Utilising predictive capabilities, mechanical engineers from the University of Wisconsin-Madison have efficiently identified a selection of highly promising high-performance polymers from a pool of 8 million candidates.
These polymers, referred to as polyimides, find extensive application in aerospace, automotive, and electronics due to their exceptional mechanical and thermal characteristics, encompassing strength, rigidity, and resistance to heat. Presently, the availability of polyimides is restricted as their design process incurs substantial costs and consumes considerable time.
Nonetheless, the University of Wisconsin-Madison engineers employ a data-oriented design framework that utilises machine learning predictions and molecular dynamics simulations. This framework enables them to significantly accelerate the discovery of novel polyimides with enhanced properties.
In a recent publication, the team extensively outlined their methodology. The paper highlights the significant implications of their findings in materials science, catalysing future investigations into advancing data-driven methodologies for materials discovery.
Ying Li, an associate professor of mechanical engineering at UW–Madison who spearheaded the research, emphasised the efficiency of their design approach in comparison to traditional trial-and-error processes. Furthermore, Li noted that their strategy can be extended to the molecular design of various other polymeric materials.
The synthesis of polyimides involves a condensation reaction between dianhydride and diamine/diisocyanate molecules. To conduct their research, the engineers initially gathered publicly available data on the chemical structures of existing dianhydride and diamine/diisocyanate molecules. Afterwards, they used this data to construct an extensive library of 8 million hypothetical polyimides.
Ying Li uses an analogy to liken this process to building with LEGO blocks. The dianhydride and diamine/diisocyanate molecules are the fundamental building blocks, with various variations available. However, manually constructing all possible structures would take much work due to the many combinations involved.
Li and the research team employed computer algorithms to systematically combine the building blocks to streamline the process, resulting in an extensive database encompassing all possible combinations.
With the comprehensive database, the researchers developed multiple machine-learning models specifically designed to predict the thermal and mechanical properties of polyimides. These models were constructed based on experimentally reported values. By employing a range of machine learning techniques, the team successfully identified essential chemical substructures that play a crucial role in determining specific properties of the polyimides.
Li explains that their machine-learning model was designed to be transparent and interpretable, enabling human experts to comprehend the reasoning behind their decisions. This transparency ensures that the model is not treated as a “black box” but instead as a comprehensible tool.
Utilising their trained machine learning models, the researchers generated property predictions for 8 million hypothetical polyimides. Afterwards, they thoroughly screened the dataset and identified the top three hypothetical polyimides that exhibited superior combined properties compared to existing polyimides.
The researchers employed several validation methods to ensure their findings’ accuracy and reliability. Firstly, they constructed detailed all-atom models for the top three candidates and conducted molecular dynamics simulations. These simulations were consistent with the predictions made by the machine learning models, providing confidence in the reliability of their predictions. Furthermore, the simulations indicated that synthesising these new polyimides would be relatively straightforward.
As an additional validation step, the team synthesised one of the newly discovered polyimides and conducted experiments to assess its thermal properties. The experimental results confirmed the material’s exceptional heat resistance, with the new polyimide remaining stable up to approximately 1,022 degrees Fahrenheit before degradation. This outcome aligned with the predictions made by the machine learning models. In comparison, existing polyimides could only withstand temperatures within the 392 to 572 degrees Fahrenheit range.
To facilitate exploration and understanding of the new high-performing polyimides, the researchers developed a web-based application that enables users to visualise and explore the properties of these materials interactively.
