Machine Learning-Enhanced Computational Reverse-Engineering Analysis for Scattering Experiments (CREASE) for Analyzing Fibrillar Structures in Polymer Solutions
Machine Learning-Enhanced Computational Reverse-Engineering Analysis for Scattering Experiments (CREASE) for Analyzing Fibrillar Structures in Polymer Solutions
复制标题
DOI：
10.1021/acs.macromol.2c02165
复制
发表时间：
2022-12-12
期刊：
影响因子：
5.5
通讯作者：
Jayaraman，Arthi
中科院分区：
文献类型：
--
作者：
Wu，Zijie;Jayaraman，Arthi
在这项工作中，我们提出了一种机器学习（ML）增强的散射实验计算逆向工程分析（CREASE）方法，以分析具有纤维直径分散的组装半柔性原纤维的聚合物溶液（例如甲基纤维素原纤维的水溶液）的小角散射曲线。这项工作是对原来的CREASE方法的改进[Beltran-Villegas， D. J.；J.]点。化学。Soc。[j] .光子学报，2019,41(1),14916−14930]，该方法通过聚合物溶液的小角度散射曲线识别出组装结构的相关尺寸，而不依赖于传统的分析模型。在这里，我们改进了原始的CREASE方法，结合ML来分析具有分散纤维直径的组装半柔性纤维结构。我们首先通过将已知尺寸（直径，库恩长度）的硅结构的散射曲线作为输入，并在误差范围内复制这些已知尺寸作为输出，来验证我们在没有ML的情况下的CREASE方法。然后，我们展示了如何在CREASE方法中结合ML（特别是人工神经网络，表示为NN）在不牺牲确定的纤维尺寸的准确性的情况下提高工作流程的速度。最后，我们将神经网络增强的CREASE应用于由Lodge， Bates和同事获得的甲基纤维素原纤维的实验小角度x射线散射剖面[Schmidt， P. W.；Macromolecules, 2018, 51, 7767−7775]来确定纤维直径分布，并将nn增强的CREASE输出与使用分析模型拟合的纤维直径分布进行比较。从nn增强的CREASE中得到的甲基纤维素原纤维的直径分布与从分析模型拟合中得到的相似，证实了Lodge， Bates和同事的结果，即甲基纤维素形成的原纤维具有一致的平均直径约15-20 nm，而与甲基纤维素链的分子量无关。神经网络增强的CREASE在处理具有维度分散的复杂大分子组装结构的实验散射剖面方面的成功实施，表明其在其他可能没有适当分析模型的非常规纤维系统中的应用潜力。
In this work, we present a machine learning (ML)-enhanced computational reverse-engineering analysis of scattering experiments (CREASE) approach to analyze the small-angle scattering profiles from polymer solutions with assembled semiflexible fibrils with dispersity in fibril diameters (e.g., aqueous solutions of methylcellulose fibrils). This work is an improvement over the original CREASE method [Beltran-Villegas, D. J.;J. Am. Chem. Soc., 2019, 141, 14916−14930], which identifies relevant dimensions of assembled structures in polymer solutions from their small-angle scattering profiles without relying on traditional analytical models. Here, we improve the original CREASE approach by incorporating ML for analyzing assembled semiflexible fibrillar structures with disperse fibril diameters. We first validate our CREASE approach without ML by taking as input the scattering profiles of in silico structures with known dimensions (diameter, Kuhn length) and reproducing as output those known dimensions within error. We then show how the incorporation of ML (specifically an artificial neural network, denoted as NN) within the CREASE approach improves the speed of workflow without sacrificing the accuracy of the determined fibrillar dimensions. Finally, we apply NN-enhanced CREASE to experimental small-angle X-ray scattering profiles from methylcellulose fibrils obtained by Lodge, Bates, and co-workers [Schmidt, P. W.;Macromolecules, 2018, 51, 7767−7775] to determine fibril diameter distribution and compare NN-enhanced CREASE’s output with their fibril diameter distribution fitted using analytical models. The diameter distributions of methylcellulose fibrils from NN-enhanced CREASE are similar to those obtained from analytical model fits, confirming the results by Lodge, Bates, and co-workers that methylcellulose form fibrils with consistent average diameters of ∼15–20 nm regardless of the molecular weight of methylcellulose chains. The successful implementation of NN-enhanced CREASE in handling experimental scattering profiles of complex macromolecular assembled structures with dispersity in dimensions demonstrates its potential for application toward other unconventional fibrillar systems that may not have appropriate analytical models.