基于大语言模型与深度学习的CO2加氢制甲醇催化剂性能筛选与预测

doi:10.12422/j.issn.1672-6952.2026.02.009

摘要/Abstract

摘要：

为解决CO2加氢制甲醇催化剂开发效率低的问题，构建并验证一种基于大语言模型 (Large Language Model,LLM)与深度学习的性能智能预测模型。首先，利用LLM设计结构化指令，实现了从大量文献中半自动化、高效率地提取多维催化剂数据；采用带梯度惩罚的Wasserstein生成对抗网络（WGAN⁃GP）对稀疏的原始数据集进行高质量增强，有效克服了数据量不足的瓶颈；在此基础上，经数据清洗、特征工程与降维处理后，采用超参数优化的多层感知机（Multi⁃layer Perceptron，MLP）构建了预测模型。结果表明，优化后的MLP模型在独立测试集上对CO2转化率与甲醇选择性的预测决定系数（R2）分别高达0.972 3和0.969 3。基于SHAP(SHapley Additive exPlanations)的特征分析结果表明，BET(Brunauer⁃Emmett⁃Teller)比表面积和铜基催化剂是影响催化性能的主导因素，且铟（In）基催化剂对金属质量分数具有独特依赖性。整合LLM与WGAN⁃GP的数据驱动模型可为新型催化剂的快速筛选与理性设计提供有力工具，展现了人工智能(Artificial Intelligence,AI)在催化研究中的巨大应用潜力。

关键词: CO₂加氢, 甲醇合成, 催化剂性能预测, 大语言模型, 机器学习

Abstract:

To address the low efficiency in developing catalysts for CO2 hydrogenation to methanol, this study constructs and validates an intelligent performance prediction model based on large language model (LLM) and deep learning. First, a Large Language Model (LLM) to design structured prompts, achieving semi⁃automated and high⁃efficiency extraction of multi⁃dimensional catalyst data from literature. Subsequently, a Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN⁃GP) is employed to augment the sparse original dataset, effectively overcoming the bottleneck of data scarcity. Following data cleaning, feature engineering, and dimensionality reduction, a hyperparameter⁃optimized Multi⁃Layer Perceptron (MLP) is constructed as the prediction model. The results show that the optimized MLP model achieves high prediction accuracy on an independent test set, with R² values for CO2 conversion and methanol selectivity reaching as high as 0.972 3 and 0.969 3, respectively. SHAP⁃based feature analysis reveals that BET surface area and Cu⁃based catalysts are the dominant factors affecting catalytic performance, and also uncovered the unique dependency of In⁃based catalysts on metal content. This data⁃driven model, integrating LLM and WGAN⁃GP, provides a powerful tool for the rapid screening and rational design of novel catalysts, demonstrating the great potential of AI in catalysis research.

Key words: CO₂ hydrogenation, Methanol synthesis, Catalyst performance prediction, Large language model, Machine learning

中图分类号:

TQ426

刘庆辉, 李子怡, 余皓, 杨思宇. 基于大语言模型与深度学习的CO₂加氢制甲醇催化剂性能筛选与预测[J]. 辽宁石油化工大学学报, 2026, 46(2): 78-87.

Qinghui LIU, Ziyi LI, Hao YU, Siyu YANG. Performance Prediction of Catalysts for CO₂ Hydrogenation to Methanol Based on Large Language Model and Deep Learning[J]. Journal of Liaoning Petrochemical University, 2026, 46(2): 78-87.

图/表 12

表1 最终输入变量及其类型和范围

Table 1 Final input variables, their types and ranges

序号	输入变量	变量性质	分类详情或数据范围	序号	输入变量	变量性质	分类详情或数据范围
1	金属类型	分类	Cu、Rh、In、Pd	12	体积空速/h^-1	连续	500~550 000
2	w（金属）/%	连续	Cu(3.0~72.0),Rh(0.5~6.0),In(0.1~100.0),Pd(0.7~10.0)	13	进料中n（H₂）/n（CO₂）	连续	0.6~9.0
3	助剂类型	分类	Ga、Mn、Zr等共36种	14	反应器	分类	固定床反应器等共7种
4	w（助剂）/%	连续	0~16.7	15	制备方法	分类	共沉淀法、浸渍法等共27种
5	载体	分类	Al₂O₃、CeO₂、ZrO₂等共17种	16	BET比表面积/(m²∙g⁻¹)	连续	Cu(2.3~72.0),Rh(N/A), In(23.7~227.0),Pd(11.0~100.0)
6	焙烧温度/K	连续	393~1 073	17	孔体积/ (cm³∙g⁻¹)	连续	Cu(0.038~1.557),Rh(N/A), In(0.142~0.548),Pd(N/A)
7	焙烧时间/h	连续	1~16	18	金属比表面积/(m² ∙g⁻¹)	连续	Cu(0.75~96.00),Rh(N/A), In(N/A),Pd(N/A)
8	还原温度/K	连续	333~973	19	金属颗粒粒径/nm	连续	Cu(2.0~44.0),Rh(N/A), In(1.0~86.0), Pd(1.4~28.3)
9	还原时间/h	连续	0.5~12.0	20	电负性	连续	Cu(1.21~1.90), Rh(1.54~1.92), In(1.33~1.93), Pd(1.27~1.92)
10	反应温度/K	连续	423~573	21	共价半径/pm	连续	Cu(112~195), Rh(112~160), In(113~175), Pd(113~194)
11	反应压力/MPa	连续	0.1~8.0

图1 主成分分析的方差贡献率

Fig.1 Variance contribution rate plot of principal component analysis

图2 WGAN?GP模型在训练过程中损失函数值的变化曲线

Fig.2 Loss function variation curve of the WGAN?GP model during the training process

图3 原始数据与WGAN生成数据的箱线图对比（前20个主成分）注：箱内横线表示中位数，箱体上下边界分别表示第75百分位数和第25百分位数（即四分位距），上下须线表示数据范围，括号内数值为方差贡献率，下同。

Fig.3 Box plot comparison of original data and WGAN generated data (Top 20 components)

图4 原始数据与WGAN生成数据的相关性矩阵对比（a）原始数据相关性矩阵（b） WGAN生成数据相关性矩阵（c）相关性差异矩阵

Fig.4 Comparison of correlation matrices between original data and WGAN generated data.

图5 优化模型在训练过程中的损失值变化曲线

Fig.5 Loss variation curve during optimized model training process

表2 优化后模型的性能评估结果

Table 2 Performance evaluation of optimization models

预测指标	R²	均方误差	均方根误差	平均绝对误差
训练集催化体积反应速率	0.863 5	0.001 9	0.043 6	0.015 3
训练集甲醇选择性	0.984 1	0.001 0	0.032 3	0.016 3
测试集催化体积反应速率	0.972 3	0.000 4	0.019 8	0.014 3
测试集甲醇选择性	0.969 3	0.001 8	0.042 8	0.023 0

图6 优化模型在测试集上的CO2转化率和甲醇选择性的预测结果

Fig.6 Prediction of CVR and methanol selectivity by the optimized model on the test set

图7 模型的预测误差交叉分析热图

Fig.7 Cross?analysis heat map of model prediction error

图8 预测误差分布的分层对比分析

Fig.8 Stratified comparative analysis of prediction error distribution

图9 模型预测的全局特征重要性分析

Fig.9 Global feature importance analysis of model prediction

图10 基于金属类型的特征重要性分解分析

Fig.10 Hierarchical decomposition analysis of feature importance based on metal types

参考文献 14

[1]	中国发展改革报社. “塞北的风点亮湾区的灯”——我国“十四五”以来能源事业取得突破性进展和历史性成就［EB/OL］. （2025⁃09⁃18）［2025⁃10⁃20］. https：//www.ndrc.gov.cn/wsdwhfz/202509/t20250918_1400562.html.
[2]	袁铁江，高玲玉，谢永胜，等. 基于氢能的风⁃火耦合多能系统设计与综合评估［J］. 电力自动化设备， 2021， 41（10）： 227⁃233.
	YUAN T J， GAO L Y， XIE Y S， et al. Design and comprehensive evaluation of wind⁃thermal power coupling multi⁃energy system based on hydrogen energy［J］. Electric Power Automation Equipment， 2021， 41（10）： 227⁃233.
[3]	NASCIMENTO DA SILVA G， ROCHEDO P R R， SZKLO A. Renewable hydrogen production to deal with wind power surpluses and mitigate carbon dioxide emissions from oil refineries［J］. Applied Energy， 2022， 311： 118631.
[4]	MARTIN O， MARTÍN A J， MONDELLI C， et al. Indium oxide as a superior catalyst for methanol synthesis by CO₂ hydrogenation［J］. Angewandte Chemie International edition， 2016， 55（21）： 6261⁃6265.
[5]	徐敏杰，朱明辉，陈天元，等. CO₂高值化利用： CO₂加氢制甲醇催化剂研究进展［J］. 化工进展， 2021， 40（2）： 565⁃576.
	XU M J， ZHU M H， CHEN T Y， et al. High value utilization of CO₂：Research progress of catalyst for hydrogenation of CO₂ to methanol［J］. Chemical Industry and Engineering Progress， 2021， 40（2）： 565⁃576.
[6]	SUVARNA M，ARAÚJO T P，PÉREZ⁃RAMÍREZ J. A generalized machine learning framework to predict the space⁃time yield of methanol from thermocatalytic CO₂ hydrogenation［J］. Applied Catalysis B⁃Environmental， 2022， 315： 121530.
[7]	AKLILU E G， BOUNAHMIDI T. Machine learning applications in catalytic hydrogenation of carbon dioxide to methanol： A comprehensive review［J］. International Journal of Hydrogen Energy， 2024， 61： 578⁃602.
[8]	FOPPA L， RÜTHER F， GESKE M， et al. Data⁃centric heterogeneous catalysis：Identifying rules and materials genes of alkane selective oxidation［J］. Journal of the American Chemical Society， 2023， 145（6）： 3427⁃3442.
[9]	DEVLIN J，CHANG M W，LEE K，et al. BERT： Pre⁃training of deep bidirectional transformers for language understanding［C］//Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics： Human Language Technologies， Volume 1 （Long and Short Papers）. Minneapolis： Association for Computational Linguistics， 2019： 4171⁃4186.
[10]	GULRAJANI I， AHMED F， ARJOVSKY M， et al. Improved training of wasserstein GANs［C］//Proceedings of the 31st International Conference on Neural Information Processing Systems. Long Beach： Curran Associates Inc.， 2017： 5769⁃5779.
[11]	BERGSTRA J， BENGIO Y. Random search for hyper⁃parameter optimization［J］. The Journal of Machine Learning Research， 2012， 13（1）： 281⁃305.
[12]	LUNDBERG S M，LEE S I.A unified approach to interpreting model predictions［C］//Proceedings of the 31st International Conference on Neural Information Processing Systems. Long Beach： Curran Associates Inc.， 2017： 4768⁃4777.
[13]	BEHRENS M， STUDT F， KASATKIN I， et al. The active site of methanol synthesis over Cu/ZnO/Al₂O₃ industrial catalysts［J］. Science， 2012， 336（6083）： 893⁃897.
[14]	SU X， XU J H， LIANG B L， et al. Catalytic carbon dioxide hydrogenation to methane： A review of recent studies［J］. Journal of Energy Chemistry， 2016， 25（4）： 553⁃565.

基于大语言模型与深度学习的CO₂加氢制甲醇催化剂性能筛选与预测

Performance Prediction of Catalysts for CO₂ Hydrogenation to Methanol Based on Large Language Model and Deep Learning

HTML

PDF

可视化

摘要/Abstract

引用本文

使用本文

图/表 12

参考文献 14

相关文章 1

编辑推荐

Metrics