The development of the generative adversarial supporting vector machine for molecular property generation

The generative adversarial network (GAN) is a milestone technique in artificial intelligence, and it is widely used in image generation. However, it has a large hyper-parameter space, which makes it difficult for training. In this work, we propose a new generative model by introducing the supporting vector machine into the GAN architecture. Such modification reduces the hyper-parameter space by half, thus making the training more accessible. The formic acid dimer (FAD) system is studied to examine the generation capacity of the proposed model. The molecular structures, molecular energies and molecular dipole moments are combined as the feature vector to train the model. It is found that the proposed model can generate new feature vectors from scratch, and the generated data agrees well with the ab initio values. In addition, each generated feature vector is unique, so the mode collapse problem is avoided, which is often encountered in the GAN model. The proposed model is extensible to incorporate any molecular properties as the feature vector is established as the direct sum of corresponding component vectors; thus, it is expected that the proposed method will have a wide range of application scenarios. Scientific contribution statement: A generative adversarial algorithm combing supporting vector machine is proposed for the first time to predict molecular properties from scratch, which agrees well with ab initio values. The new model is more efficient than generative adversarial networks, and it is convenient to extend for application in different scenarios.