Machine learning for thermal energy storage in buildings: A Physics-first systematic review of sensible, latent and sorption storage

Machine learning (ML) is widely used for reduced-order modeling, prediction and control of thermal energy storage (TES) in buildings, yet the field lacks a critical synthesis linking modeling choices to storage physics. The objective of this review is to establish, quantitatively, how ML is actually used for building-scale TES and where practice diverges from the physics of the storage process. Its novelty is twofold: it is the first review to organize the TES–ML literature by storage mechanism rather than by algorithm or task, and the first to perform a cross-study quantitative synthesis of architectures, metrics and validation practice. This review critically analyzes 40 peer-reviewed studies published between 1999 and 2025, identified through a systematic, PRISMA 2020-style screening (last search update: December 2025), using a physics-first taxonomy organized by storage mechanism, i.e. sensible, latent and sorption/thermochemical, rather than by algorithm or material. Sensible and latent storage dominate (about 42% and 48% of studies); sorption and thermochemical storage remain underrepresented (about 10%). Of 38 catalogued models, 25 (66%) belong to the feedforward/multilayer-perceptron family, 10 are recurrent (RNN, LSTM, GRU, NARX), 2 use deep reinforcement learning and only 1 is physics-informed; of the 27 architectures whose depth is explicitly reported, all but two use three or fewer hidden layers. Where reported, models reproduce reference data accurately with correlation coefficients frequently above 0.99 and, in one surrogate, a speed-up from roughly 120 h to 5 s. However, performance is communicated through at least seven different metric conventions (R/R², RMSE, CVRMSE, MSE, MAE, relative error and system-level savings), only four studies report a model-fidelity metric together with a system-level outcome, and validation, dataset construction and computational baselines are reported too heterogeneously for direct comparison. The central research gap is a mismatch between the inherently transient physics of TES and the static quantities most models predict; recurrent, hybrid and physics-informed approaches that could close it remain sparsely validated. We translate these findings into a prioritized research agenda: shared benchmark problems, open datasets, a minimum reporting checklist for TES-ML studies, and an explicit mapping from storage mechanism to suitable model family.