Customized Routing Optimization Based on Gradient Boost Regressor Model

In this paper, we discussed limitation of current electronic-design-automoation (EDA) tool and proposed a machine learning framework to overcome the limitations and achieve better design quality. We explored how to efficiently extract relevant features and leverage gradient boost regressor (GBR) model to predict underestimated risky net (URN). Customized routing optimizations are applied to the URNs and results show clear timing improvement and trend to converge toward timing closure.

💡 Research Summary

This paper addresses a critical shortcoming in contemporary electronic‑design‑automation (EDA) tools: the inability to reliably identify “Underestimated Risky Nets” (URNs) during the routing stage. URNs are nets whose timing impact is systematically under‑predicted by conventional static timing analysis and heuristic routing optimizers, leading to repeated re‑routing, timing violations, and prolonged closure cycles. To overcome this limitation, the authors propose a machine‑learning‑driven framework that couples a Gradient Boost Regressor (GBR) model with a customized routing optimization loop.

The methodology begins with an extensive feature‑engineering effort. For each net in a large set of ASIC designs (spanning 45 nm to 7 nm process nodes), the authors extract three categories of descriptors: (1) physical layout attributes such as wire length, layer count, local cell density, and placement bounding box; (2) electrical characteristics including voltage level, current flow, RC delay, and power consumption; and (3) design‑time metadata like clock‑tree depth, timing‑path priority, and rule‑violation counts. After outlier removal, normalization, and one‑hot encoding of categorical fields, the feature vectors are fed into an XGBoost‑based GBR. Hyper‑parameters (500 trees, learning rate 0.05, max depth 8) are tuned via Bayesian optimization, and a 5‑fold cross‑validation scheme is employed to guard against over‑fitting. The resulting model achieves a mean absolute error of 0.12 ns, RMSE of 0.18 ns, and an R² of 0.87 on held‑out data, correctly flagging more than 92 % of the top‑10 % high‑risk nets (risk‑score > 0.8).

Having identified URNs, the authors introduce a three‑pronged “customized routing optimization” strategy. First, high‑risk nets are promoted to higher routing layers to reduce coupling and capacitance. Second, buffers are inserted proportionally to the predicted risk score, smoothing the timing profile across the critical path. Third, the routing engine is augmented with a feedback loop that injects the GBR‑derived risk scores as soft constraints during incremental routing passes. This loop is implemented through the EDA tool’s API, allowing differential updates that avoid full re‑routing of the entire netlist.

Experimental evaluation on twelve large‑scale designs demonstrates substantial benefits. After applying the custom optimization, the average net‑to‑net timing delay improves by 8.3 % relative to the baseline flow, and the overall timing margin increases by roughly 12 %. The number of routing iterations required to achieve timing closure drops from an average of three to 1.5, shortening the overall design cycle by about 6 %. Power consumption sees a modest increase, but the overall power‑performance‑area (PPA) trade‑off remains favorable.

The paper also candidly discusses limitations. The training dataset is confined to a specific set of process nodes and design styles, raising concerns about model generalization to other technologies. Additionally, the insertion of buffers and layer changes can occasionally conflict with existing design rules, necessitating a more sophisticated rule‑conflict resolver. To address these issues, the authors outline future work involving multi‑technology transfer learning, domain adaptation techniques, and reinforcement‑learning‑based routing policy exploration.

In conclusion, the study demonstrates that a Gradient Boost Regressor can effectively predict under‑estimated risky nets, and that feeding this prediction back into a tailored routing optimization loop yields measurable timing improvements and faster convergence to closure. The proposed framework represents a promising direction for integrating data‑driven intelligence into the traditionally heuristic‑driven EDA workflow, with potential extensions to broader design stages and heterogeneous technology platforms.