The problem of learning an unknown function from given data, i.e., construct an approximation to the function that predicts its values away from the data is ubiquitous in modern science. There are numerous settings for this learning problem depending on what additional information is provided about the unknown function, how the accuracy is measured, what is known about the data and data sites, and whether the data observations are polluted by noise. We consider the specific case where the learning problem consists of determining the solution to a second order elliptic partial differential equation (PDE) without information on its boundary condition. The lack of sufficient information necessary to uniquely determine the targeted function is alleviated by given finitely many linear noiseless measurements. The recovery performance is measured in energy norm and for the recovery problem to be tractable, we assume that the function to be recovered belongs to a compact subset of the energy space. The latter is referred to as the model class assumption. Among all functions satisfying the measurements, the PDE, and the model class assumption, the proposed algorithm constructs an approximation of the representative with minimal norm. For each measurement, this requires the approximation of a fractional diffusion problem on the boundary of the computational domain. We present and discuss the performances of the Dunford-Taylor method employed for the space discretization. In addition, we show how the resulting recovery algorithm is near-optimal and asymptotically optimal in the limit of the vanishing space discretization error.