HAADF ‐ STEM and EDS tilt‐series simulations of 25x25x25 nm semiconductors

Abstract

In recent years significant progress has been made in moving towards a more quantitative analysis of electron microscope images. Computational simulations have played a large part in this evolution, allowing insights into the image formation process and informing the optimization of experimental acquisitions [1‐3]. Recent innovations in X‐ray detector technology have prompted a renewal of interest in EDS mapping by substantially improving signal to noise ratios and offering commensurate reductions in acquisition times [4]. To continue the trend in quantitative analysis, it is vital for computational simulations to incorporate this resurging modality. Here, a new simulation program is introduced which extends the well‐founded multislice protocol to include the ionization of atoms by the electron beam, and the subsequent generation of characteristic X‐rays. In addition to describing the formulation of the multislice simulation program, this presentation will describe how the program is being used on the Dutch national supercomputer to address challenges faced by the semiconductor fabrication industry. The ongoing miniaturisation of computer chips is putting an increasing strain on fabrication techniques, resulting in greater occurrences of manufacturing defects. In order to increase yields, a technique is needed which is capable of identifying and characterizing these defects. Typically, the defects have dimensions on the order of a nanometre, making HAADF‐STEM one of the few viable options. Whilst the atomic number sensitivity of HAADF‐STEM can be used to differentiate some elements, modern semiconductor devices make use of electronically disparate elements with similar atomic numbers that cannot be readily distinguished by HAADF‐STEM alone. For example, the chemical sensitivity of EDS is required in cases such as the substitution of the high dielectric constant element hafnium (Z=72), with a typical gate metal, tantalum (Z=73). Modern finFET transistors have complex 3‐dimensional structures, so defect detection must be performed in the framework of tilt‐series tomography. To meet the needs of industry, the multi‐ modal reconstruction and analysis procedure must be accurate, robust, and fast. The starting point is to develop an algorithm that accurately reconstructs the non‐linear images produced using HAADF‐ STEM+EDS. To achieve this, a tomographic dataset is required for which the true nature of the specimen is well defined. This can only be achieved through simulation, and requires substantial computational cost. This presentation will describe how two such datasets have been produced, in which tilt series simulations of a 30x30x30 nm region of a finFET device have been calculated. The finFET device consists of a crystalline silicon fin with a thin oxide layer at the surface, coated with a 20 nm thick amorphous hafnium dioxide layer. On top of this is a gate metal layer of 20 nm of amorphous tantalum. The remaining volume is filled by polycrystalline titanium aluminium nitride. The first tilt‐series features an ideal device, whilst the second includes roughening at interfaces, pinhole defects in the dielectric layer, and a 7 nm carbon nanoparticle contaminant. Each dataset consists of 179 projections in 2 degree increments (no missing wedge) with 8 elemental maps and a number of annular detector geometries. This large calculation was made possible through the use of both multiple CPUs and multiple GPUs. The construction of the model shown in figure 1 will be described as will the computational techniques that were employed to simulate the tomographic projections. An example of a HAADF‐STEM image and EDS maps from one projection can be seen in figures 2‐4.