Machine learning force fields based on local parametrization of dispersion interactions: Application to the phase diagram of <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:msub><mml:mi mathvariant="normal">C</mml:mi><mml:mn>60</mml:mn></mml:msub></mml:math>

We present a comprehensive methodology to enable addition of van der Waals (vdW) corrections machine learning (ML) atomistic force fields. Using Gaussian approximation potential (GAP) [Bart\'ok et al., Phys. Rev. Lett. 104, 136403 (2010)] as baseline, we accurately learn local model atomic polarizabilities based on Hirshfeld volume partitioning the charge density [Tkatchenko and Scheffler, 102, 073005 (2009)]. These environment-dependent are then used parametrize screened London-dispersion vdW interactions. Our ML only needs implicitly, by reference volumes from functional theory (DFT). In practice, can predict accurate knowledge environment (atomic positions) alone, making highly computationally efficient. For additional efficiency, our reuses same many-body descriptors for underlying GAP bonded interatomic also show how method enables straightforward computation gradients observables, even when these remain challenging (e.g., calculating in DFT). Finally, demonstrate approach studying phase diagram C$_{60}$, where effects important. The need vdW-inclusive reactive field is highlighted modeling decomposition C$_{60}$ molecules taking place at high pressures temperatures.