<p>The pervasive employment of machine learning (ML) in experimental physics raises questions about its epistemic impact. While ML techniques are accelerating scientific progress at an unprecedented pace, the concern is that they may limit or even impede scientific understanding, and that, therefore, they should be regarded as mere instrumental tools. Drawing on a case study from accelerator physics where ML techniques successfully optimize beam intensity, this paper argues that ML-based experimental strategies can be epistemically salient with respect to understanding. In particular, they can provide ‘attributive understanding’, a genuine form of scientific understanding that is experiment-driven and centered on grasping robust modal relations among experimental variables. This conclusion not only challenges the view that ML techniques cannot provide scientific understanding, but also offers new insights into the epistemology of experimentation in general, urging a reconsideration of the role of ML in experimental physics.</p>