Abstract Selective laser melting processes deposit and join metal powders to near net shape in a layer-by-layer manner. The process of melting and re-solidification of several layers of deposited material can result in geometric deviations, and the impact is particularly significant for sub-millimetre structures oriented at a wide range of overhang angles with respect to the building platform. This paper assesses and benchmarks the capabilities of a neural network-based geometric compensation approach for truss lattice structures with circular cross-sections. The neural network method is capable to generate free-form cross-sections with enhanced geometric freedom for compensation compared to more established analytical compensation approaches limited to predefined geometric shapes. For neural network training, lattice dome structures composed of trusses with different overhang angles were designed and printed by selective laser melting and measured via X-ray computed tomography, resulting in point cloud data sets containing more than 20,000 data points for each overhang angle. For experimental validation, neural network-compensated dome structures were benchmarked against dome structures with elliptical parameter compensation. Results show that the neural network compensated lattice trusses achieve higher printing dimensional accuracy compared to the uncompensated structures and those compensated based on elliptical parameter estimates.