Directional design of chemical reaction, a prerequisite for efficient synthetic planning, has been challenged by the slow trial and error process in the laboratory and the high computational expense of ab initio calculations. It is desirable to develop an artificial intelligence algorithm to predict chemical reactions. Here, we propose an Unsupervised Assisted Directional Design of Chemical Reactions (UADDCR) software to determine whether a chemical reaction can proceed smoothly under given chemical equations, surface compositions, and facets, based on the unsupervised assisted neural network. A database with five catalytic products is trained and tested to predict more than 100,000 chemical reactions, with a lowest predicted mean absolute error (MAE) of 0.18 eV. The case studies show that the UADDCR software can facilitate chemical reactions by adjusting the inputs, a process which is millions of times faster than ab initio calculations and can accurately predict and design more than 100,000 chemical reactions.

化学反应的定向设计是进行有效合成计划的先决条件，但由于实验室中缓慢的试验和错误过程以及从头算的高计算量而受到挑战计算。期望开发一种人工智能算法来预测化学反应。在这里，我们提出了一种基于无监督辅助神经网络的无监督化学反应辅助定向设计（UADDCR）软件，以确定在给定的化学方程式，表面成分和构面下化学反应是否可以顺利进行。训练并测试了包含五种催化产物的数据库，以预测100,000多个化学反应，且最低预测平均绝对误差（MAE）为0.18 eV。案例研究表明，UADDCR软件可以通过调整输入值来促进化学反应，该过程比从头算要快几百万倍，并且可以准确地预测和设计100,000多个化学反应。