Israrul H. Hashmi, Himanshu, Rahul Karmakar, Tarak K Patra
{"title":"Extrapolative ML Models for Copolymers","authors":"Israrul H. Hashmi, Himanshu, Rahul Karmakar, Tarak K Patra","doi":"arxiv-2409.09691","DOIUrl":null,"url":null,"abstract":"Machine learning models have been progressively used for predicting materials\nproperties. These models can be built using pre-existing data and are useful\nfor rapidly screening the physicochemical space of a material, which is\nastronomically large. However, ML models are inherently interpolative, and\ntheir efficacy for searching candidates outside a material's known range of\nproperty is unresolved. Moreover, the performance of an ML model is intricately\nconnected to its learning strategy and the volume of training data. Here, we\ndetermine the relationship between the extrapolation ability of an ML model,\nthe size and range of its training dataset, and its learning approach. We focus\non a canonical problem of predicting the properties of a copolymer as a\nfunction of the sequence of its monomers. Tree search algorithms, which learn\nthe similarity between polymer structures, are found to be inefficient for\nextrapolation. Conversely, the extrapolation capability of neural networks and\nXGBoost models, which attempt to learn the underlying functional correlation\nbetween the structure and property of polymers, show strong correlations with\nthe volume and range of training data. These findings have important\nimplications on ML-based new material development.","PeriodicalId":501146,"journal":{"name":"arXiv - PHYS - Soft Condensed Matter","volume":null,"pages":null},"PeriodicalIF":0.0000,"publicationDate":"2024-09-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"arXiv - PHYS - Soft Condensed Matter","FirstCategoryId":"1085","ListUrlMain":"https://doi.org/arxiv-2409.09691","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"","JCRName":"","Score":null,"Total":0}
引用次数: 0
Abstract
Machine learning models have been progressively used for predicting materials
properties. These models can be built using pre-existing data and are useful
for rapidly screening the physicochemical space of a material, which is
astronomically large. However, ML models are inherently interpolative, and
their efficacy for searching candidates outside a material's known range of
property is unresolved. Moreover, the performance of an ML model is intricately
connected to its learning strategy and the volume of training data. Here, we
determine the relationship between the extrapolation ability of an ML model,
the size and range of its training dataset, and its learning approach. We focus
on a canonical problem of predicting the properties of a copolymer as a
function of the sequence of its monomers. Tree search algorithms, which learn
the similarity between polymer structures, are found to be inefficient for
extrapolation. Conversely, the extrapolation capability of neural networks and
XGBoost models, which attempt to learn the underlying functional correlation
between the structure and property of polymers, show strong correlations with
the volume and range of training data. These findings have important
implications on ML-based new material development.
机器学习模型已逐渐被用于预测材料特性。这些模型可以利用已有数据建立,可用于快速筛选材料的物理化学空间,而这一空间在经济上是巨大的。然而,ML 模型本质上是内插模型,其在材料已知性能范围之外搜索候选材料的功效尚未得到解决。此外,ML 模型的性能与其学习策略和训练数据量密切相关。在此,我们将确定 ML 模型的外推能力、训练数据集的大小和范围与其学习方法之间的关系。我们将重点放在一个典型问题上,即根据单体序列预测共聚物的性质。树状搜索算法可以学习聚合物结构之间的相似性,但外推法的效率很低。相反,神经网络和 XGBoost 模型的外推能力与训练数据的数量和范围有很强的相关性。这些发现对基于 ML 的新材料开发具有重要意义。