Improved detector resolution can aid in the identification and non-destructive assay (NDA) of radioisotopes, which is crucial for nuclear safeguards applications. This improvement would be particularly useful in separating closely-spaced characteristic photopeaks within spectra used for enrichment measurements. A data-driven approach was developed using unsupervised machine learning to cluster segments of an H3D M400 pixelated Cadmium Zinc Telluride (CZT) detector. The candidate clusters were ranked by their resolvability, defined as the square root efficiency divided by the Full Width at Half Max (FWHM), to optimally trade off a modest amount of detector efficiency for large improvements in peak resolution. The unsupervised model was fitted using data collected from long-dwell (64 hour) measurements of a 100 𝜇Ci Eu-154 source placed 30 cm away from the front face of the H3D M400 CZT detector. The resulting model can then be applied to spectra outside the training dataset. In one example, a model was applied to spectra obtained from uranium standard measurements from various enrichments at Lawrence Berkeley National Laboratory, demonstrating that the model can generalize to newly seen data from different radioactive sources. Ongoing work will continue to quantify spectral improvement for safeguards-relevant measurement scenarios. This data-driven approach offers a real-time algorithmic solution to improve gamma-ray spectrometry in pixelated CZT detectors. In the future, the model will be accessible to external stakeholders, such as the IAEA, via a python software package, to allow inspectors to select the desired resolvability improvement while conducting measurements in the field.