Comprehensive autonomous R2R using Agentic AI to address today’s rising CMP processing challenges

Copper (Cu) dishing and erosion continue to pose significant and persistent challenges in Chemical Mechanical Planarization (CMP), particularly in the context of advanced semiconductor packaging technologies such as 3D integration, hybrid bonding, and ultra-fine-pitch interconnects. As semiconductor devices evolve with tighter geometries and greater integration complexity, ensuring surface planarity and minimizing topographic defects have become mission critical. For instance, in hybrid bonding applications, even minimal surface non-uniformities — such as dishing greater than 5–10 nanometers — can result in bond voids and electrical discontinuities, ultimately degrading device yield, reliability, and long-term performance.

While traditional Run-to-Run (R2R) process control strategies have historically provided some degree of process stabilization by adjusting thickness targets and polish times, these methods are frequently based on fixed linear models and rely heavily on offline measurements. As such, they fall short in addressing modern manufacturing realities that include complex material pattern dependencies, consumable-induced drift, and highly dynamic process conditions. Moreover, the continued reliance on Send Ahead wafers and manual engineering analysis introduces undesirable delays, increased cycle times, and additional cost burdens.

This paper presents a next-generation, autonomous CMP R2R control solution built on the Sentient platform using Agentic AI. By integrating Digital Twin modeling, real-time metrology, Fault Detection and Classification (FDC), Statistical Process Control (SPC), and inline yield learning, the framework establishes a unified, intelligent control environment. Central to this architecture is a CMP-specific ontology that enables the system to understand and adapt to evolving tool states, pad wear dynamics, slurry chemistry shifts, and inter-wafer variability. The platform continuously learns and refines process behavior, enabling predictive corrections that are executed autonomously.

A notable advancement introduced in this work is the development of “smart reset” models that intelligently recalibrate the process without requiring Send Ahead wafers following pad conditioning or replacement events. Additionally, the system incorporates predictive models to forecast post-CMP metrics — including within-wafer thickness uniformity, dishing, erosion, and defect density — thereby enabling real-time process adjustments that directly optimize product quality and downstream yield.

The paper concludes with a real-world case study demonstrating how an Autonomous R2R framework significantly enhances CMP process stability, reduces defectivity, improves throughput, and supports broader digital transformation initiatives. This marks a pivotal step toward agentic semiconductor manufacturing where data-to-action efficiency is achieved through closed-loop intelligence and adaptive automation.