Intelligent design method of organic amine for carbon dioxide absorption based on machine learning

Abstract

The invention provides an intelligent design method of organic amine for carbon dioxide absorption based on machine learning based on a Kolmogorov-Arnold Networks (KAN) machine learning model and a chemical interpretation descriptor (such as MATSd, RPCG, BCUTdv-11 and the like). By screening the chemical descriptors of amine molecules and analyzing the correlation between the chemical descriptors and the upper and lower limits of a cycle period, possible combination speculation of potential organic amine chemical descriptors is finally realized. Compared with a traditional experiment-driven trial and error method, high-throughput prediction of molecular properties is realized through a machine learning technology, the descriptor speculation efficiency is greatly improved, and meanwhile, the research and development cost is reduced. By utilizing a machine learning model, chemical descriptor combinations of other potential amine molecules can be speculated, and theoretical support and technical reserve are provided for innovative development of a carbon capture technology.

Description

Technical Field

[0001] This invention belongs to the field of carbon capture and computer-aided molecular property prediction technology, and crosses chemical reaction engineering, cheminformatics and data-driven modeling. Specifically, it relates to an intelligent design method for the use of organic amine solvents for carbon dioxide absorption based on interpretable machine learning. Background Technology

[0002] Carbon capture technology has received widespread attention in recent years, becoming one of the important solutions for addressing climate change and reducing greenhouse gas emissions globally. Currently, the intensification of global climate change and the continuous increase in greenhouse gas concentrations have severely impacted human production, lives, and ecosystems. Carbon dioxide (CO2), as a major greenhouse gas, has its emission control a key step in achieving carbon neutrality. Among existing carbon capture technologies, organic amine solvents have become one of the commonly used CO2 absorbents due to their excellent CO2 absorption performance. However, the application effects of existing organic amine solvents still have many shortcomings. Determining the cycling capacity of organic amines using traditional experimental methods requires a long time, which severely restricts the improvement of carbon capture technology efficiency and cost reduction. Therefore, more quickly estimating the cycling cycle of organic amine solvents has become a key breakthrough point for promoting the development of carbon capture technology.

[0003] This invention proposes a machine learning model for predicting the upper and lower limits of the cycle time of organic amines, based on the Kolmogorov–Arnold Networks (KAN) machine learning model and chemical interpretive descriptors (such as MATSd, RPCG, BCUTdv-11, etc.). By screening chemical descriptors of amine molecules and analyzing their correlation with the upper and lower limits of the cycle time, the possible combinations of chemical descriptors for potential organic amines can be predicted. Compared with traditional experimentally driven trial-and-error methods, this invention achieves high-throughput prediction of molecular properties through machine learning technology, significantly improving the efficiency of descriptor prediction while reducing R&D costs. Using the machine learning model, other potential combinations of chemical descriptors for amine molecules can be predicted, providing theoretical support and technical reserves for the innovative development of carbon capture technology. Summary of the Invention

[0004] This invention aims to provide a small-sample-friendly and engineering-interpretable intelligent prediction method for the cycling cycle of amine molecules: after data cleaning, an analytical formula is obtained by training a KAN model and an upper / lower boundary model of the cycling cycle is constructed, which significantly reduces the uncertainty of experimental rounds and energy consumption assessment. This method includes the following specific steps:

[0005] A machine learning-based intelligent design method for organic amines in carbon dioxide absorption, with the following specific steps:

[0006] Step 1: Perform data construction and cleaning to form clean data of "molecule-chemical descriptor-cycle upper and lower limits" that can be trained, and ensure consistency across sources;

[0007] Step 2: Using the KAN model combined with chemical descriptors, interpretable analytical formulas and robust generalization performance are obtained under small sample conditions, and a cycle period upper and lower limit boundary model is established.

[0008] Furthermore, step 1 specifically includes:

[0009] Step 1.1: Input publicly available papers / patents / reports and experimental data; process the data, standardize the units, create an Excel spreadsheet, and record the data;

[0010] Step 1.2: Perform data cleaning, clean the descriptors, and repeat the experiment to input the data into the database with the mean ± standard deviation;

[0011] Step 1.3: Characterize the structure and physicochemical properties of molecules using molecular descriptors;

[0012] Step 1.4: Perform feature selection, variance threshold filtering, and select suitable descriptors; calculate feature correlation; and perform normalization.

[0013] Furthermore, step 2 specifically includes:

[0014] Step 2.1: Construct a predictive model for the cycle life of amine molecules using the KAN model combined with chemical interpretive descriptors;

[0015] Step 2.2: Construct a training strategy, derive analytical formulas, derive formulas for sparse / symbolic models, and label variables and coefficients item by item;

[0016] Step 2.3: Use the P-value to assess uncertainty and obtain a distribution table of P-values ​​for all descriptors;

[0017] Step 2.4: Train the model to obtain the upper / lower boundary models of the cycle period respectively;

[0018] Step 2.5: Perform optimization using Bayesian optimization to recommend the next possible combination of descriptors.

[0019] The beneficial effects of this invention are as follows: This invention achieves high-throughput prediction of molecular properties through machine learning technology, significantly improving the efficiency of descriptor inference while reducing research and development costs. Utilizing machine learning models, it is possible to infer chemical descriptor combinations for other promising amine molecules, providing theoretical support and technological reserves for the innovative development of carbon capture technology. Attached Figure Description

[0020] Figure 1 : Characteristic standard deviation distribution (Top-500 highlighted).

[0021] Figure 2 Correlation heatmap (after filtering).

[0022] Figure 3 : p-value distribution (α=0.05 reference line). Detailed Implementation

[0023] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0024] Step 1.1 (Literature Retrieval and Experimental Data Import): Input published papers / patents / reports and experimental data; process the data, standardize the units, create an Excel spreadsheet, and record the data.

[0025] Step 1.2 (Denoising and Standardization): Perform data cleaning, clean the descriptors, and repeat the experiment to store the data with mean ± standard deviation.

[0026] Step 1.3: Characterize the structure and physicochemical properties of molecules using molecular descriptors.

[0027] Step 1.4: Perform feature selection and variance threshold filtering to select suitable descriptors (combined with...) Figure 1 "Standard deviation Top-500 highlighted"); calculate feature correlation (combined with...) Figure 2 Correlation heatmap); normalization processing (z-score / quantile transformation).

[0028] Step 2.1: Construct a predictive model for the cycle life of amine molecules using the KAN model in combination with chemical interpretive descriptors (such as MATSd, RPCG, BCUTdv-11, etc.).

[0029] Step 2.2: Construct training strategies, derive analytical formulas, derive formulas for sparse / symbolic models, and label variables and coefficients item by item.

[0030] Step 2.3: Use the p-value to assess uncertainty and obtain a p-value distribution table for all descriptors. (The determination coefficients R² and Q² obtained during training meet the threshold, supporting the correctness of the formula and its engineering usability; Figure 3 The "p-value distribution" is used to corroborate significance.

[0031] Step 2.4: Train the model to obtain the upper / lower boundary models of the cycle period.

[0032] Step 2.5: Perform optimization using Bayesian optimization to recommend the next possible combination of descriptors.

[0033] Cycle upper / lower limit models and indicators

[0034]

[0035] R²_train is the coefficient of determination for the training set. The closer R² is to 1, the better the model fits the data.

[0036] MAE_train is the mean absolute error of the training set. The smaller the MAE, the smaller the difference between the predicted value and the true value, and the better the model fits.

[0037] Q²_cv is the cross-validation determination coefficient. The closer Q² is to 1, the better the model fit.

[0038] Verification shows that the coefficients of determination R² and Q² meet the engineering prediction requirements, proving that the obtained analytical formula is correct and reliable.

[0039] Appendix A: Analytical Formulas for the Upper Limit Model (Excerpt)

[0040] y = -5.03835171989169*sin(0.330875861295597*sin(1.77210350713432*x_1- 4.18236398603359) - 0.154719420634304*cos(1.98021393535876*x_2 -2.80064879770362) + 0.242589124250965*cos(1.77201298680901*x_6 -6.3850927190636) + 0.100251928288786) + 4.73489024610483*sin(0.741636772484263*(1 - 0.0407156066657769*x_3)**5 - 0.121185553771268*(-0.768182060111469*x_6 - 1)**2 + 0.179561555123906*sin(1.1519618334489*x_1 +2.20021960738131) + 0.388364534228067*cos(1.11913922531497*x_5 +7.22311039859431) - 3.48099891963712) + 4.99633409139691*sin(2.08870992597668*(0.64070308290503 - x_3)**2 - 4.55109301254073*cos(2.21890796559286*x_2 + 2.57759379888509) + 0.585283324621086*tan(1.36680436139377*x_5 - 3.5877393910317) - 0.66816244772983 +0.316006047475207*exp(-11.7812073625596*(0.671649534090562 - x_4)**2) -2.60852789963131*exp(-5.77655171289829*(0.372139436950323 - x_6)**2)) -2.54744479729621*tan(-0.364433923049727*(0.520002558552035 - x_1)**2 +0.434739573979487*(x_4 + 0.520904479456935)**(3 / 2) + 0.878468836680769*sin(1.72236592768487*x_5 - 4.37301496540513) + 0.728858880210459*cos(1.3338717497249*x_2 + 4.02716080145554) - 0.529080503085737*tanh(2.24554764708753*x_6 - 2.26895132352688) + 5.14123673860734) +10.9372856105756 + 14.3465739962967*exp(-5.52200293616082*(0.127923859059937*sin(2.43856270485689*x_1 - 1.58817372063389) + 0.238514437389765*sin(3.90236852291352*x_3 - 4.84369404220581) - sin(1.35459165144294*x_4 -1.82919853253907) + 0.0733687184939283*cos(2.22099817270538*x_2 +6.40143692743445) - 0.79981231046168*cos(1.08429078825334*x_6 +0.967983047814844) + 0.234168855905931*tanh(1.79927346798233*x_5 -1.52408238327567) - 0.120847124868323)**2)R2_train: 0.989MAE_train: 0.450Q2_cv: 0.960.

[0041]

[0042] Appendix B: Analytical Formula for the Lower Bound Model (Excerpt)

[0043] y = -0.349428578528038*sin(3.49962899813505*sin(2.68990833302538*x_1- 8.62408056149364) - 2.70517982625662*atan(2.94611457091691*x_3 -1.02840982928049) + 4.38531902511491) - 2.44923948776345*cos(15.7476252621745*(1 - 0.0635890270231108*x_4)**5 + 1.14659210246834*sin(3.79172681096967*x_5 + 0.0108426486482458) + 3.23474762392004*cos(1.41626643754552*x_6 - 5.80092031809972) - 18.4730327230356) +5.81995115027801*tan(0.138885424516073*cos(6.3043080372373*x_6 -1.14349810452236) - 0.16725065584885*tanh(4.99940800804515*x_3 -3.13339879045364) + 2.98869276438863 - 0.759817942105766*exp(-24.6293105965414*(0.926202363506932*x_5 - 1)**2) + 0.85003193882956*exp(-1.76860777097075*(1 - 0.90856557358717*x_2)**2) - 0.166495989649419*exp(-5.19593452203231*(1 - 0.841963148769481*x_1)**2)) + 5.44554699910821 +3.9408993284118*exp(-14.2829848671775*(-0.666585326718011*sin(1.24711442676031*x_3 + 2.3575946445362) - 0.343388769478227*cos(3.30857505903258*x_1 - 5.25145451170291) - 0.183992441826103*tanh(2.91880100141035*x_2 - 1.66922716901805) + 0.605333745467643*atan(1.57030213121319*x_5 - 0.422911255218474) + 1 - 0.490670787084809*exp(-1.22008528429048*(-x_6 - 0.182259156807783)**2))**2) + 5.03111769312307*exp(-386.532419954995*(-0.990551858083693*(-0.0222190902428438*x_4 - 1)**4 -0.0228496673651229*sin(1.72314897002189*x_3 - 7.36795877711595) +0.0205456015196282*tan(2.33377383600747*x_1 + 2.02100809654115) +0.0225116015905844*atan(5.13756746071617*x_5 - 1.22229872706578) + 1 +0.0290967313368916*exp(-35.5436390243928*(1 - 0.972358987659252*x_2)**2))**2)+ 2.61750781223201*exp(-71.475120845404*(-(-0.0342535990727945*x_4 - 1)**5 +0.180742230773053*atan(2.39952150610765*x_1 - 0.930740395276446) +0.125180851533644*atan(2.98447706003727*x_6 - 1.15131200189021) -0.683906653340301 - 0.37517256792901*exp(-6.14034543309356*(-x_5 -0.0666114001552542)**2))**2)R2_train: 0.987MAE_train: 0.371Q2_cv: 0.973.

[0044]

Claims

1. An intelligent design method for organic amines for carbon dioxide absorption based on machine learning, characterized in that, Specific steps: Step 1: data construction and cleaning, forming a clean data of "molecule-chemical descriptor-cycle period upper and lower limit" that can be trained, and ensuring consistency across sources; Step 2: using KAN model combined with chemical descriptors to obtain interpretable analytical formula and robust generalization performance under small sample conditions, and establishing cycle period upper and lower limit boundary model.

2. The machine learning based intelligent design method for organic amine for carbon dioxide absorption according to claim 1, wherein, Step 1 specifically includes: Step 1.1: input public papers / patents / reports, experimental data; data processing, unit conversion, excel table establishment, data recording; Step 1.2: data cleaning, descriptor cleaning, repeated experiments with mean ± standard deviation into database; Step 1.3: describe the structure and physicochemical properties of molecules through molecular descriptors; Step 1.4: feature selection, variance threshold filtering, selecting appropriate descriptors; calculating feature correlation; normalization processing.

3. The method for intelligent design of organic amines for carbon dioxide absorption based on machine learning as claimed in claim 1, wherein, Step 2 specifically includes: Step 2.1: using KAN model combined with chemical explanatory descriptors to build amine molecule cycle period capacity prediction model; Step 2.2: build training strategy, derive formula, derive formula for sparse / symbolic model, label variables and coefficients one by one; Step 2.3: use P value to evaluate uncertainty, get P value distribution table of all descriptors; Step 2.4: model training, get cycle period upper / lower limit boundary model respectively; Step 2.5: optimization, using Bayesian optimization to recommend the next possible descriptor combination.

Images