Machine learning (ML) can replace mechanism-based solutions, such as first-principle calculation, for speeding up fundamental researches. Although ML has the benefits of representing the material's properties with critical descriptors without involving the physical/chemical mechanisms, the reliability of data-driven models remain a great challenge because of the scarcity and irregular distribution of data sets. Here, we develop several models with different input features and ML methods. We find the artificial neural network (ANN) with reasonable features that can greatly alleviate these two challenges by a case study of early prediction of supercapacitors (SCs) cycle lives. We generate a comprehensive data set consisting 88 commercial SCs cycled under different charging strategies, with widely varying cycle lives up to 10,000 cycles. Based on the ANN model, we achieve the early prediction of SCs' cycle life with the test errors less than 10.9%, only using the first 16% cycles, and such error could be further tuned by the data set. The proposed model is suitable for training widely distributed data set and has accurate early diagnosis and prediction ability for the performance of complex SC systems.

机器学习（ML）可以替代基于机制的解决方案，例如第一性原理计算，以加快基础研究的速度。尽管ML具有在不涉及物理/化学机制的情况下用关键描述符表示材料特性的好处，但是由于数据集的稀缺性和不规则分布，数据驱动模型的可靠性仍然是一个巨大的挑战。在这里，我们开发了几种具有不同输入功能和ML方法的模型。通过对超级电容器（SCs）循环寿命的早期预测的案例研究，我们发现了具有合理特征的人工神经网络（ANN），可以大大缓解这两个挑战。我们生成了一个包含88个商业SC的综合数据集，这些SC在不同的充电策略下循环运行，循环寿命变化很大，可达10,000个循环。基于ANN模型，我们仅使用前16％的循环就可以实现SC循环寿命的早期预测，且测试误差小于10.9％，并且可以通过数据集进一步调整此类误差。该模型适用于训练分布广泛的数据集，并具有对复杂SC系统的性能的准确的早期诊断和预测能力。