Efficient approach to early detection of lithographic hotspots using machine learning systems and pattern matching

Abstract

Early lithographic hotspot detection has become increasingly important in achieving lithography-friendly designs and manufacturability closure. Fast physical verification tools employing pattern matching or machine learning techniques have emerged as great options for detecting hotspots in the early design stages. In this work, we propose a characterization methodology that provides measurable quantification of a given hotspot detection tool's capability to capture a previously seen or unseen hotspot pattern. Using this methodology, we conduct a side-by-side comparison of two hotspot detection methods-one using pattern matching and the other based on machine learning. The experimental results reveal that machine learning classifiers are capable of predicting unseen samples but may mispredict some of its training samples. On the other hand, pattern matching-based tools exhibit poorer predictive capability but guarantee full and fast detection on all their training samples. Based on these observations, we propose a hybrid detection solution that utilizes both pattern matching and machine learning techniques. Experimental results show that the hybrid solution combines the strengths of both algorithms and delivers improved detection accuracy while sacrificing little runtime efficiency.

EPIC: Efficient prediction of IC manufacturing hotspots with a unified meta-classification formulation

“…In other words, EPIC gives more enhancement in accuracy and false-alarm on C1. This is mostly because C1 class represents the set of lithography hotspot with ANN [10,13] SVM [10,13] PM [14][15][16][17] Hybrid [18] C0 H it rate C0 run time (hour) C1 H it rate C1 run time (hour) (a) ( EPIC ANN [10,13] SVM [10,13] PM [14][15][16][17] Hybrid [18] C0 Extra ratio C0 run time (hour) C1 Extra ratio C1 run time (hour) 32 25 Figure 8: Overall performance comparison in Hit rate, Extra ratio and run-time over C0 and C1 hotspot data 4.5nm to 6.0nm of EPE, while C0 is the set of hotspots with above 6.0nm EPE values. Under our employed RETs, C0 translates to a set of hotspots that have high variability and small quantity (a few hundred out of a layout with totally millions of patterns); whereas C1 is a set of hotspots with less severe variability and much larger quantity.…”

“…(3) will be equivalent to the hybrid flow in [18]. Here EPIC 's advantage lies in the dynamic/automated optimization techniques, thus it can easily reach an optimized solution.…”

“…Moreover, EPIC 's unified formulation covers the static hybrid detection flow proposed in [18] as a special case, i.e., when weighting function f M LK1 (+1) = 0.5 and 0 elsewhere, f M LK2 (+1) = 0.5 and 0 elsewhere, f P M (+1) = 1.0 and 0 elsewhere, θ = 1.0, then EPIC 's formulation Eqn. (3) will be equivalent to the hybrid flow in [18].…”

“…Recently in [18], a hotspot detection flow was proposed to hybridize the strengths of machine learning models and pattern matching models. Such a flow feeds data samples to a pattern matcher first, then employs machine learning classifiers to further examine the non-hotspot data set produced by the pattern matcher.…”

Layout pattern-driven design rule evaluation

“…Several works [6][7][8][9][10][11][12][13] used various techniques of machine learning or fuzzy pattern matching to identify the hotspots in the design. References 14 and 15 suggested a flow that integrates a pattern checker, a pattern fixer, and a router, such that the router completes its job, then the pattern check and fix are performed if needed, and some routes are tentatively redone.…”