Exploring the tabular foundation model TabPFN for performance map prediction of variable-speed heat pumps and compressors
Division Heat and Buildings
Energy and AI | Volume 25, September 2026 | 100871
Sahil Vishram Vadadkar, Beatrice Rodenbücher, Michael Kropp, Katharina Morawietz, Hans-Martin Henning, Manuel Lämmle, Andreas Velte-Schäfer
This study demonstrates the application of TabPFN, a transformer-based Tabular Foundation Model, to predict performance maps of variable-speed heat pumps and compressors. Led by PhD researcher Sahil Vishram Vadadkar at the Albert-Ludwigs-Universität Freiburg, the work evaluates prediction quality in low-data regimes compared to conventional machine learning methods and commonly applied polynomial models.
Polynomial models are limited to single variants and require recalibration for each new application. Physics-based models require a large number of parameters that are often not known and difficult to obtain. In contrast, TabPFN enables knowledge transfer across different heat pump variants and manufacturers. The model achieved highly accurate performance map predictions for both heat pumps and compressors, with reliable predictions possible using as few as one or two data points from the target variant. TabPFN consistently outperformed state-of-the-art methods including XGBoost, Random Forest, and Gradient Boost.
This data efficiency will reduce measurement requirements and accelerate modelling workflows for new component variants. The results indicate that TabPFN learns inherent relationships between input features and successfully transfers them to unknown compressor or heat pump variants through group indices in training datasets. This advanced training strategy paves the way for creating a transferable "learning environment" for building energy components.
While physics-based models remain necessary for transient state analysis, this data-driven approach offers a competitive alternative for steady-state performance prediction. The methodology may prove particularly valuable in practical engineering contexts during the design phase, where extensive measurement data is unavailable or costly to obtain.