Geologically informed machine learning for neutron porosity reconstruction from legacy well logs: implications for high-temperature reservoir thermal energy storage

High-temperature thermal energy storage in subsurface reservoirs (HT-RTES) can help balance seasonal heat demand and supply by storing excess thermal energy in subsurface reservoirs. Mature hydrocarbon fields are promising storage targets because their reservoir properties are relatively well constrained; however, many legacy fields lack modern petrophysical data. In the Leopoldshafen oil field (northern Upper Rhine Graben), complete log coverage is limited to legacy self-potential (SP) and short-normal (SN) resistivity logs, whereas neutron porosity (NPHI) logs are available only in two wells (LH-5 A and Stutensee-1). Here, we present an integrated workflow to reconstruct NPHI in legacy wells using machine-learning methods. SP and SN logs from 24 boreholes were digitized, standardized, and cleaned to remove acquisition artefacts. Lithological classes were derived from SP–SN cross-plot clustering and used as a rock-physics-guided lithology flag. Five supervised algorithms were tested using two input configurations: Model 1, based on scaled SP and SN logs, and Model 2, which additionally included the lithology flag. Random Forest with Model 2 yielded the best performance in the training well Stutensee-1 and generalized well to the independent validation well LH-5 A. The main prediction errors occurred in gas-bearing intervals, but a lithology-specific post-processing correction markedly improved agreement with measured NPHI. The results show that lithological information substantially improves porosity prediction from sparse legacy logs and enables geologically consistent NPHI reconstruction for reservoir characterization at Leopoldshafen.