【作者】 张雪；

【导师】 刘晓燕；

【作者基本信息】 哈尔滨工业大学 ， 计算机科学与技术， 2025， 硕士

【摘要】 近年来,多孔晶态材料在气体吸附、催化、传感和能量存储等各个应用领域大展拳脚,尤其是其中的共价有机框架材料由于其突出的易于设计性、高表面积、高热稳定性和拓扑结构多样性,引起了广泛的关注。但是目前共价有机框架材料的设计大多依赖于研发人员的经验和试错方法,研发周期长、试错成本高,提出快速高效的材料设计方法迫在眉睫。面对此类功能性材料研发任务,飞速发展的机器学习,尤其是深度学习无疑是非常有力的工具。随着人们对新分子、新材料的需求越来越大,近年来出现了不少关于分子设计的方法研究,这些研究提出的方法在药物研发领域大显身手,但这些方法大多是只是搜索具有药物相似性等期望性质的新分子,同时满足化合价等一些简单的化学限制,在结构限制严格的多孔晶态材料研发领域表现欠佳。本文将材料分子设计任务建模为一个马尔可夫决策过程,并在强化学习算法的基础上,将分子语法融入到环境设计当中,以此方法来对生成的分子结构进行约束,并在这样的结构限制内通过最大化奖励函数优化分子的性能,提出了一个能够将分子的结构要求和性能需求相结合并共同优化设计的新分子生成方法。本文提出了两种针对多孔晶态材料构建单元分子的语法嵌入方法。首先使用硬性的强制语法对生成分子的过程进行固定的结构限制,以基序为基础单元,在状态空间中引入对称列表并在动作空间中引入对称动作元组,通过对当前对称状态的记录以及对称动作的执行来确保生成的分子的对称性。实验表明,该方法在保证分子符合结构要求的同时,在Penalized log P属性优化的任务中取得了领先的效果;而且在以共价有机框架材料的氧气吸附属性作为优化目标的实验中,使用该方法生成的构建单元分子进行拼接得到了上万种合理的共价有机框架材料并对它们的氧气吸附属性进行预测。不过这种方法比较依赖于初始提供的分子基序的数据集,并因此牺牲了生成分子库的多样性。为了提升生成分子库的多样性,本文在该方法的基础上对语法的嵌入方式进行了改进,将邻域控制嵌入语法根据分子的特点拓展为分子邻域控制嵌入语法,将一个新分子的生成解释为一个解析树的生成,其中树中的每个节点代表一条语法,并将从多孔晶态材料构建单元分子中抽取得到的合理的语法用作环境中的动作空间,将状态空间设计为一个可以解码为分子的语法序列。实验表明,该方法在Penalized log P属性优化的任务中表现更好,而且针对共价有机框架材料的设计问题能够生成多个符合结构要求的构建模块分子库,并且相较于第一种语法嵌入方法更为优雅,生成的分子数据库多样性更高。总的来说,本文提出了两种分子语法嵌入方法,补足了当前针对强结构要求的分子设计领域研究的不足,不仅考虑了分子属性的优化,而且考虑到了分子结构层面的需求。更多还原

【Abstract】 Currently,crystalline porous materials have made great strides in various application fields such as gas adsorption,catalysis,sensing and energy storage,especially covalent organic framework materials,which have attracted widespread attention due to their out-standing ease of design,high surface area,high thermal stability and topological diversity.However,at present,the design of covalent organic framework materials mostly relies on the experience and trial and error methods.The research and development cycle is long and the trial and error cost is high.So it is urgent to propose a fast and efficient material design method.In the face of such functional material research and development tasks,the rapidly developing machine learning,especially deep learning,is undoubtedly a very powerful tool.With the increasing demand for new molecules and new materials,a lot of research on molecular design has emerged in recent years.The methods proposed in these stud-ies have shown great promise in the field of drug research and development.But most of these methods are only to search for new molecules with desired properties such as drug similarity and meet some simple chemical limitations such as valency.Almost all of them perform poorly in the field of porous crystalline materials research with strict structural restrictions.This thesis models the material molecule design task as a Markov decision process,and the molecular grammar is integrated into the environment design based on the re-inforcement learning algorithm.In this way,the generated molecular structure is con-strained,and the molecular performance is optimized by maximizing the reward function within such structural constraints.In this thesis,a new molecular generation method is proposed,which can combine the structural requirements and performance requirements of molecules and optimize together.In this thesis,two grammatical embedding methods for building unit molecules of porous crystalline materials are proposed.Firstly,a rigid mandatory grammar is used to fix the structural constraints of the process of generating molecules.Based on the mo-tifs,a symmetric list is introduced into the state space and a symmetric action tuple is introduced into the action space.The symmetry of the generated molecules is ensured by recording the current symmetric state and executing the symmetric action.Experiments show that the proposed method is leading in the optimization of Penalized log P prop-erties while ensuring that the molecules meet the structural requirements.Moreover,in the experiment with oxygen adsorption properties of covalent organic frame materials as the optimization target,tens of thousands of reasonable covalent organic frame materials were obtained by splicing the building block molecules generated by this method and their oxygen adsorption properties were predicted.However,this approach relies on the initially provided dataset of molecular motifs,and thus sacrifices the diversity of the generated molecular library.In order to improve the diversity of the generative molecule library,this thesis improves the embedding mode of grammar on the basis of this method,The neighborhood controlled embedding grammar is expanded into a molecular neighborhood controlled embedding grammar according to the characteristics of molecules,and the generation of a new molecule is interpreted as the generation of a parse tree,where each node in the tree represents a grammar.The reasonable grammar extracted from the building unit molecules of porous crystalline ma-terials is used as the action space in the environment,and the state space is designed as the sequences of production rules that can be decoded into molecules.Experiments show that this proposed method performs better in the task of Penalized log P attribute optimiza-tion,and can generate multiple building unit molecular libraries that meet the structural requirements for the design of covalent organic framework materials.And this method is more elegant than the first grammar embedding method,and the molecular database diversity is higher.In general,two molecular grammatical embedding methods are proposed in this the-sis,which complement the current research in the field of molecular design with strong structure requirements,and consider not only the optimization of molecular properties,but also the needs of molecular structure level.更多还原