Artificial intelligence (AI) is revolutionizing polymer science by enabling accurate prop- erty prediction of ideally an infinite number of materials. An example of property which my be problematic for polymers is the solubility parameter, a key factor in material com- patibility and formulation. Traditional methods for its determination, such as group contribution models, have lim- itations due to empirical assumptions and the lack of experimental data for complex macromolecules. In this work, we explore the application of machine learning (ML) for predicting the sol- ubility parameter of a specific class of polymers: waterborne polyurethanes (WPUs). These compounds are receiving incredible attention due to their potential application in bioprinting fields; unfortunately, their characterization is often impossible due to their non-soluble nature in most solvents, requiring complex specific experimental procedures. Three datasets composed of polymers and solvents were built, integrating SMILES rep- resentations, molecular weight, and solubility parameters obtained from multiple group contribution methods (Hansen, Van Krevelen and Hoy) with the respective experimental values. Various ML algorithms, including linear regression, random forest, and extreme gradient boosting, were trained and evaluated to determine the most effective predictive model. The results highlight the potential of ML approaches in improving accuracy over tra- ditional methods, providing a valuable tool for screening novel polymers, particularly in cases where experimental determination is challenging, such as waterborne polyurethanes. Finally, the study highlights the limitations and differences of available datasets and pro- poses potential improvements concerning both the datasets and the models. The findings demonstrate once again AI’s potential in polymer design and analysis, paving the way for new applications in materials science.
L’intelligenza artificiale ha rivoluzionato la scienza dei polimeri, permettendo la predizione di proprietà e la progettazione di nuovi materiali. Questo lavoro si concentra sullo studio delle potenzialità di modelli di intelligenza artificiale applicati in particolare alla predi- zione del parametro di solubilità, essenziale per valutare la compatibilità tra sostanze. La determinazione sperimentale di questo parametro è complessa, soprattutto per i polimeri, e i metodi teorici tradizionali, come quelli dei contributi di gruppo, risultano spesso limi- tanti per strutture avanzate. L’obiettivo finale della tesi è la predizione del parametro di solubilità di poliuretani in dispersione acquosa, composti di crescente interesse per la loro biocompatibilità e versa- tilità. Tuttavia, la loro caratterizzazione sperimentale è spesso difficile essendo insolubili in quasi tutti i solventi più diffusi, richiedendo metodi di caratterizzazione realizzati su misura nei vari casi specifici. Il machine learning (ML) viene esplorato come alternativa per ottenere risultati accurati senza test sperimentali, permettendo di esaminare molte strutture diverse in pochissimo tempo. Sono stati costruiti tre dataset contenenti polimeri e solventi, combinando peso moleco- lare, SMILES, valori sperimentali e teorici di solubilità calcolati con i metodi di Hansen, Van Krevelen e Hoy (uno per dataset). Su questi dati sono stati addestrati tre modelli di ML: regressione lineare, random forest e XGBoost. I risultati vengono poi confrontati con metodi tradizionali, dati di letteratura e Polymer Genome, una delle principali pi- attaforme di predizione delle proprietà polimeriche. L’analisi finale evidenzia limiti e potenzialità dei modelli e dei dataset proposti, sug- gerendo miglioramenti in termini di qualità dei dati e di approcci predittivi più avanzati.
AI for polymers property prediction: investigation on machine learning and deep learning for the prediction of polymers solubility parameter
Ancora, Nicolò
2023/2024
Abstract
Artificial intelligence (AI) is revolutionizing polymer science by enabling accurate prop- erty prediction of ideally an infinite number of materials. An example of property which my be problematic for polymers is the solubility parameter, a key factor in material com- patibility and formulation. Traditional methods for its determination, such as group contribution models, have lim- itations due to empirical assumptions and the lack of experimental data for complex macromolecules. In this work, we explore the application of machine learning (ML) for predicting the sol- ubility parameter of a specific class of polymers: waterborne polyurethanes (WPUs). These compounds are receiving incredible attention due to their potential application in bioprinting fields; unfortunately, their characterization is often impossible due to their non-soluble nature in most solvents, requiring complex specific experimental procedures. Three datasets composed of polymers and solvents were built, integrating SMILES rep- resentations, molecular weight, and solubility parameters obtained from multiple group contribution methods (Hansen, Van Krevelen and Hoy) with the respective experimental values. Various ML algorithms, including linear regression, random forest, and extreme gradient boosting, were trained and evaluated to determine the most effective predictive model. The results highlight the potential of ML approaches in improving accuracy over tra- ditional methods, providing a valuable tool for screening novel polymers, particularly in cases where experimental determination is challenging, such as waterborne polyurethanes. Finally, the study highlights the limitations and differences of available datasets and pro- poses potential improvements concerning both the datasets and the models. The findings demonstrate once again AI’s potential in polymer design and analysis, paving the way for new applications in materials science.File | Dimensione | Formato | |
---|---|---|---|
2024_04_Ancora_Executive_Summary_02.pdf
non accessibile
Dimensione
1.28 MB
Formato
Adobe PDF
|
1.28 MB | Adobe PDF | Visualizza/Apri |
2024_04_Ancora_Tesi_01.pdf
non accessibile
Dimensione
11.54 MB
Formato
Adobe PDF
|
11.54 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/236426