4-18. AIMS – A database system of physical properties of molecular liquids built by combining artificial intelligence and molecular simulation

Huai Sun*, Zheng Gong, Yan Xiang

Shanghai Jiao Tong University, School of Chemistry and Chemical Engineering & Center of Material Genome Engineering

Abstract: Molecular dynamics simulations can obtain the physical properties of condensed matters more efficiently than experimental measurements. However, limited by the quality of the molecular force field and sampling efficiency, this method has so far failed to be used to quantitatively predict molecular liquid properties on a large scale. The release of AIMS will change this situation. For the first time, combining artificial intelligence and molecular simulation techniques, we can quantitatively predict the physical properties of molecular fluids on a large scale, which supplements experimental measurements to obtain basic data.

By comparing with the experimental data, we show that the data obtained by the simulation is consistent with the experimental data, and the deviation of the predicted values is equivalent to the uncertainty reported by the experimental data as shown in Figure 1. The method of predicting the physical properties as listed above has been established. On this basis, we have developed a workflow that implements high-throughput molecular force field simulation as shown in Figure 2. Only the identifier of the given molecule (SMILES string) and the temperature and pressure conditions are required as input, and the entire process of building the model, running calculations, verifying and correcting simulation data, and outputting results are automatically performed on super computer. Up to now, we have calculated more than 10,000 molecular fluids and obtained more than 2.5 million data points.

Using the big data of simulation, we developed a machine learning model to predict physical properties as shown in Figure 3. The significance of artificial intelligence is not only to expand the scope of prediction, but also to have a two-way feedback mechanism between machine learning and molecular simulation. Combined with the evolving molecular simulation technology and force field method, a sustainable learning and predicting mechanism is formulating, which is expected to eventually realize the desire of obtaining the basic thermodynamic data solely from computations.

In this talk, we will present our work, demonstrate the AIMS system, and discuss the challenges and issues.

AIMS - 结合人工智能和分子模拟技术构建的分子液体物理性质数据库系统

孙淮*，龚正，向衍

上海交通大学，化学化工学院及材料基因组中心

摘要：分子动力学模拟可以比实验测量更高效地获得凝聚态的物理性质。然而，受到分子力场的质量和采样效率的限制，迄今为止这种方法未能用来大规模地定量地预测分子液体性质。AIMS的发布将改变这一状态。第一次，结合人工智能和分子模拟技术，我们可以大规模地定量地预测分子液体的物理性质，补充和替代实验手段获得基础数据。

利用大量的模拟计算数据，我们开发了一个人工智能机器学习模型，初步数据如图三所示。人工智能的意义不仅是可以扩大预测的范围，更在于机器学习和分子模型之间具有双向反馈机制，结合不断发展的分子模拟技术和力场方法，可以形成一个能够不断学习增强预测能力的自洽系统，有望最终实现人类用计算机获得基本热力学数据的愿望。

在本报告中，我们将介绍上述研究成果，演示AIMS系统，并讨论面临的挑战和问题。

关键词：分子模拟，人工智能，分子液体，热力学性质，数据库