Functional materials, especially those that largely differ from known materials, are not easily discoverable because both human experts and supervised machine learning need prior knowledge and datasets. An autonomous system can evaluate various properties a priori, and thereby explore unknown extrapolation spaces in high-throughput simulations. However, high-throughput evaluations of molecular dynamics simulations are unrealistically demanding. Here, we show an autonomous search system for organic molecules implemented by a reinforcement learning algorithm, and apply it to molecular dynamics simulations of viscosity. The evaluation is dramatically accelerated (by three orders of magnitude) using a femto-second stress-tensor correlation, which underlies the glass-transition model. We experimentally examine one of 55,000 lubricant oil molecules found by the system. This study indicates that merging simulations and physical models can open a path for simulation-driven approaches to materials informatics.

功能材料，尤其是与已知材料有很大差异的功能材料，由于人类专家和受监督的机器学习都需要先验知识和数据集，因此不容易发现。自治系统可以事先评估各种属性，从而在高通量仿真中探索未知的外推空间。但是，对分子动力学模拟的高通量评估要求不切实际。在这里，我们展示了通过强化学习算法实现的有机分子自主搜索系统，并将其应用于粘度的分子动力学模拟。使用飞秒应力张量相关性可以大大加快评估速度（三个数量级），这是玻璃转换模型的基础。我们通过实验检查了55个中的一个，系统发现000个润滑油分子。这项研究表明，将模拟和物理模型合并可以为材料信息学的模拟驱动方法开辟一条道路。