A Multi-Resolution Physics-Informed Neural Network Framework for Sustainable Assessment and Remediation of Hydrocarbon-Contaminated Soils: A Small-Sample Study at Kuwait’s Al-Ahmadi Field

The 1991 Gulf War contaminated more than 49 km 2 of Kuwaiti desert with hydrocarbon spills, a persistent threat to soil resources, infrastructure and the United Nations Sustainable Development Goals embedded in Kuwait Vision 2035. Managing these legacy lands calls for predictive tools that capture spatial variability while remaining computationally tractable and statistically defensible at the small sample sizes typical of post-conflict monitoring. This study develops a multi-resolution physics-informed neural network that combines wavelet-based parameter encoding, scale-dependent regularisation and a progressive upsampling training protocol. The framework is evaluated on nine trial-pit observations at a single depth of 30 cm in the Al-Ahmadi field, where the contaminated pits show a mean internal friction angle of 26.8° compared with 36.0° at co-located control pits sampled at the same time. Generalisation is assessed by leave-one-out cross-validation across the nine locations. The framework attains a friction-angle root-mean-square error of 1.29°. Under the same data and compute budget, ordinary kriging and a standard physics-informed neural network remain statistically competitive. This outcome indicates that the physics residual acts as a mass-conservation-consistent smoothness regulariser rather than a site-calibrated transport predictor. A multi-objective remediation workflow produces a cost-versus-residual-risk Pareto front for a scenario-specific 1–2 km 2 case, presented as an illustrative decision-support envelope pending external pilot calibration. A projected pathway from these outcomes to six Sustainable Development Goals and two pillars of Kuwait Vision 2035 is also discussed; quantitative attribution at this sample size is beyond scope. The small-sample, single-depth and single-locality limitations that bound the admissible inference are stated explicitly.