logo

SCIENCE CHINA Chemistry, Volume 64 , Issue 6 : 1039-1046(2021) https://doi.org/10.1007/s11426-020-9969-y

Machine learning-assisted systematical polymerization planning: case studies on reversible-deactivation radical polymerization

More info
  • ReceivedDec 21, 2020
  • AcceptedFeb 24, 2021
  • PublishedMay 17, 2021

Abstract


Funded by

the National Natural Science Foundation of China(21704016,21971044)


Acknowledgment

This work was supported by the National Natural Science Foundation of China (21971044, 21704016) and Fudan University and State Key Laboratory of Molecular Engineering of Polymers.


Interest statement

The authors declare no conflict of interest.


Author information







Supplement

Supporting information

The supporting information is available online at http://chem.scichina.com and http://link.springer.com/journal/11426. The supporting materials are published as submitted, without typesetting or editing. The responsibility for scientific accuracy and content remains entirely with the authors.


References

[1] Corey EJ. Pure Appl Chem, 1967, 14: 19-38 CrossRef Google Scholar

[2] Gentekos DT, Sifri RJ, Fors BP. Nat Rev Mater, 2019, 4: 761-774 CrossRef Google Scholar

[3] Doncom KEB, Blackman LD, Wright DB, Gibson MI, O’Reilly RK. Chem Soc Rev, 2017, 46: 4119-4134 CrossRef Google Scholar

[4] Lynd NA, Meuler AJ, Hillmyer MA. Prog Polym Sci, 2008, 33: 875-893 CrossRef Google Scholar

[5] Fetters LJ, Lohse DJ, Richter D, Witten TA, Zirkel A. Macromolecules, 1994, 27: 4639-4647 CrossRef ADS Google Scholar

[6] Lin B, Hedrick JL, Park NH, Waymouth RM. J Am Chem Soc, 2019, 141: 8921-8927 CrossRef Google Scholar

[7] Corrigan N, Almasri A, Taillades W, Xu J, Boyer C. Macromolecules, 2017, 50: 8438-8448 CrossRef ADS Google Scholar

[8] Rubens M, Junkers T. Polym Chem, 2019, 10: 6315-6323 CrossRef Google Scholar

[9] Walsh DJ, Schinski DA, Schneider RA, Guironnet D. Nat Commun, 2020, 11: 3094 CrossRef ADS Google Scholar

[10] Leibfarth FA, Mattson KM, Fors BP, Collins HA, Hawker CJ. Angew Chem Int Ed, 2013, 52: 199-210 CrossRef Google Scholar

[11] Nicolas J, Guillaneuf Y, Lefay C, Bertin D, Gigmes D, Charleux B. Prog Polym Sci, 2013, 38: 63-235 CrossRef Google Scholar

[12] Moad G, Rizzardo E, Thang SH. Aust J Chem, 2005, 58: 379-410 CrossRef Google Scholar

[13] Ouchi M, Sawamoto M. Macromolecules, 2017, 50: 2603-2614 CrossRef ADS Google Scholar

[14] Matyjaszewski K. Macromolecules, 2012, 45: 4015-4039 CrossRef ADS Google Scholar

[15] Whitfield R, Parkatzidis K, Rolland M, Truong NP, Anastasaki A. Angew Chem Int Ed, 2019, 58: 13323-13328 CrossRef Google Scholar

[16] Ahneman DT, Estrada JG, Lin S, Dreher SD, Doyle AG. Science, 2018, 360: 186-190 CrossRef ADS Google Scholar

[17] Badowski T, Gajewska EP, Molga K, Grzybowski BA. Angew Chem Int Ed, 2020, 59: 725-730 CrossRef Google Scholar

[18] Lin TS, Coley CW, Mochigase H, Beech HK, Wang W, Wang Z, Woods E, Craig SL, Johnson JA, Kalow JA, Jensen KF, Olsen BD. ACS Cent Sci, 2019, 5: 1523-1531 CrossRef Google Scholar

[19] Rizkin BA, Shkolnik AS, Ferraro NJ, Hartman RL. Nat Mach Intell, 2020, 2: 200-209 CrossRef Google Scholar

[20] Gómez-Bombarelli R, Aguilera-Iparraguirre J, Hirzel TD, Duvenaud D, Maclaurin D, Blood-Forsythe MA, Chae HS, Einzinger M, Ha DG, Wu T, Markopoulos G, Jeon S, Kang H, Miyazaki H, Numata M, Kim S, Huang W, Hong SI, Baldo M, Adams RP, Aspuru-Guzik A. Nat Mater, 2016, 15: 1120-1127 CrossRef ADS Google Scholar

[21] Hatakeyama-Sato K, Tezuka T, Umeki M, Oyaizu K. J Am Chem Soc, 2020, 142: 3301-3305 CrossRef Google Scholar

[22] Xie Y, Zhang C, Hu X, Zhang C, Kelley SP, Atwood JL, Lin J. J Am Chem Soc, 2020, 142: 1475-1481 CrossRef Google Scholar

[23] Kim B, Lee S, Kim J. Sci Adv, 2020, 6: eaax9324 CrossRef ADS Google Scholar

[24] Dong Y, Li D, Zhang C, Wu C, Wang H, Xin M, Cheng J, Lin J. Carbon, 2020, 169: 9-16 CrossRef Google Scholar

[25] Butler KT, Davies DW, Cartwright H, Isayev O, Walsh A. Nature, 2018, 559: 547-555 CrossRef ADS Google Scholar

[26] Segler MHS, Preuss M, Waller MP. Nature, 2018, 555: 604-610 CrossRef ADS arXiv Google Scholar

[27] Gong H, Gu Y, Zhao Y, Quan Q, Han S, Chen M. Angew Chem Int Ed, 2020, 59: 919-927 CrossRef Google Scholar

[28] Corrigan N, Yeow J, Judzewitsch P, Xu J, Boyer C. Angew Chem Int Ed, 2019, 58: 5170-5189 CrossRef Google Scholar

[29] Dadashi-Silab S, Doran S, Yagci Y. Chem Rev, 2016, 116: 10212-10275 CrossRef Google Scholar

[30] Chen M, Zhong M, Johnson JA. Chem Rev, 2016, 116: 10167-10211 CrossRef Google Scholar

[31] Gu Y, Wang Z, Gong H, Chen M. Polym Chem, 2020, 11: 7402-7409 CrossRef Google Scholar

[32] Han S, Gu Y, Ma M, Chen M. Chem Sci, 2020, 11: 10431-10436 CrossRef Google Scholar

[33] Vega MP, Lima EL, Pinto JC. Polymer, 2001, 42: 3909-3914 CrossRef Google Scholar

[34] Breiman L. Machine Learning, 2001, 45: 5-32 CrossRef Google Scholar

[35] Haghighatlari M, Li J, Heidar-Zadeh F, Liu Y, Guan X, Head-Gordon T. Chem, 2020, 6: 1527-1542 CrossRef Google Scholar

[36] Tu K, Huang H, Lee S, Lee W, Sun Z, Alexander-Katz A, Ross CA. Adv Mater, 2020, 32: 2005713 CrossRef Google Scholar

[37] Siebert M, Krennrich G, Seibicke M, Siegle AF, Trapp O. Chem Sci, 2019, 10: 10466-10474 CrossRef Google Scholar

[38] Han H, Wang W Y, Mao B H. Borderline-Smote: A New Over-Sampling Method in Imbalanced Data Sets Learning. Berlin, Heidelberg: Springer, 2005. 878–887. Google Scholar

[39] Xu J, Shanmugam S, Duong HT, Boyer C. Polym Chem, 2015, 6: 5615-5624 CrossRef Google Scholar

[40] Xu J, Jung K, Atme A, Shanmugam S, Boyer C. J Am Chem Soc, 2014, 136: 5508-5519 CrossRef Google Scholar

[41] Quan Q, Gong H, Chen M. Polym Chem, 2018, 9: 4161-4171 CrossRef Google Scholar

[42] Kottisch V, Gentekos DT, Fors BP. ACS Macro Lett, 2016, 5: 796-800 CrossRef Google Scholar

[43] Gentekos DT, Dupuis LN, Fors BP. J Am Chem Soc, 2016, 138: 1848-1851 CrossRef Google Scholar

[44] Corrigan N, Manahan R, Lew ZT, Yeow J, Xu J, Boyer C. Macromolecules, 2018, 51: 4553-4563 CrossRef ADS Google Scholar

[45] Rubens M, Vrijsen JH, Laun J, Junkers T. Angew Chem Int Ed, 2019, 58: 3183-3187 CrossRef Google Scholar

[46] Fors BP, Hawker CJ. Angew Chem Int Ed, 2012, 51: 8850-8853 CrossRef Google Scholar

[47] Anastasaki A, Nikolaou V, Zhang Q, Burns J, Samanta SR, Waldron C, Haddleton AJ, McHale R, Fox D, Percec V, Wilson P, Haddleton DM. J Am Chem Soc, 2014, 136: 1141-1149 CrossRef Google Scholar

[48] Ma W, Zhang X, Ma Y, Chen D, Wang L, Zhao C, Yang W. Polym Chem, 2017, 8: 3574-3585 CrossRef Google Scholar

[49] Tian C, Wang P, Ni Y, Zhang L, Cheng Z, Zhu X. Angew Chem Int Ed, 2020, 59: 3910-3916 CrossRef Google Scholar

[50] Xia L, Cheng B, Zeng T, Nie X, Chen G, Zhang Z, Zhang W, Hong C, You Y. Adv Sci, 2020, 7: 1902451 CrossRef Google Scholar

[51] Li S, Han G, Zhang W. Polym Chem, 2020, 11: 1830-1844 CrossRef Google Scholar

[52] Li R, An Z. Angew Chem Int Ed, 2020, 59: 22258-22264 CrossRef Google Scholar

[53] Xu J, Jung K, Boyer C. Macromolecules, 2014, 47: 4217-4229 CrossRef ADS Google Scholar

[54] Tu K, Xu T, Zhang L, Cheng Z, Zhu X. RSC Adv, 2017, 7: 24040-24045 CrossRef ADS Google Scholar

[55] Lee IH, Discekici EH, Anastasaki A, de Alaniz JR, Hawker CJ. Polym Chem, 2017, 8: 3351-3356 CrossRef Google Scholar

  • Figure 1

    Outline of the workflow. The SPP platform proceeds as follows. (a) Collecting the data set and selecting appropriate inputs and outputs for investigation (which in this work are initial feed ratios and the resulted Mw, respectively). (b) Screening and applying eligible ML algorithms to analyze the data set and establishing relationships between the conditions and results. (c) Using the model to acquire synthetic instructions inversely from the target Mw result by traversing the entire condition space. (d) Screening out optimal conditions for achieving Ð requirements through the outputted diversified solutions. (e) Employing transfer functions to incorporate different reaction substrates into the SPP platform, tailoring for polymers with various chemical structures. (a–c) constitute a feedback loop of model establishment, (c–e) delineate the three-dimensional tailoring for polymer synthesis (color online).

  • Figure 1

    Outline of the workflow. The SPP platform proceeds as follows. (a) Collecting the data set and selecting appropriate inputs and outputs for investigation (which in this work are initial feed ratios and the resulted Mw, respectively). (b) Screening and applying eligible ML algorithms to analyze the data set and establishing relationships between the conditions and results. (c) Using the model to acquire synthetic instructions inversely from the target Mw result by traversing the entire condition space. (d) Screening out optimal conditions for achieving Ð requirements through the outputted diversified solutions. (e) Employing transfer functions to incorporate different reaction substrates into the SPP platform, tailoring for polymers with various chemical structures. (a–c) constitute a feedback loop of model establishment, (c–e) delineate the three-dimensional tailoring for polymer synthesis (color online).

  • Figure 2

    The optimized model performances under different ML algorithms. The examined ML algorithms include the widely used (a) Ridge, (b) SVM, (c) kNN, (d) eXtreme Gradient Boosting (XGB), (e) Neural network, (f) Random forest models. (g) RMSE and (h) R2 (coefficient of determination) values measuring the goodness-of-fit for the training set and test set were also displayed (color online).

  • Figure 2

    The optimized model performances under different ML algorithms. The examined ML algorithms include the widely used (a) Ridge, (b) SVM, (c) kNN, (d) eXtreme Gradient Boosting (XGB), (e) Neural network, (f) Random forest models. (g) RMSE and (h) R2 (coefficient of determination) values measuring the goodness-of-fit for the training set and test set were also displayed (color online).

  • Figure 3

    Trellis diagrams of the relationships between model-predicted Mw and three reaction parameters. In each plot one of the parameters was fixed to display the influence of other two parameters on Mw. The fixed value for (a) [M], (b) [CTA], (c) [PC] was varied in each row to track the co-effects between parameters (color online).

  • Figure 3

    Trellis diagrams of the relationships between model-predicted Mw and three reaction parameters. In each plot one of the parameters was fixed to display the influence of other two parameters on Mw. The fixed value for (a) [M], (b) [CTA], (c) [PC] was varied in each row to track the co-effects between parameters (color online).

  • Figure 4

    (a) Experimental Mw results (blue solid dots) and the RMSE values of different trial sets (red hollow dots) during the ML-assisted optimization for Mw at 2.0 × 106 Da. Solid line: Mw = 2.0 × 106 Da line, dashed lines were used to separate different trial sets. (b) Experimental versus predicted Mw plot (color online).

  • Figure 4

    (a) Experimental Mw results (blue solid dots) and the RMSE values of different trial sets (red hollow dots) during the ML-assisted optimization for Mw at 2.0 × 106 Da. Solid line: Mw = 2.0 × 106 Da line, dashed lines were used to separate different trial sets. (b) Experimental versus predicted Mw plot (color online).

  • Figure 5

    Heatmaps involving (a) percentage of propagating group (PG%) and (b) Ð results using CTA-1 under different [CTA]/[M] and [PC]/[CTA] settings ([M] = 0.75 mol/L for all the data points). Black dots are the actual positions of the data points (color online).

  • Figure 5

    Heatmaps involving (a) percentage of propagating group (PG%) and (b) Ð results using CTA-1 under different [CTA]/[M] and [PC]/[CTA] settings ([M] = 0.75 mol/L for all the data points). Black dots are the actual positions of the data points (color online).

  • Figure 6

    Schematic view for the application of SPP platform on PET-RAFT polymerization method. The ML model was built on dataset of experimental and literature results, concerning diversified substrate structures and Mn/Ð results ranging from 1.8×103–2.2×106 Da/1.03–2.10. Upon the establishment of model, suitable conditions of feed ratios, light source and reaction time can be inversely prescribed for synthesizing various polymer targets through PET-RAFT polymerization, with mean absolute percentage error (MAPE) at 4.78%/5.70% for Mn and Ð, respectively (color online).

  • Figure 6

    Schematic view for the application of SPP platform on PET-RAFT polymerization method. The ML model was built on dataset of experimental and literature results, concerning diversified substrate structures and Mn/Ð results ranging from 1.8×103–2.2×106 Da/1.03–2.10. Upon the establishment of model, suitable conditions of feed ratios, light source and reaction time can be inversely prescribed for synthesizing various polymer targets through PET-RAFT polymerization, with mean absolute percentage error (MAPE) at 4.78%/5.70% for Mn and Ð, respectively (color online).

qqqq

Contact and support