AVS 71 Session CPS+MS1-MoM: Metrology for Semiconductor Manufacturing

Hafnium and zirconium oxide based thin films deposited by atomic layer deposition (ALD) are used as dielectric layers in advanced semiconductor devices. These films can also be stabilized in a ferroelectric phase for applications in memory, logic, and synaptic devices. ALD typically produces small-grained polycrystalline films containing a mixture of ferroelectric and non-ferroelectric phases with varying crystallographic orientations. Routine characterization of these films is critical for the research, development, and manufacturing of next-generation devices. While X-ray diffraction (XRD) is widely used for phase identification, it is limited to large-area, unpatterned thin films. Electron microscopy-based methods, in contrast, enable site-specific characterization within device structures, where local phase distributions may differ from blanket film samples.

This presentation discusses automated analysis of four-dimensional scanning transmission electron microscopy (4D-STEM) precession electron diffraction (PED) datasets for hafnium zirconium oxide (HZO) thin films in TiN/HZO/TiN capacitor structures. STEM lamellae are often thicker than the average HZO grain size, resulting in dynamical diffraction contributions from multiple grains at many probe positions. Additionally, distinguishing between HZO crystal phases is challenging due to small differences in lattice parameters and the potential presence of multiple orthorhombic polymorphs, making automated phase mapping particularly difficult. PED offers advantages over nanobeam electron diffraction (NBED) for phase and orientation analysis, and we find that PED is necessary for reliable automated template matching in HZO diffraction data.

Although automated phase and orientation mapping of HZO films using 4D-STEM has been previously demonstrated, a detailed assessment of different analysis methods has been lacking. Here, we compare results from a commercial software package (NanoMEGAS ASTAR) with an open-source framework (py4DSTEM). Correlation between automated phase maps and electrical verification of ferroelectricity confirms the identification of the non-centrosymmetric orthorhombic space group 29 phase of HZO.

Atom probe tomography (APT) is a sensitive analytical tool capable of providing 3D atomic reconstructions with isotopically resolved elemental and sub-nm spatial resolution. In the field of semiconductor manufacturing, it is used for failure analysis, process development, and competitive engineering. The highly complex, multi-element heterogenous structures of modern electronic devices require APT tools capable of quantitative analysis of samples containing layers with different optical, thermal, and electrical properties. Current state-of-the-art commercial APT instruments use a deep ultraviolet (DUV) light source with a wavelength of 257 nm (4.8 eV photon energy) to trigger the process of field ion evaporation from the samples under study. Historical progression of commercial APT instruments has shown that using shorter wavelength light sources results in improvements in sample survivability, especially for heterogenous samples, which is crucial for the study of semiconductor devices. In support of the semiconductor manufacturing industry, our group proposed and was tasked with the development of an APT instrument that employs a variety of wavelength sources. At NIST we are designing a beamline capable of delivering several different wavelengths (515 nm, 343 nm, 257 nm, and 206 nm) to the APT sample with similar beam parameters. This new multiple wavelength APT instrument will enable a careful study of the benefits and potential drawbacks of each triggering wavelength, with special emphasis on sample survivability, background signal, and mass resolving power. Our mission is to explore the APT wavelength parameter space using different triggering light sources to identify the ideal wavelength, or combination of wavelengths, best suited to improve the study of complex heterogenous semiconductor devices with APT.

The next big wave of innovation in microelectronics will likely stems from the successful integration of heterogeneous materials and devices into new 3D constructs, driven by improvements in yields, energy efficiency and costs. Engineering chips for this novel architecture requires precise knowledge of the thermal conductivity and interfacial thermal conductance of thin films and interfaces, but current thermal metrology is inadequate, especially in terms of spatial resolution and throughput. For example, time-domain thermo-reflectance (TDTR), measures thermal conductivity, by reconstructing the sample thermalization as a function of the probe delay time and fitting the sample decay rate to a thermal model. However, TDTR is not well adapted to mapping applications because of low spatial resolution (> 1 μm) and long measurement times (≈120 s/pixel). Furthermore, TDTR typically requires coating the sample with a metallic transducer layer to increase the signal to noise, which is undesirable.

Photothermal induced resonance (PTIR), also known-as AFM-IR, by the combination of atomic force microscopy (AFM) with IR spectroscopy, is an emergent technique that yield IR spectra and maps with nanoscale resolution using the AFM tip to bypass the light diffraction limit. However, while commercially available AFM probes transduce the sample photothermal expansion, they lack the sensitivity and bandwidth required to measure the fast sample thermalization directly.

Here I will introduce, a wide bandwidth (WD) version of PTIR, pioneered at NIST, that enables imaging of thermal conductivity (η), interfacial thermal conductance (G) and chemical composition at the nano scale, concurrently and with high throughput (20 ms/pixel). This new measurement paradigm leverages custom optomechanical AFM cantilevers with very low detection-noise (≈1 fm/Hz^1/2) over a wide (>125 MHz) bandwidth. Thanks to these characteristics the entire, time-domain, thermal expansion and contraction of a sample is measured with high spatial (≈10 nm) and temporal (≈ 4 ns) resolutions and at once, rather than as a function of pump-probe delay as in TDTR. Fitting the time-domain thermalization of the sample to a thermal model enable mapping η and G at the nanoscale.

Compared to TDTR, WB-PTIR is ≈6000x faster and does not require a transducer layer. Such WB-PTIR measurements are particularly well adapted for measuring samples with low to moderate thermal conductivities, like many packaging materials, and excel in measuring interfacial thermal conductance at the nanoscale with high precision (e.g., ΔG = 2-5% for SU-8 on ZnSe).