Sequences of projection images collected during in situ tomography experiments can capture the formation of patterns in crystallization and yield their three-dimensional growth morphologies. These image streams generate enormous and high dimensional datasets that span the full extent of a phase transition. Detecting from the continuous image stream the characteristic times and temperatures at which the phase transition initiates is a challenge because the phase change is often swift and subtle. Here, we show a flexible unsupervised binary classification algorithm to identify a change point during data intensive experiments. The algorithm makes a prediction based on statistical metrics and has a quantifiable error bound. Applied to two in situ X-ray tomography experimental datasets collected at a synchrotron light source, the developed method can detect the moment at which the solid phase emerges from the parent liquid phase upon crystallization and without performing computationally expensive volume reconstructions. Our approach is verified using a simulated X-ray phantom and its performance evaluated with respect to solidification parameters. The method presented here can be broadly applied to other Big Data problems where time series can be classified without the need for additional training data.

在原位层析成像实验中收集的投影图像序列可以捕获结晶过程中图案的形成，并产生其三维生长形态。这些图像流生成了巨大的高维数据集，这些数据集涵盖了整个相变范围。从连续图像流中检测相变开始的特征时间和温度是一个挑战，因为相变通常是迅速而微妙的。在这里，我们展示了一种灵活的无监督二进制分类算法，可以在数据密集型实验中识别变化点。该算法基于统计指标进行预测，并具有可量化的误差范围。应用于在同步加速器光源下收集的两个原位X射线断层摄影实验数据集，所开发的方法可以检测结晶时固相从母液相中出现的时刻，而无需进行计算上昂贵的体积重建。我们的方法已使用模拟的X射线体模进行了验证，并针对凝固参数评估了其性能。本文介绍的方法可以广泛应用于其他大数据问题，在这些问题中可以对时间序列进行分类而无需其他训练数据。