<p>Surface roughness is an important quality characteristic of components produced by machining processes. However, effectively modeling the mapping from process parameters to surface roughness remains challenging when training data are limited, despite its importance for informed production decisions and efficient manufacturing. Existing small-sample learning paradigms, including model-agnostic meta-learning (MAML), have shown promise but often exhibit insufficient prediction accuracy and reliability when applied to manufacturing scenarios. These limitations mainly stem from two factors: overconfident outcomes caused by point-estimate parameterization and inadequate incorporation of physical consistency. In response to these challenges, this article develops a physics-guided Bayesian meta-learning (PGBML) method that integrates MAML-style meta-learning, Bayesian inference, and physical prior knowledge of surface roughness formation to improve predictive modeling. In PGBML, Bayesian inference enables uncertainty-aware parameter estimation, while the mechanistic component is incorporated as a soft physical constraint to guide learning under limited data conditions. Experiments on robotic disc grinding verify that PGBML surpasses existing methods in both prediction accuracy and reliability. Its cross-process applicability is further validated using datasets obtained from multiple machining operations, including wheel grinding and turning, as well as milling. Overall, PGBML provides a feasible solution for predicting surface roughness in industrial settings where available data are scarce.</p>