能源是人类社会发展的物质基础，能源安全是国家安全的命脉所在，是实现国家可持续发展和民族复兴的重要保障，当前我国能源对外依存度极高，时刻可能成为被人卡住的“最细的脖子”。我国已探明的页岩气和页岩油储量分别居世界第一、第三，但是由于我国页岩油气藏在地质学上相对较为年轻，整体处于中低成熟度，不适合直接开采，需要开发原位催熟技术促进发育。干酪根是石油和天然气的母质，也是地球上最丰富的有机质存在形式，深入理解干酪根结构和熟化机理是指导页岩油气原位催熟与油气藏储量评估的基础。干酪根分子结构模型是从分子层面自下而上地研究干酪根化学-力学性质的基石，但在分子模型重构领域存在“组合爆炸”问题，分子重构难度随着分子量的增大指数增长。干酪根具有起源复杂，分子量大，官能团种类多样的特性，组合爆炸问题尤为突出，导致传统干酪根分子重构方法需要专业人员综合分析多种实验数据，并反复试错逼近真实分子结构，需要花费大量的人力物力，但效率极低，严重制约了对干酪根化学-力学性质的研究。因此，亟待开发新的干酪根分子高通量重构方法，并实现对干酪根分子化学-力学性质的智能化预测。

本学位论文根据上述背景，针对干酪根分子结构模型重构中的组合爆炸和干酪根化学-力学性质智能化预测难题，在力能学指导下，围绕干酪根分子模型智能化重构与干酪根裂解化学-力学机理两大关键科学问题，采用机器学习结合分子动力学模拟、核磁谱模拟计算以及实验分析的方法开展研究。

建设干酪根分子的机器学习数据库。经过标注的海量样本数据是实现机器学习方法的前提条件，但是尚无可满足机器学习高通量重构未知分子模型的数据库，针对此问题，本文通过实验和模拟建设了包含超过 200 万组分子样本的机器学习数据库，其中收录了分子结构、量子轨道构成、热解时序以及拉伸模拟应力-应变曲线等干酪根各项力学、化学特征信息，并且设计了 ¹H 和 ¹³C 核磁谱的一维、二维离散化重构方法，为机器学习方法高通量重构干酪根分子模型和干酪根化学-力学性质研究奠定了坚实基础。

干酪根分子组分、类型和成熟度的机器学习智能化预测。由于干酪根分子模型的复杂度极高，本文设计了从原子到分子层层递进的策略。首先通过机器学习方法，智能化预测了干酪根分子的骨架组分，并实现了对干酪根类型和成熟度指数的高精度分析。验证了通过机器学习方法结合实验数据智能化获取干酪根分子结构信息的可行性，可在不需要人为干预的情况下直接给出分析结果，极大降低干酪根样品的分析和测试成本。

干酪根分子模型的机器学习法智能化重构。为提高干酪根分子模型的预测性能，解决单一谱图特征对机器学习模型训练效率低下的问题，设计了组合核磁谱特征重构方案和与其相匹配的机器学习神经网络模型，实现了多种谱图输入，使得机器学习模型可同时对多种谱图综合分析，获得了显著高于单一谱图分析的预测性能，实现了对干酪根分子模型的高通量重构，有助于缩短对干酪根的熟化机理和力学性能的研究周期。

干酪根分子裂解化学-力学性质分析及预测。通过密度泛函理论和分子动力学等分子模拟方法，从官能团层面研究不同 sp²/sp³ 原子占比干酪根分子模型的高温裂解和拉伸裂解机理，分析了干酪根受热裂解和拉伸裂解的主导机制。并针对干酪根热解行为分析困难的问题，设计机器学习方案，实现了对干酪根热解点位的智能化预测，可为页岩油气藏的生烃潜能评估提供指导。

本论文结合机器学习设计了干酪根分子模型智能化高通量重构方案，并从官能团尺度揭示了干酪根裂解化学-力学行为主导机理，实现对干酪根类型、成熟度及热解位点的智能化预测。为促进非常规油气原位熟化技术的开发提供了更为经济高效的新方法。

Energy is the material basis for the development of modern society. Energy security is the lifeblood of national security and an essential guarantee for sustainable development. The energy dependence of China is extremely high and may become the bottleneck for national rejuvenation. China’s shale oil/gas reserve ranks among the top in the world. However, these resources are unsuitable for exploitation because the shale oil/gas reservoirs are in medium and low maturity. It is necessary to develop in-situ ripening technology to promote the production of shale oil/gas. Kerogen is the primary hydrocarbon source of shale oil/ gas and the most abundant organic matter. Therefore, an in-depth understanding of the kerogen structure and maturation mechanism is fundamental to guiding the in-situ shale oil/gas ripening. Molecular model is the cornerstone for the bottom-up research of kerogen chemo-mechanical properties. Unfortunately, there is a “combinatorial explosion” problem in molecular model reconstruction. Reconstruction complexity increases exponentially with the molecular weight. The combinatorial explosion problem is even more severe for kerogen because of its complex origin, significant molecular weight, and diverse functional groups. Almost all the molecular experimental measurement methods are carried out to reconstruct the kerogen model. And many outstanding ways have been developed. However, Trial-and-error processing is required to approximate the actual structure by traditional methods, which is not only time- and material-consuming but also extremely inefficient. So, it is urgent to develop new high-throughput reconstruction methods of kerogen model and intelligently predict the kerogen chemo-mechanical properties.

This dissertation focuses on two critical academic issues: The high-throughput reconstruction of the kerogen molecular model and the chemo-mechanical mechanism of kerogen cleavage. Machine learning (ML), simulation, and experimental analysis methods are combined to address the challenge.

Massive labeled samples are the prerequisite for the ML method. However, no database now satisfies the need to reconstruct the molecular model by ML. In response to this problem, the ML database containing more than two million molecular samples through experiments and simulations is designed. Nuclear magnetic resonance (NMR) spectra, structural formulas, molecular hybridization information, etc., are redesigned in the database through feature engineering. Thus, the solid foundation is laid for the high-throughput reconstruction of the kerogen molecular model and the high-precision prediction of kerogen chemo-mechanical properties.

By considering the high complexity of the kerogen molecular model, the layer-by-layer progressive strategy from atoms to the whole molecule is adopted in the ML reconstruction method. Kerogen skeleton components are firstly predicted by combining the ML and ¹³C NMR spectra, and the high-precision analysis of the kerogen type and maturity index is fulfilled. Then the feasibility of obtaining kerogen molecular information through ML is verified. Analysis results of kerogen structure can be provided by the ML method without human intervention, significantly reducing the costs of kerogen analysis.

A combined NMR feature reconstruction scheme and the matching ML model are designed to solve the inefficient training of ML models with a single spectral feature. Multiple spectra can be comprehensively analyzed by the developed ML model simultaneously. The prediction accuracy significantly higher than single-spectrum is obtained. The high-throughput reconstruction of the kerogen molecular model is thus achieved, which is helpful in shortening the research cycle of the kerogen maturation mechanism and mechanical properties.

The pyrolysis and tensile cleavage mechanisms of kerogen with different sp²/sp³ atomic ratios are studied at the functional group level through molecular simulations. The dominant factors of thermal and tensile cleavage of kerogen are clarified. An ML method is suggested to realize intelligent prediction of the kerogen cleavage site. The work can guide the evaluation of the hydrocarbon generation potential of shale oil/gas reservoirs.

In this dissertation, a high-throughput reconstruction method of the kerogen molecular model is designed based on ML. And the dominant mechanism for the kerogen chemo-mechanical behavior of cleavage is discussed at the level of functional groups. The high-precision prediction of kerogen type, maturity, and cleavage site is realized. A more efficient and economical novel method is provided to promote the development of unconventional oil/ gas in-situ ripening technology and the evaluate oil/gas reservoirs.