Deep learning (DL) is one of the fastest-growing fields in artificial intelligence (AI). While still in its early stages of adoption, DL has already shown it has the potential to make significant changes to the lithography and photomask industries through the automation or optimization of equipment and processes. The key element required for application of DL techniques to any process is a large volume of data to adequately train the DL neural networks. The accuracy of the classification or prediction of any DL system is dependent on the depth and breadth of the training data to which it is exposed. For semiconductor manufacturing, finding adequate data – especially for corner cases – can be difficult and/or expensive. In this paper, we will present two digital twins that are themselves built from DL as a part of a DL Starter Kit. We will demonstrate the creation of DL-based digital twins for a mask scanning electron microscope (SEM) and for curvilinear inverse lithography technology (ILT).
We have recently demonstrated that curvilinear shapes and multi-beam mask writing are necessary to minimize the impact of mask variability on wafer hotspots. Several key challenges and opportunities remain. We ask how we update mask-inspection rules, and how we correct for mask-process systematics for extreme ultraviolet (EUV), where the optical response must be taken into account. This paper proposes updated mask rule checks (MRC), derived from a mask variability perspective. We also demonstrate the need for MRC-aware inverse lithography technology (ILT) metrics as a pre-requisite to ensure the reticle shapes are what the wafer-side lithographer desires. Armed with a fully curvilinear ILT and mask data preparation (MDP) system, there is an opportunity to relax the restrictions on fully Manhattan designs where possible.
Deep Learning (DL) is one of the most exciting fields in artificial intelligence (AI) right now. It’s still early days, but DL will completely change the lithography and photomask industry to automate or optimize the efficiency of equipment and processes. The key element required for building applied DL is a GPU-accelerated simulation environment. In this paper, we will present a Deep Learning Kit (DLK), an artificial intelligence platform that allows semiconductor manufacturing companies and mask shops to do such simulations for DL training, and show a case study with DLK. DLK provides accurate physical models for masks and lithography that are fully accelerated by CUDA on GPUs, the de facto DL training platform, a GPU accelerated Computational Design Platform (CDP), fully integrated and distributed TensorFlowTM on CDP, and pre-trained neural network models for wafer and mask problems. Using DLK, semiconductor manufacturing companies and mask shops can quickly build their deep neural network model, connect the simulator of their choice (either provided by D2S or its partners), and train the neural network model in that environment to learn a desired behavior.
It has been known for quite a long time that the best possible process window obtainable for 193i layers is by using ILT correction. These are typically converted (at a high runtime cost) to Manhattan masks both for reasons of mask manufacturability as well as computational efficiency; OPC M3D and rigorous simulators have, until very recently, been optimized for speed to run primarily with Manhattan shapes. We have recently shown that the insertion into production of multibeam mask writers makes writing curvilinear masks possible and that it is preferable to move toward a completely curvilinear paradigm, not only because the ILT is better, but because the mask manufactured will have reduced variability. Recent studies have shown a similar need for ILT-style corrections for EUV, mainly due to more complex thick mask effects. We extend the work using Monte-Carlo methods for mask variability to show that EUV layers more strongly require curvilinear approaches to mask writing in order to minimize the wafer PV bands due to both the tighter overall tolerances combined with the smaller wavelength (13.5 vs, 193) which transfers mask defects to wafer over smaller lengthscales.
With both 193i multiple patterning and EUV technologies, the constraints on the mask manufacturability are becoming increasingly stringent. The necessity for understanding curvilinear shapes implicitly in design (for ILT and EUV) or OPC correction (corner-rounding effects) along with new multi-beam mask writing systems mean the mask manufacturers are at an inflection point: whether the mask shapes are described as curvilinear targets or complex rectilinear targets, the actual mask shapes after exposure are curvilinear and must be accounted for correctly for wafer lithography. We present a GPU-accelerated intrinsically curvilinear mask data preparation system, compatible with both VSB and multi-beam systems, that is capable of full-ship simultaneous shape and dose correction using arbitrary (non-Gaussian) kernels for model shape and dose effects.