Efficient Adversarial Generation of Thermally Activated Delayed Fluorescence Molecules

Abstract

Adversarial generative models are becoming an essential tool in molecular design and discovery due to their efficiency in exploring the desired chemical space with the assistance of deep learning. In this article, we introduce an integrated framework by combining the modules of algorithmic synthesis, deep prediction, adversarial generation, and fine screening for the purpose of effective design of the thermally activated delayed fluorescence (TADF) molecules that can be used in the organic light-emitting diode devices. The retrosynthetic rules are employed to algorithmically synthesize the D–A complex based on the empirically defined donor and acceptor moieties, which is followed by the high-throughput labeling and prediction with the deep neural network. The new D–A molecules are subsequently generated via the adversarial autoencoder, with the excited-state property distributions perfectly matching those of the original samples. Fine screening of the generated molecules, including the spin–orbital coupling calculation and the excited-state optimization, is eventually implemented to select the qualified TADF candidates within the novel chemical space. Further investigation shows that the created structures fully mimic the original D–A samples by maintaining a significant charge transfer characteristic, a minimal adiabatic singlet–triplet gap, and a moderate spin–orbital coupling that are desirable for the delayed fluorescence.

De novo creation of fluorescent molecules via adversarial generative modeling

“…The generator and discriminator losses start to oscillate after the pre-set threshold epoch and converge to a level of 5 and 0.2 respectively. The finding is slightly different from the convergence profile in our previous adversarial training 24 where the classification loss can approach the theoretical value of BCE for indistinguishable samples. The phenomenon can be attributed to the relativistic approach applied on the discriminator which can possibly lead to a state-of-the-art during the convergence.…”

“…As shown in Table 1, given the small training set (which is 2923 in this case), generation sizes of 1 K to 20 K are considered. The uniqueness is found to be comparable with the profiles in Polykovskiy et al 37 (approaching 1.0 for LatentGAN at 1 K generation size) and Tan et al 24 (0.99 for AAE at 1 K generation size). For the novelty ratio, the current generation result is even better than that in Tan et al 24 (where around 0.65 novelty is reported), possibly owing to the addition of the generator network in AGAN which can potentially enhance the probability of creating new molecules not appearing in the original dataset.…”

“…The performance of generations is measured with the molecular validity, novelty, uniqueness and overall efficiency, as defined in our previous work. 24 To briefly summarize, validity represents the fraction of molecules within the generated samples that can pass the RDKit's sanitization check so that the atomic valency and the consistency of bonds in aromatic rings can be maintained. Novelty is the fraction of the generated molecules that are not present in the training set.…”

“…One is for the insufficient data within the histogram which can lead to the irregularity of distribution curves. In addition, the diversity in the training set (where samples are known to be experimentally synthesized and collected from different literature), can be another cause for the distribution mismatch, while in Tan et al 24 both original and generated sets are of the same molecular type. More details regarding the diversity can be seen in the following analyses.…”

“…Adversarial generation, known as a powerful generative model in machine learning, has attracted great attentions in physical and chemical fields for the innovative design of drug molecules and optoelectronic materials. 24 The family of adversarial models basically includes adversarial autoencoder (AAE), generative adversarial network (GAN), reinforcement learning augmented GAN, etc. , where the model framework generally consists of a neural network generator and discriminator for the sample generation and discrimination respectively.…”