# Multifunctional polyimide performance prediction based on explainable machine learning

## TL;DR

This study uses machine learning to predict key performance properties of polyimides, helping design better materials for microelectronics.

## Contribution

The novel contribution is combining interpretable machine learning with predictive modeling for polyimide properties to accelerate material discovery.

## Key findings

- Predictive models for Tg, CW, and CTE achieved high accuracy and stability.
- Interpretability analysis using SHAP provided insights into model decision-making.
- 135 novel PIs were designed and validated computationally without experiments.

## Abstract

Polyimides (PIs) are widely used in the microelectronics field due to their excellent comprehensive performance and the diversity and designability of their structures. In flexible substrate applications, designing the molecular structure to balance thermodynamic and optical properties is the most critical part of the PI design process. To accelerate the discovery of high‐performance PIs, we established predictive models for glass transition temperature (Tg), cut‐off wavelength (CW), and coefficient of thermal expansion (CTE) using various machine learning algorithms. The optimal predictive models for the three properties demonstrated high accuracy and stability in both test set predictions and cross‐validation results. Additionally, the interpretability of the three optimal models was analyzed using the SHAP method, and the accuracy and generalization ability of the models were validated using several novel PIs. By combining the three models, predictions were made for multiple PIs, leading to the selection and synthesis of PIs with excellent comprehensive performance. 135 novel PIs were designed and their key properties were obtained without the need for experimental verification. The predictive models established in this study can assist researchers in quickly determining the Tg, CW and CTE of PIs, thereby facilitating the swift identification of promising candidates for further development.

Utilizing interpretable machine learning algorithms to develop predictive models for the glass transition temperature, cut‐off wavelength, and coefficient of thermal expansion of polyimides can assist in the design of novel high‐performance polyimides. This approach facilitates the swift identification of promising candidates for further development and provides new insights into the discovery of high‐performance polymers for the future.