3D Chemical Mapping via P-FIB Tomography and Machine Learning

Serial sectioning and chemical mapping via energy dispersive spectroscopy (EDS) in a focused ion-beam (FIB) based on Ga can access only limited volumes of material [1]. Xe-based plasma FIBs (PFIB) can access volumes two orders of magnitude volumes greater than Ga [2] and the addition of a femtosecond laser can further increase that by three orders of magnitude more [3]. In this work we discuss the extension of 3D chemical mapping to large volumes of material in the PFIB with additional ongoing work to extend this to even larger volumes accessed with the femtosecond laser. Further we apply machine learning to effectively segment the chemical data [1], pushing such analyses to data sets in excess of 30 Gbytes. A 3D EDS spectral image data set was acquired on Thermo Fisher Scientific Helios 5 Laser PFIB from a ceramic-to-metal braze [4]. The block lift-out analysis geometry was utilized [4], where a volume of material, 60 μm by 40 μm by 25 μm, was removed from a bulk metallographic section and mounted on a Cu TEM grid as shown in the SEM image in Figure 1a and Figure 1b after clean-up of the analytical surface for 3D EDS measurements. The ceramic milled much more slowly than the metals and was removed for this proof-of-principle experiment. Future work will look at utilizing the femtosecond laser to mitigate this effect. Fig. 1b shows the sample ready for 3D EDS after the removal of the ceramic. In addition to the Cu (intermediate gray)/Ag (lightest) eutectic on the left and Fe-Co-Ni alloy on the right, there is an apparent dark intermetallic phase winding through the Cu/Ag. Additionally, there is channeling contrast in several of the phases which could make image segmentation of such an image or series of images problematic. For this purpose, we collected a full X-ray spectral image at each slice in the volume with the intent of applying Machine Learning (ML) to the full tomographic spectral image dataset as was done previously for much smaller volumes [1]. In order to maximize spatial resolution of the X-ray signal, an electron accelerating voltage of 7 kV was used. 244 sections were acquired at a spacing of 100 nm and x-y pixel sizes were ∼160 nm thus sampling the expected interaction volume. The overall data set was 512 by 352 pixels by 750 channels (10 eV/channel) by 244 slices. Stored as one-byte unsigned integers this was ∼34 Gbytes. The data were analyzed with Sandia’s Automated eXpert Spectral Image Analysis (AXSIA) machine learning code [1, 5] in ∼12 minutes on a laboratory workstation (Dell Precision 5820 Intel i7 Xeon W-2155 CPU @ 3.30GHz, 10 cores, and 64 Gb RAM). Figure 2 shows the AXSIA results where two intermetallic phases were identified, one Ni-rich Ti and another Fe-rich Ti. The electron image shows no contrast between these two phases while AXSIA has chemically segmented these phases with no operator input. [6] Block lift-out mounted to a TEM grid and b. as ion-polished for PFIB sectioning and 3D EDS. AXSIA results showing the spectral factors and the corresponding spatial factors for the first slice in the inset component image overlay.