Answers are grounded in this paper and cite their sources.
In atom probe tomography (APT), the path to knowledge can be divided into three main operations: specimen preparation, data acquisition, and data analysis [1]. Over the decades, significant advancements in both hardware and software have reduced the time required for these operations from days or weeks to mere minutes, depending on the maturity of the application and the frequency of performing the same analysis. The obstacles to knowledge vary greatly depending on the time distribution in each major step. For new applications, such as a finFET in a modern microelectronics device, specimen preparation may initially dominate the timeline. However, with each iteration, the time required for specimen preparation can be significantly reduced. Subsequently, optimizing data acquisition conditions to maximize success rates and data quality can become the primary focus. Once high-quality data is collected, data analysis can then become the most time-consuming and expertise-driven step. Identifying peaks in the mass spectrum (commonly called “ranging”) and conducting three-dimensional data analysis are predominantly manual processes. The identification and ranging of mass-spectral peaks significantly influence the results of all subsequent analyses. Previous round robin experiments have demonstrated that subjective user bias in mass spectral analysis can greatly affect the results, limiting accuracy and precision [2]. Ranging and phase identification are often iterative processes, requiring adjustments and revisions to previous steps. CAMECA is collaborating with groups worldwide to develop new tools for peak identification, ranging, and volume segmentation to create a faster and more reproducible workflow using the Atom Probe Suite™ 6.3.3 platform that includes spatial ranging and an interactive extensions platform that can incorporate third-party algorithms. This workflow, summarized in Figure 1, includes automated ranging using a supervised machine learning method that recognizes peaks and determines their boundaries as shown in Figure 2. The next step involves leveraging natural isotopic abundance patterns to identify elemental and molecular peaks through model training. Confidence scores are reported to allow manual review of peaks of uncertain identity. Finally, multivariate statistical analysis methods are used to identify unique spectral components throughout the volume, leading to volume-segmented ranging and compositional analysis in unique phases identified by these components [3,4]. In the last step, automated ranging can be re-applied to each of the independent identified phases which typically increased confidence in peak assignments due to fewer overlaps in the independent phases and less background noise. The AP Suite 6.3.3 spatial ranging and extensions platform has allowed independent algorithm development to automatically range, identify peaks, and identify spectrally unique regions of the 3D data. Each of these steps takes just seconds and removes user bias (and fatigue etc.) from peak ranging and identification leading to at least more precision and likely more accuracy in data analysis. The machine learning workflow will allow retraining based on new datasets and the training data will be available for transparency. Automated workflow starting from the full mass spectrum, automated ranging and peak ID for the whole dataset. Automated spectral component analysis using MVSA followed by final ranging and peak identification for each identified volume for precise and rapid compositional analysis. Automated peak ranging based on supervised machine learning from ‘expert’ ranged datasets.
No takes yet. Share an insight, caveat, or question.
Ulfig et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: