Extreme ultraviolet (EUV) lithography is expected to succeed in 193-nm immersion multi-patterning technology for -sub-10-nm critical layer patterning. In order to be successful, EUV lithography has to demonstrate that it can satisfy the industry requirements in the following critical areas: power, dose stability, etendue, spectral content, and lifetime. Currently, development of second-generation laser-produced plasma (LPP) light sources for the ASML's NXE: 3300B EUV scanner is complete, and first units are installed and operational at chipmaker customers. We describe different aspects and performance characteristics of the sources, dose -stability results, power scaling, and availability data for EUV sources and also report new development results.

As the critical dimensions required in mask making and direct write by electron beam lithography become ever smaller, correction for proximity effects becomes increasingly important. Furthermore, the problem is beset by the fact that only a positive energy dose can be applied with an electron beam. We discuss techniques such as chopping and dose shifting, which have been proposed to meet the positivity requirement. An alternative approach is to treat proximity correction as an optimization problem. Two such methods, local area dose correction and optimization using a regularizer proportional to the informational entropy of the solution, are compared. A notable feature of the regularized proximity correction is the ability to correct for forward scattering by the generation of a ''firewall'' set back from the edge of a feature. As the forward scattering width increases, the firewall is set back farther from the feature edge. The regularized optimization algorithm is computationally time consuming using conventional techniques. However, the algorithm lends itself to a microelectronics integrated circuit coprocessor implementation, which could perform the optimization faster than even the fastest work stations. Scaling the circuit to larger number of pixels is best approached with a hybrid serial/parallel digital architecture that would correct for proximity effects over 10(8) pixels in about 1 h. This time can be reduced by simply adding additional coprocessors. (C) 1996 Society of Photo-Optical Instrumentation Engineers.

As the critical dimensions required for masks and e-beam direct write become ever smaller, the correction of proximity effects becomes more necessary. Furthermore, the problem is beset by the fact that only a positive energy dose can be applied with the e-beam. We discuss here approaches such as chopping and dose shifting which have been proposed to meet the positivity requirement. An alternative approach is to treat proximity correction as an optimization problem. Two such methods, local area dose correction and optimization using a regularizer proportional to the informational entropy of the solution, are compared. A notable feature of the regularized proximity correction is the ability to correct for forward scattering by the generation of a 'firewall' set back from the edge of a feature. As the forward scattering width increases, the firewall is set back further from the feature edge. The regularized optimization algorithm is computationally time consuming using conventional techniques. However, the algorithm lends itself to a microelectronics integrated circuit coprocessor implementation which could perform the optimization much faster than even the fastest work stations. Scaling the circuit to larger number of pixels is best approached with a hybrid serial/parallel digital architecture which would correct for proximity effects over 108 pixels about one hour. This time can be reduced by simply adding additional coprocessors.

This is the first of a series of articles aimed at laying the foundations for construction of special purpose computing machinery for solving the proximity effect problem. In this article, we compare three different types of proximity correction approach (basic matrix inversion, gradient descent optimization, and optimization using a Shannon entropy regularizer). We show that entropy regularization does not result in physically unrealizable negative dose requests as does the matrix inversion and related methods, and always provides optimizations closer to target than those derived from ‘‘least-mean-square’’ type gradient descent. The implications of this approach for integrated co-processor design are outlined.

In this paper, we present a neural net co-processor capable of performing computed tomographic image reconstruction. The circuit performs the Radon transformation using a cost function gradient descent method. The unique aspect of this co-processor is the incorporation of informational entropy as a regularizer in the optimization problem. A 10 pixel×10 pixel array was designed and fabricated in 2 μm CMOS technology. Convergence time of the array was less than 5 μs. Issues relating to scaling the array to larger sizes are discussed in this paper