Accelerating the Discovery of Low-Dielectric Polyimides Based on Interpretable Machine Learning

Rapidly advancing computer technology has demonstrated
great potential
in the development of promising materials. Polyimide (PI), an important
dielectric material, is widely applied in modern industries. The demand
for devices with low dielectric and low energy consumption in the
microelectronics industry has led to a growing need for low dielectric
constant (Dk) PI. Developing Dk prediction models using machine learning (ML) algorithms
can help construct quantitative structure–property relationships
and deeply understand the molecular nature that affects the dielectric
properties of PI. In this study, we assembled a data set comprising
970 PIs and extracted key structural features using RDKit. Among the
12 popular ML algorithms, the Extra Trees-based model yielded the
most accurate results, achieving a coefficient of determination of
0.897 for the test set, a mean absolute error of 0.194, and a root-mean-square
error of 0.267. SHapley Additive exPlanation analysis was then employed
to explain the optimal model for Dk prediction
from a physicochemical point of view and structural aspects. The study
identified the BCUT2D_LOGPHI descriptor as being particularly influential
on Dk, showing a negative correlation.
More importantly, eight potential PI candidates with low Dk values were designed according to the chemical insights
of the key descriptors, which were verified through all-atom molecular
dynamics (MD) simulation. By comparison of the calculated and predicted Dk values, the lowest prediction deviation was
found to be approximately 1.06%. The proposed methodology achieved
a prediction accuracy comparable to that of traditional MD simulation,
but the computational time and resource consumption were dramatically
reduced. And Schuffenhauer’s synthetic accessibility scores
were used to evaluate the ease of synthesis of each PI before the
experiment. This research demonstrates the feasibility of using ML
methods to accelerate the property prediction and molecular design
and provides a powerful tool for discovering high-performance dielectric
materials in the future.