基于分子表面电荷密度分布与机器学习的混合物设计方法研究

doi:10.11949/0438-1157.20190795

摘要/Abstract

摘要：

由于混合物性能的可调控性，当前市场对其关注与日俱增。对于这类产品，基于模型的设计方法由于具有高效性以及普适性，相较于其他产品设计方法得到了更快的发展。但是对于很多性质，如气味、颜色等，准确且普适的模型尚不可得。因此，本文提出了一种基于分子表面电荷密度分布描述符（S描述符）和机器学习模型的混合物设计方法，采用描述符表征产品、再通过机器学习模型将其与性质关联，直接用于混合物产品设计。具体地，根据给定的产品性质需求，机器学习模型直接预测/设计混合物产品的S描述符；然后以欧几里德距离为指标，在给定的数据库中筛选出S描述符满足要求的候选混合物组成。最后，对候选混合物及其组分性质进行实验验证，完成设计。本文以香精的混合替代物设计作为算例，设计得到丙酸叶醇酯的两种混合香精替代物，通过实验对混合物进行了验证。结果表明，混合替代物的气味及其组分的各理化性质均与丙酸叶醇酯相近，证实本文所提出方法的有效性。

关键词: 系统工程, 产品设计, 神经网络, 分子表面电荷密度分布, 混合物设计, 香精

Abstract:

Modern business pays increasing attentions to mixture products due to its adjustable characteristics. For the design methods towards such kinds of products, the development of model-based design methods is faster than others, because of its efficiency and wide application. However, the models for some properties, like odor and color, with acceptable accuracy or general application range are still not available. Therefore, an application methodology of machine learning (ML) with molecular surface charge density distribution (Sdescriptors) for mixture product design is proposed in this study, where descriptors are employed to represent the product and ML is responsible for correlating them to the target properties, for the purpose of designing product directly. Specifically, machine learning model is expected to predict Sdescriptors of candidate products according to the assigned property value, and ingredients are screened out using Euclidean-based method according to the predicted descriptors. Finally, the properties of the candidate mixtures and its ingredients are verified by experiments. This methodology is introduced using a case study of mixture substitution fragrance design for cis-3-hexenyl propionate and two mixture fragrances are obtained ultimately. The odor properties of mixtures and physicochemical properties of their components are similar to the target, which highlights the effective of the proposed method.

Key words: systems engineering, product design, neural network, molecular surface charge density distribution, mixture design, fragrance

中图分类号:

TQ 021.8

毛海涛, 王璐, 许志颖, 解万翠, 都健, 张磊. 基于分子表面电荷密度分布与机器学习的混合物设计方法研究[J]. 化工学报, 2020, 71(S1): 282-292.

Haitao MAO, Lu WANG, Zhiying XU, Wancui XIE, Jian DU, Lei ZHANG. Mixture product design based on molecular surface charge density distribution and machine learning[J]. CIESC Journal, 2020, 71(S1): 282-292.

图/表 13

图1 分子表面电荷密度分布谱图以及对其极性区域积分得到的S描述符

Fig.1 Molecular surface charge density distribution spectrum and S descriptors integrated from the polar area of spectrum

图2 机器学习模型的建立流程

Fig.2 Workflow for establishment of machine learning model

图3 混合香精设计流程

Fig.3 Workflow of mixed fragrance design

表1 基团贡献法以及阈值

Table 1 Group contribution method and corresponding thresholds

组分性质	基团贡献法公式	阈值
$δ$ /MPa^1/2	$δ = 1000 H 298 V a p - R T V 298 m 12$	$δ ≥ 158.9$
$T b$ /K	$T b = 244.7889 l n ∑ i ∈ G n i T b, i$	$T b ≥ 333$
$T f$ /K	$T f = 150.0218 + ∑ i ∈ G n i T f, i$	$T f ≥ 323$
$K o / w$	$l n K o / w = 0.4876 + ∑ i ∈ G n i K o / w, i$	$l n K o / w ≤ 2.91$
$L C 50$ /(mol·L^-1)	$- l n L C 50 = 2.18 + ∑ i ∈ G n i L C 50, i$	$- l n L C 50 ≤ 3.70$

表1 基团贡献法以及阈值

Table 1 Group contribution method and corresponding thresholds

组分性质	基团贡献法公式	阈值
$δ$ /MPa^1/2	$δ = 1000 H 298 V a p - R T V 298 m 12$	$δ ≥ 158.9$
$T b$ /K	$T b = 244.7889 l n ∑ i ∈ G n i T b, i$	$T b ≥ 333$
$T f$ /K	$T f = 150.0218 + ∑ i ∈ G n i T f, i$	$T f ≥ 323$
$K o / w$	$l n K o / w = 0.4876 + ∑ i ∈ G n i K o / w, i$	$l n K o / w ≤ 2.91$
$L C 50$ /(mol·L^-1)	$- l n L C 50 = 2.18 + ∑ i ∈ G n i L C 50, i$	$- l n L C 50 ≤ 3.70$

表2 PEN3电子鼻的检测对象以及检出限

Table 2 Detected objections and detection limits of PEN3 E-nose

传感器名称	检测气味类型	检出限/ (mg/kg)
W1C	有机化合物	10
W5S	氮氧化物	1
W3C	氨类	10
W6S	氢气	0.1
W5C	烷烃与非极性有机化合物	1
W1S	甲烷	100
W1W	含硫有机化合物	1
W2S	酒精	100
W2W	无机硫	1
W3S	有机化合物以及脂肪族有机化合物	10

图4 基于10折交叉验证法建立的IML模型的平均验证性能与平均测试性能

Fig.4 Average performances of validation and testing for IML models established by 10-folds cross validation

图5 IML模型在24个测试样本上的预测性能

Fig.5 Prediction performances of inverse machine learning towards 24 testing samples

表3 丙酸叶醇酯的物性需求

Table 3 First consumer needs of cis-3-hexenyl propionate

参数	数值	参数	数值
可食用	32.42	馊味	25.93
烘焙味	43.00	愉悦度	60.19
甜味	34.37	蒸气压/Pa	53.86
水果味	31.54	扩散系数/（m²/h）	0.16
花香味	29.95

图6 IML模型预测结果与Keller数据库中组合的替代混合物之间的S描述符欧氏距离差

Fig.6 Euclidean distance differences between S descriptors predicted by IML models and screened ingredients from Keller database for cis-3-hexenyl propionate substitution

表4 丙酸叶醇酯替代香精的设计结果

Table 4 Results for substitution design of cis-3-hexenyl propionate

参数	丙酸叶醇酯	替代混合物1			替代混合物2
参数	丙酸叶醇酯	4-异丙基苯甲醇	左旋香芹酮	混合物偏差	2-甲基戊酸	2-乙基丁酸烯丙酯	混合物偏差
CAS No.	33467-74-2	536-60-7	6485-40-1	—	97-61-0	7493-69-8	—
体积分数	1	0.2	0.8	—	0.4	0.6	—
溶解度 $δ$ ^①(298 K,水)/(mg/L)	158.9	1687	367.1	472.18	15000	157.3	5935.48
沸点T_b^①(101.325 kPa) /K	453-455	512.66	499.47	47.11-49.11	468.36	449.96	2.32-4.32
闪点T_f^①/K	333	498.15	465.15	138.75	364.26	327.59	9.258
K_o/w^①	2.909	2.37	2.71	-0.267	1.8	2.972	-0.4058
LC₅₀^②/(mol·L^-1)	3.36	3.25	3.39	0.002	2.45	4.03	0.038

表4 丙酸叶醇酯替代香精的设计结果

Table 4 Results for substitution design of cis-3-hexenyl propionate

参数	丙酸叶醇酯	替代混合物1			替代混合物2
参数	丙酸叶醇酯	4-异丙基苯甲醇	左旋香芹酮	混合物偏差	2-甲基戊酸	2-乙基丁酸烯丙酯	混合物偏差
CAS No.	33467-74-2	536-60-7	6485-40-1	—	97-61-0	7493-69-8	—
体积分数	1	0.2	0.8	—	0.4	0.6	—
溶解度 $δ$ ^①(298 K,水)/(mg/L)	158.9	1687	367.1	472.18	15000	157.3	5935.48
沸点T_b^①(101.325 kPa) /K	453-455	512.66	499.47	47.11-49.11	468.36	449.96	2.32-4.32
闪点T_f^①/K	333	498.15	465.15	138.75	364.26	327.59	9.258
K_o/w^①	2.909	2.37	2.71	-0.267	1.8	2.972	-0.4058
LC₅₀^②/(mol·L^-1)	3.36	3.25	3.39	0.002	2.45	4.03	0.038

图7 不同温度下丙酸叶醇酯及其替代混合物组分的蒸气压与扩散系数

Fig.7 Vapor pressure and diffusion coefficient for cis-3-hexenyl propionate and its substituted mixtures’ components of diverse temperatures

表5 丙酸叶醇酯替代混合物的测试实验设备以及试剂

Table 5 Equipment and materials for substitution of cis-3-hexenyl propionate

试剂	说明
丙酸叶醇酯	北京迈瑞达科技有限公司，纯度>98%
2-甲基戊酸	北京迈瑞达科技有限公司，纯度>98%
2-乙基丁酸烯丙酯	北京迈瑞达科技有限公司，纯度>97%
4-异丙基苯甲醇	河南郑州阿尔法化工有限公司，纯度>99%
左旋香芹酮	河南郑州阿尔法化工有限公司，纯度>97%
95%乙醇	天津市富宇化工有限公司

图8 丙酸叶醇酯以及两组混合香精替代物的电子鼻检测雷达图

Fig.8 Rader pictures of E-nose for cis-3-hexenyl propionate and its mixed fragrance substitutions

参考文献 34

1	Rodríguez O, Gomes P, Mata V, et al. Chapter 1 - A Product Engineering Approach in the Perfume Industry [M]//Teixeira M A. Perfume Engineering. Oxford: Butterworth-Heinemann, 2019: 1-13.
2	Wibowo C, Ng K M. Product-oriented process synthesis and development: creams and pastes[J]. AIChE Journal, 2001, 47(12): 2746-2767.
3	Fung K Y, Ng K M. Product-centered processing: pharmaceutical tablets and capsules[J]. AIChE Journal, 2003, 49(5): 1193-1215.
4	Gani R, Brignole E A. Molecular design of solvents for liquid extraction based on UNIFAC[J]. Fluid Phase Equilibria, 1983, 13(83): 331-340.
5	Joback K G. Designing molecules possessing desired physical property values[D]. Massachusetts: Massachusetts Institute of Technology, 1989
6	Conte E, Gani R, Ng K M. Design of formulated products: a systematic methodology[J]. AIChE Journal, 2011, 57: 2431-2449.
7	Kontogeorgis G M, Michele M, Ng K M, et al. An integrated approach for the design of emulsified products[J]. AIChE Journal, 2019, 65: 75-86.
8	张磊, 刘琳琳, 都健. 替代燃油的计算机辅助设计方法[J]. 化工进展, 2018, 37(6): 2438-2444.
	Zhang L, Liu L L, Du J. A computer-aided design methodology for tailor-made surrogate fuels[J]. Chemical Industry and Engineering Progress, 2018, 37(6): 2438-2444.
9	Hornic K. Multilayer feedforward networks are universal approximators[J]. Neural Networks, 1989, 2: 359-366.
10	Raccuglia P, Elbert K C, Adler P D, et al. Machine-learning-assisted materials discovery using failed experiments[J]. Nature, 2016, 533: 73-76.
11	苏荣欣, 邹龙花, 齐崴, 等. 酪蛋白-胰酶水解历程分子量变化模拟与三维表征[J]. 化工学报, 2013, 64(1): 346-351.
	Su R X, Zou L H, Qi W, et al. Simulation and 3D plot of molecular weight distribution of released peptides from pancreatic hydrolysis of casein[J]. CIESC Journal, 2013, 64(1): 346-351.
12	黄凯, 陈勇, 母志为, 等. 基于人工神经网络和遗传算法的甲烷制氢催化剂设计[J]. 化工学报, 2016, 67(8): 3481-3490.
	Huang K, Chen Y, Mu Z W, et al. Catalyst design for production of hydrogen from methane based on artificial neural network and genetic algorithm[J]. CIESC Journal, 2016, 67(8): 3481-3490.
13	安爱民, 刘云利, 张浩琛, 等. 微生物燃料电池的动态性能分析及其神经网络预测控制[J]. 化工学报, 2017, 68(3): 1090-1098.
	An A M, Liu Y L, Zhang H C, et al. Dynamic performance analysis and neural network predictive control of microbial fuel cell[J]. CIESC Journal, 2017, 68(3): 1090-1098.
14	林生岭, 徐绍芬, 王俊德, 等. 钙钛矿型La_xSr_1-_xNi_1-_yCo_yO₃光电催化活性研究[J]. 化学学报, 2005, 63(5): 385-390.
	Lin S L, Xu S F, Wang J D, et al. Study on photo-electro catalytic activity of perovskite type oxides La_xSr_1－_xNi_1－_yCo_yO₃[J]. Acta Chimica Sinica, 2005, 63(5): 385-390.
15	Zhang L, Mao H, Liu L, et al. A machine learning based computer-aided molecular design/screening methodology for fragrance molecules[J]. Computers & Chemical Engineering, 2018, 115: 295-308.
16	Sanchez-Lengeling B, Aspuru-Guzik A. Inverse molecular design using machine learning: generative models for matter engineering[J]. Science, 2018, 361: 360-365.
17	Klamt A, Schueuermann G J. COSMO: a new approach to dielectric screening in solvents with explicit expressions for the screening energy and its gradient[J]. Journal of the Chemical Society, Perkin Transactions II, 1993, 5: 799-805.
18	Rossiter K J. Structure-odor relationships[J]. Chemical Review, 1996, 96: 3201-3240.
19	Klamt A, Reinisch J, Eckert F, et al. Polarization charge densities provide a predictive quantification of hydrogen bond energies[J]. Physical Chemistry Chemical Physics, 2011, 14(2): 955-963.
20	Lin S T, Sandler S I. A priori phase equilibrium prediction from a segment contribution solvation model[J]. Industrial & Engineering Chemistry Research, 2002, 41(5): 899-913.
21	Klamt A, Eckert F, Arlt W. COSMO-RS: an alternative to simulation for calculating thermodynamic properties of liquid mixtures[J]. Annual Review of Chemical and Biomolecular Engineering, 2010, 1: 101-122.
22	Kang X, Liu X, Li J, et al. Heat capacity prediction of ionic liquids based on quantum chemistry descriptors[J]. Industrial & Engineering Chemistry Research, 2018, 57(49): 16989-16994.
23	Kang X, Zhao Z, Qian J, et al. Predicting the viscosity of ionic liquids by the ELM intelligence algorithm[J]. Industrial & Engineering Chemistry Research, 2017, 56(39): 11344-11351.
24	Palomart J, Torrecilla J S, Ferro V R, et al. Development of an a priori ionic liquid design tool (Ⅱ): Ionic liquid selection through the prediction of COSMO-RS molecular descriptor by inverse neural network[J]. Industrial & Engineering Chemistry Research, 2009, 48(4): 2257-2265.
25	Keller A, Vosshall L B. Olfactory perception of chemically diverse molecules[J]. BMC Neuroscience, 2016, 17: 55.
26	周志华. 机器学习[M]. 北京: 清华大学出版社, 2016: 23-53.
	Zhou Z H. Machine Learning[M]. Beijing: Tsinghua University Press, 2016: 23-53.
27	刘方, 徐龙, 马晓迅. BP神经网络的发展及其在化学化工中的应用[J]. 化工进展, 2019, 38(6): 2559-2573.
	Liu F, Xu L, Ma X X. Development of BP neural network and its application in chemistry and chemical engineering[J]. Chemical Industry and Engineering Progress, 2019, 38(6): 2559-2573.
28	de Bruyne M, Foster K, Carlson J R. Odor coding in the Drosophila antenna[J]. Neuron, 2001, 30(2): 537-552.
29	Tamir A. In Applications of Markov Chains in Chemical Engineering[M]. Amsterdam, Netherlands: Elsevier, 1998.
30	Marrero J, Gani R. Group-contribution based estimation of pure component properties[J]. Fluid Phase Equilibria, 2001, 183: 183-208.
31	Hukkerikar A S. Development of pure component property models for chemical product-process design and analysis[D]. Denmark: Technical University of Denmark, 2013.
32	马琦, 伯继芳, 冯莉, 等. GC-MS结合电子鼻分析干燥方式对杏鲍菇挥发性风味成分的影响[J]. 食品科学, 2019, 40(14): 276-282.
	Ma Q, Bo J F, Feng L, et al. Effect of drying method on volatile components of pleurotus eryngii analyzed by combined use of GC-MS and electronic nose[J]. Food Science, 2019, 40(14): 276-282.
33	Reid R C, Prausnitz J M, Poling B E. The Properties of Gases & Liquids[M]. New York: McGrawHill, 1988.
34	Lee B I, Kesler M. A generalized thermodynamic correlation based on three-parameter corresponding states[J]. AIChE Journal, 1975, 21: 510-527.

[1]	温凯杰, 郭力, 夏诏杰, 陈建华. 一种耦合CFD与深度学习的气固快速模拟方法[J]. 化工学报, 2023, 74(9): 3775-3785.
[2]	尹刚, 李伊惠, 何飞, 曹文琦, 王民, 颜非亚, 向禹, 卢剑, 罗斌, 卢润廷. 基于KPCA和SVM的铝电解槽漏槽事故预警方法[J]. 化工学报, 2023, 74(8): 3419-3428.
[3]	诸程瑛, 王振雷. 基于改进深度强化学习的乙烯裂解炉操作优化[J]. 化工学报, 2023, 74(8): 3429-3437.
[4]	闫琳琦, 王振雷. 基于STA-BiLSTM-LightGBM组合模型的多步预测软测量建模[J]. 化工学报, 2023, 74(8): 3407-3418.
[5]	徐野, 黄文君, 米俊芃, 申川川, 金建祥. 多源信息融合的离心式压缩机喘振诊断方法[J]. 化工学报, 2023, 74(7): 2979-2987.
[6]	高学金, 姚玉卓, 韩华云, 齐咏生. 基于注意力动态卷积自编码器的发酵过程故障监测[J]. 化工学报, 2023, 74(6): 2503-2521.
[7]	黄磊, 孔令学, 白进, 李怀柱, 郭振兴, 白宗庆, 李平, 李文. 油页岩添加对准东高钠煤灰熔融行为影响的研究[J]. 化工学报, 2023, 74(5): 2123-2135.
[8]	贠程, 王倩琳, 陈锋, 张鑫, 窦站, 颜廷俊. 基于社团结构的化工过程风险演化路径深度挖掘[J]. 化工学报, 2023, 74(4): 1639-1650.
[9]	吴心远, 刘奇磊, 曹博渊, 张磊, 都健. Group2vec：基于无监督机器学习的基团向量表示及其物性预测应用[J]. 化工学报, 2023, 74(3): 1187-1194.
[10]	王子宗, 索寒生, 赵学良. 数字孪生智能乙烯工厂研究与构建[J]. 化工学报, 2023, 74(3): 1175-1186.
[11]	张江淮, 赵众. 碳三加氢装置鲁棒最小协方差约束控制及应用[J]. 化工学报, 2023, 74(3): 1216-1227.
[12]	吴选军, 王超, 曹子健, 蔡卫权. 数据与物理信息混合驱动的固定床吸附穿透深度学习模型[J]. 化工学报, 2023, 74(3): 1145-1160.
[13]	王雅琳, 潘雨晴, 刘晨亮. 基于GSA-LSTM动态结构特征提取的间歇过程监测方法[J]. 化工学报, 2022, 73(9): 3994-4002.
[14]	高学金, 程琨, 韩华云, 高慧慧, 齐咏生. 基于中心损失的条件生成式对抗网络的冷水机组故障诊断[J]. 化工学报, 2022, 73(9): 3950-3962.
[15]	周乐, 沈程凯, 吴超, 侯北平, 宋执环. 深度融合特征提取网络及其在化工过程软测量中的应用[J]. 化工学报, 2022, 73(7): 3156-3165.