Pipe-jacking is fast becoming the primary construction method for buried utility infrastructure. A key concern in both design and construction is that the jacking force required to advance the tunnel boring machine and pipe string may exceed the prepared-for capacity, resulting in jacking pipe/launch shaft damage and/or costly tunnel boring machine (TBM) recovery. Uncertainty in jacking forces – amplified by the influence of work stoppages – necessitates redundant intermediate shafts and/or inter-jacks at great expense [1]. Empirical methods are popular in industry but are well known to lack accuracy; whilst there is a strong desire to supplement such approaches with robust data-driven techniques, typically small construction datasets present significant challenges. To address this challenge, this paper develops a physics-constrained neural network predictive model for pipe-jacking forces. Information used as input into the model includes principal design information and soil type. Novel contributions include (a) a bespoke framework to constrain a neural network using a pipe-jacking mechanistic model which includes stoppage-induced friction increases, (b) built-in model uncertainty for greater confidence in model outputs, (c) new historical drive data for model training, and (d) one-hot encoding of soil type as a model input. The model is calibrated and validated against fourteen tunnel drives across four different sites with four distinctive ground types. The physics constrained model was found to predict jacking force to a higher accuracy than current industry practice and better discern meaningful patterns in data than a purely data driven artificial neural network. The results reveal promising performance for this initial dataset such that there is motivation to train the present approach on a more comprehensive drive database for more reliable and cost effective solutions for new projects.