Recent developments in computer processing power lead to new paradigms of how problems in many-body physics and especially polymer physics can be addressed. Parallel processors can be exploited to generate millions of molecular configurations in complex environments at a second, and concomitant free-energy landscapes can be estimated. Databases that are complete in terms of polymer sequences and architecture form a powerful training basis for cross-checking and verifying machine learning-based models. We employ an exhaustive enumeration of polymer sequence space to benchmark the prediction made by a neural network. In our example, we consider the translocation time of a copolymer through a lipid membrane as a function of its sequence of hydrophilic and hydrophobic units. First, we demonstrate that massively parallel Rosenbluth sampling for all possible sequences of a polymer allows for meaningful dynamic interpretation in terms of the mean first escape times through the membrane. Second, we train a multi-layer neural network on logarithmic translocation times and show by the reduction of the training set to a narrow window of translocation times that the neural network develops an internal representation of the physical rules for sequence-controlled diffusion barriers. Based on the narrow training set, the network result approximates the order of magnitude of translocation times in a window that is several orders of magnitude wider than the training window. We investigate how prediction accuracy depends on the distance of unexplored sequences from the training window.

计算机处理能力的最新发展带来了新的范式，说明了如何解决多体物理学，尤其是高分子物理学中的问题。可以利用并行处理器在一秒钟内在复杂的环境中生成数百万个分子构型，并且可以估算伴随的自由能态势。完整的聚合物序列和体系结构数据库为交叉检查和验证基于机器学习的模型提供了强大的培训基础。我们采用了详尽的聚合物序列空间枚举来对神经网络的预测进行基准测试。在我们的例子中，我们认为共聚物通过脂质膜的移位时间是其亲水和疏水单元序列的函数。第一，我们证明，对于聚合物所有可能的序列进行大规模平行的Rosenbluth采样，就可以通过膜的平均首次逸出时间进行有意义的动态解释。其次，我们在对数易位时间上训练了多层神经网络，并且通过将训练集减少为易位时间的狭窄窗口，表明神经网络为序列控制的扩散障碍建立了物理规则的内部表示。基于狭窄的训练集，网络结果近似于比训练窗口宽几个数量级的窗口中的移位时间的数量级。我们调查预测准确性如何取决于训练窗口中未探索序列的距离。