Polymer Informatics for Energy Applications

Polymers play a critical role in advancing energy technologies due to their unique properties and broad applicability in mass and energy transport, particularly in gas separation and thermal management. This thesis presents advancements in polymer informatics that integrate computational simulations, predictive machine learning (ML), and generative design to accelerate the discovery of high-performance polymers for energy applications.

We first introduce a graph-augmented, imbalanced ML framework for predicting polymer gas permeability and identifying promising polymers that surpass conventional performance limits. We then extend this approach to ladder polymers—materials with superior separation performance but limited prior exploration—by developing PolyLand, an ML-powered platform that combines predictive modeling and three generative strategies: template-based design, graph diffusion transformers, and optimization-aware large language models. The selected ladder-like candidates are validated via molecular simulations and structure-property analysis.

Shifting focus to thermal transport, we explore polymer blends using high-throughput molecular dynamics simulations coupled with an active learning framework. This work reveals how inter- and intra-molecular interactions influence thermal conductivity, identifying blends with improved performance.

To support broader discovery efforts, we introduce POINT², a benchmark suite combining predictive accuracy, uncertainty quantification, interpretability, and synthesizability. POINT² features diverse ML models and representations trained on labeled datasets across multiple polymer properties, including density, glass transition temperature, melting temperature, fractional free volumn, gas permeability, and thermal conductivity.

Together, these contributions demonstrate how data-driven methods can accelerate polymer discovery, enabling more efficient and targeted design of materials for energy-related applications.

聚合物凭借其独特的性能与在质量、能量输运领域的广泛适用性，尤其在气体分离与热管理方向，对能源技术的发展起到关键支撑作用。本论文提出了整合计算模拟、预测型机器学习（Machine Learning, ML）与生成式设计的聚合物信息学研究进展，以加速面向能源应用的高性能聚合物发现进程。

首先，本文提出一种图增强型不平衡机器学习框架，用于预测聚合物气体渗透率，并筛选出突破传统性能上限的高性能聚合物候选体。随后，本文将该方法拓展至梯形聚合物领域——这类材料具备优异的分离性能但此前研究探索不足——通过构建PolyLand平台：这是一款依托机器学习的集成平台，融合了预测建模与三类生成式策略：基于模板的设计、图扩散Transformer以及感知优化的大语言模型（Large Language Model, LLM）。所筛选出的类梯形聚合物候选体通过分子模拟与结构-性能关联分析完成验证。

随后将研究重心转向热输运领域，本文借助高通量分子动力学模拟结合主动学习框架，对聚合物共混体系展开研究。本研究揭示了分子间与分子内相互作用对热导率的调控机制，并筛选出性能优化的聚合物共混体系。

为支撑更广泛的材料发现工作，本文提出POINT²基准测试套件：该套件整合了预测精度、不确定性量化、可解释性与可合成性四大维度。POINT²涵盖了多样化的机器学习模型与表征方式，其训练数据集覆盖了聚合物的多项核心性能指标，包括密度、玻璃化转变温度、熔融温度、自由体积分数、气体渗透率与热导率。

综上，本研究的各项工作证明了数据驱动方法可加速聚合物发现进程，从而为能源相关应用实现更高效、更具针对性的材料设计提供支撑。

搜集汇总

背景与挑战

背景概述

该数据集聚焦聚合物信息学在能源领域的应用，整合了计算模拟、机器学习和生成设计方法，包含多种聚合物性质数据，旨在加速高性能聚合物的发现和设计。数据集由Notre Dame大学发布，采用开放许可。