Protein crystallisation is critical for structural biology yet finding the right chemical conditions which support crystal growth remains challenging and time-consuming. One key obstacle is the limited understanding about the underlying chemical search space. The similarity of chemical conditions refers to the extent which two different conditions will yield comparable effects on the crystallisation process and therefore yield predictable outcomes. This does not refer to merely the chemical similarity, but to the biophysical effect that the chemicals exert on the crystallisation process. Similarity is a critical property for developing a protein crystallisation search strategy. However, beyond the use of simple heuristics it was previously unknown whether the outcome of two conditions would be similar or not. In this work, we learn a model which characterises the protein crystallisation search space and learns the similarity between conditions by capturing the similarities across experimental outcomes. Our model is trained using triplets, each composed of two successful crystallisation conditions, expected to be similar, and one unsuccessful condition, expected to be dissimilar. By leveraging deep metric learning, we teach the model to cluster conditions that are more likely to lead to similar outcome. This approach creates a feature space where outcome-similar conditions are closely grouped. Our model successfully learns to identify patterns of crystallisation condition similarity and significantly improves that ability to recognise co-hits compared to previous methods, even for proteins that were unseen during training. Our results indicate that our model can be used inform condition selection in new crystallisation trials, offering a tool for researchers to search the experimental design space more efficiently.