There is a growing need for accurate and interpretable machine learning models of thermal comfort in buildings. Physics-informed machine learning could address this need by adding physical consistency to such models. This paper presents metamodeling of thermal comfort in non-air-conditioned buildings using physics-informed machine learning. The studied metamodel incorporated knowledge of both quasi-steady-state heat transfer and dynamic simulation results. Adaptive thermal comfort in an office located in cold and hot European climates was studied with the number of overheating hours as index. A one-at-a-time method was used to gain knowledge from dynamic simulation with TRNSYS software. This knowledge was used to filter the training data and to choose probability distributions for metamodel forms alternative to polynomial. The response of the dynamic model was positively skewed; and thus, the symmetric logistic and hyperbolic secant distributions were inappropriate and outperformed by positively skewed distributions. Incorporating physical knowledge into the metamodel was much more effective than doubling the size of the training sample. The highly flexible Kumaraswamy distribution provided the best performance with R² equal to 0.9994 for the cold climate and 0.9975 for the hot climate. Physics-informed machine learning could combine the strength of both physics and machine learning models, and could therefore support building design with flexible, accurate and interpretable metamodels.