded implementations respectively). Almost 15% of the molecules were from the validation set, i.e., novel. Note also that the size and chemical composition of the training and validation sets are identical, so an unbiased generator should suggest a similar number of molecules from both sets. Autoencoder models indeed demonstrate nearly equal number of generated molecules from both datasets. On the contrary, the CharRNN model is extremely biased towards the training set, which could be an indication of overfitting; here we do not investigate that question in depth since we used default suggested values of training parameters from MOSES (Polykovskiy et al., 2020).

The most interesting and important part for the problem of the discovery of novel TTF molecules is the number of novel discovered leads. Recall that a lead is a molecule with excitation energies suitable for blue OLED applications. We see that all models performed poorly in the conventional random sampling mode: CharRNN suggested one lead from the validations set, while all other models suggested none. The situation is very different for seeded generation. The AAE model was capable of suggesting four leads from the validation set that have not been seen by the model during training. And this is exactly where the JT-VAE model shines: it was able to generate 11 TTF candidates from the validation set, much higher than any other model in the comparison.

Note also that the number of unique molecules generated by AAE and VAE is nearly identical to the total number of valid SMILES both in random and seeded implementations. For JT-VAE, this holds under random sampling, but not in the seeded mode, where nearly 65% of generated molecules turned out to be duplicates. At first glance it might seem to be a drawback, but in fact this property means that fewer molecules need to be checked for excitation energies favorable for TTF, and it shows that the sampling space of JT-VAE in seeded mode is much more concentrated. Ultimately, we are interested in the number of suggested leads, not just the number of unique molecules, and indeed, despite lower number of unique SMILES, the JT-VAE model suggested by far the most leads from the training and validation sets. This advantage is especially striking if we consider the fraction of generated molecules that need to be checked to find a new lead (shown as percentages in Table 2): the probability to find a lead with JT-VAE is ≈ 2.4 % for each new suggested SMILES string, which represents a more than 15 times higher rate than for the VAE and AAE models and 75 times higher than picking molecules from the validation set at random.

Note that for all autoencoder models we have observed a presumably linear dependence between the number of unique molecules and the number of molecules from the training and validation sets, as indicated by comparing the random and seeded generators. This is indicative of the fact that seeded generation does not alter the fraction of generated molecules belonging to the desired region of the chemical space, in our case π-conjugated systems with 4 rings. This observation allows to suggest that application of seeded sampling does not disturb the predefined constraints on molecular composition, but allows to further accelerate lead discovery.

Based on the above analysis, we suggest the JT-VAE model with seeded sampling as the best generative model for the discovery of realistic candidates for deep blue OLED applications.

To better understand why and how the JT-VAE model generates an increased number of leads with seeded sampling, we have investigated the latent space of the model using the t-distributed Stochastic Neighbor Embedding (t-SNE) to generate a two-dimensional visual representation (van der Maaten and Hinton, 2008). The results obtained for 25, 000 molecules randomly sampled from the dataset are shown in Figure 5. Color corresponds to the “fitness” of a molecule for TTF applications: red indicates a better fit, blue, a worse fit. This means that leads are shown in red.

Figure 5 shows that the distribution of red, blue, and intermediate points is far from uniform in the latent space: leads and generally molecules with higher fitness tend to cluster together. Therefore, if we choose a random linear combination of latent space vectors for three red points (leads), it will have an increased chance to end up in the close vicinity of another red point (especially when all leads are sampled from the same cluster). This demonstrates that the JT-VAE model in seeded sampling mode allows to discover most of the leads in the chemical space with fewer iterations than other approaches. We note that there are in total only 56 leads in the validation area of the chemical space comprising 170, 716 molecules, and JT-VAE was capable of discovering 11, i.e,. ≈ 20 % of those leads in just the first 3,500 unique samples (out of 10, 000 total first samples).

As discussed above, we have analyzed the models’ performance in a relatively small region of chemical space with up to 4 rings without side chains, a region with less than 0.5 million molecules in total. This has allowed us to assess the potential of the models to extrapolate beyond the training set and discover leads in the entire constrained chemical space. In what follows, we apply our conclusions to the discovery of TTF candidates in much larger chemical spaces that cannot be sampled exhaustively.

Table 2 shows that in our restricted subset of the chemical space, the seeded JT-VAE model generates leads at a rate of (11 + 16)/(559, +, 558) ≈ 2.4%. Though one could apply quantum chemistry methods to all generated molecules to discover the leads, there is a more computationally efficient alternative. We have trained the JT-E network as discussed in Section 2.6 to predict excitation energies for molecules supplied as SMILES strings from the generator. Calculations of excitation energies for the training dataset were done using PM3 (Stewart, 1989), the same SE method as we have used above. Mean absolute errors for the excitation energies for validation set are 0.104, 0.054, and 0.086 eV for the S ₁, T ₁, and T ₂ energies respectively. Note that this is a remarkable accuracy, comparable to the accuracy of the PM3 method itself. The JT-E model is trained independently on the same dataset as the JT-VAE model (see Supplementary Material for details). Based on predicted energies, we have filtered generated molecules according to the same criteria from (Eq. 2): δ _a > 2.3 eV, δ _b = δ _c >—0.4 eV. Those are looser criteria than for lead selection since we wanted to give a very safe margin of error for the JT-E model (exceeding 2σ). Geometries of the molecules satisfying these criteria are then optimized, and excitation energies are computed using PM3. In the next section, we show and discuss the overall workflow for TTF molecules discovery.

In order to promote the discovery of real TTF materials, we assembled a multi-step workflow acting in the space of π—conjugated compounds. The training datasets consisted of previously used core structures decorated with side groups. We used two types of side groups: tert-butyl and mesityl moieties, as a model of different bulky, but chemically inert substituents that are often used to prevent flat cores from aggregation. We also replaced all amine hydrogens with phenyls, as required to ensure operational stability of the material. The size of the chemical space for compounds with side groups is much larger than for cores only. Therefore the datasets were not exhaustive and included ≈ 450,000 molecules with positions and types of the side groups selected at random.

We have investigated three different chemical spaces corresponding to cores: decorated with none, one, and two side chains. We had tried to train a single network on the entire dataset, but our experiments showed that training separate neural networks for each number of side chains allows to increase reconstruction accuracy and accuracy of energies prediction, as well as to achieve better clustering of leads of each type in the chemical space, so we have chosen this strategy. For all molecules from generated chemical spaces, we optimized their geometries and calculated S ₁, T ₁, and T ₂ excitation energies using the SE PM3 approach. We have excluded molecules with unreasonably low energies from each dataset, using criteria outlined in (Eq. 3). The datasets were used to train three JT-VAE generative models and three JT-E energy predicting models (see Supplementary Material for additional data on the architecture and accuracy of JT-VAE and JT-E models). We have selected molecules most suitable for TTF leads from the datasets and utilized them in seeded sampling of the latent space of the autoencoders, using procedures discussed in Section 3.2. The overall discovery workflow for TTF materials is shown in Figure 6.

We have found 75, 58, and 55 leads directly in the dataset with 0, 1, and 2 side chains respectively. The leads were used to seed the generator of JT-VAEs. During each discovery loop, we generated another 100, 000 samples and predicted excitation energies for them using the JT-E network. The molecules satisfying criteria discussed in Section 3.4 were selected for quantum chemical calculation with PM3. Results obtained for each iteration of the workflow cycle are summarized in Table 3.

Table 3 shows that all three datasets with nearly 1.5 million molecules initially contained only 188 leads. However, each iteration of the discovery workflow brings ≈ 400 more new leads. After just two iterations, we have obtained more than a thousand compounds that appear promising for deep blue TTF applications. SMILES notation for those molecules along with calculated PM3 energies are provided in Supplementary Material. Although one can easily continue the discovery cycles, we stopped at the current stage since more than 1,000 leads is already a substantial amount that is not easy to verify experimentally.