Discovery of High-Performance Low-k Interconnect Materials via Multi-Target Machine Learning Model

Abstract

With the continuous development of the semiconductor industry in terms of miniaturization and performance improvement, the demand for high-frequency and energy-efficient semiconductor devices, traditional dielectric materials, represented by dense amorphous silicon dioxide (SiO2) grown via thermal oxidation or CVD, which is widely used in conventional interconnect and gate oxide processes with an industry-standard dielectric constant k≈3.9, can no longer meet the increasingly stringent requirements for signal delay reduction and power consumption management. While low-dielectric constant (low-k) materials are essential, they often suffer from an inherent trade-off in mechanical strength, which is critical for surviving chemical mechanical polishing (CMP) and packaging stresses. This study focuses on designing a machine learning (ML) model to identify high-performance low-k materials that can be utilized in advanced interconnects.Our research presents a machine learning framework to identify next-generation low-k materials that balance low-k and high Young’s modulus (E). We curated a dataset of 3,206 materials from the Materials Project database, incorporating a diverse set of structural, electronic, and elemental descriptors. Data preprocessing involved robust data cleaning, including missing value imputation and standardized outlier pruning, to ensure dataset integrity. We developed a multi-task regression model based on the Gradient Boosting algorithm (XGBoost) to concurrently predict the k and E. The model was optimized using automated hyperparameter optimization protocols to ensure high predictive fidelity across the low-k regime.The predictive model achieved a coefficient of determination (R2) of 0.87 for k and 0.98 for E on the test set. Through high-throughput screening of 729 candidates (from Materials Project), three materials: Mg3Si4(HO6)2, SiO2, and B2S2O9 were identified and subsequently validated through density functional theory (DFT) calculations, showing high consistency with ML predictions. These materials exhibit k values below 2.6 and excellent mechanical properties (E>70 GPa).The discovery of these high-stiffness, low-k materials offers promising methods for mitigating resistance-capacitance (RC) delay while ensuring structural integrity during the CMP process or constructing porous structures in advanced packaging and 3D integration.