We propose a rigorous construction of a 1D path collective variable to sample structural phase transformations in condensed matter. The path collective variable is defined in a space spanned by global collective variables that serve as classifiers derived from local structural units. A reliable identification of local structural environments is achieved by employing a neural network based classification. The 1D path collective variable is subsequently used together with enhanced sampling techniques to explore the complex migration of a phase boundary during a solid-solid phase transformation in molybdenum.Efficient sampling of high-dimensional conformational spaces represented by rough potential energy landscapes constitutes a significant challenge in the computational molecular sciences, particularly when different basins on the landscape are separated by energy barriers significantly higher than k B T . In order to address this challenge, various enhanced sampling techniques have been developed including accelerated molecular dynamics [1][2][3][4][5], transition path sampling [6-9], metadynamics [10][11][12][13], and (driven) adiabatic free energy dynamics (d-AFED) [14][15][16] or temperature accelerated molecular dynamics (TAMD) [17] and combinations of these [18,19]. In many cases the sampling, and in essentially all cases the analysis of the resulting high-dimensional free-energy landscapes, requires a projection onto a low-dimensional collective variable (CV) space. Indeed, the choice of the CVs is not always intuitive, but a meaningful representation in the low-dimensional space is crucial for capturing the correct mechanisms.Machine learning (ML) can provide a powerful approach to address the aforementioned challenges. The last decade has seen significant advances in the use of electronic structure calculations to train ML potentials for atomistic simulations capable of reaching large systems sizes and long time scales with accurate and reliable energies and forces. More recently, ML approaches have proved useful in learning high-dimensional freeenergy surfaces (FESs) [20,21], and in providing a lowdimensional set of CVs [22,23]. In such approaches, however, it is often difficult to interpret the low-dimensional CVs that emerge from the learning procedure as they emerge as abstract outputs of the ML model employed.In this letter, we overcome this difficulty by exploiting an ML model to identify local atomic structures and then using the ML output to construct a physically motivated one-dimensional CV. The latter is then employed with an enhanced configurational sampling scheme to char-acterise structural phase transformations in condensed matter. The basic idea of our approach is generally applicable to tackle different kinds of phases transformations.A structural phase transformation can be viewed as a global change of the entire system that is associated with and driven by changes in the local structural environment around each atom (or other structural building blocks such as molecules). Furthermore, f...