Evolving preferential dissolution channels are common features formed during reactive fluid flow in carbonate rocks. Understanding these is of particular importance in applications involving subsurface engineered reservoirs but predicting their progression is currently challenging and poorly understood. Here, we propose a new approach to predict both the spatial distribution and extent of dissolution using a combination of experimental work, X-ray microtomography (μCT) and machine learning. We have conducted experiments, under reservoir conditions of temperature and pressure, involving pre- and post-flooding μCT characterisations, and coupled the outputs with a neural network to predict locations where carbonate was most likely to be dissolved. Our simulations demonstrate that our new solution can identify the key geometrical features that are important during dissolution, and can accurately predict the location and spread of dissolution. An important benefit of this approach is that it can accurately predict dissolution channels through forward prediction, while it does not require further chemical parameters, using instead common and accessible variables.