<p>Heavy metal migration in soil poses significant risks to geoenvironmental safety and human health, making risk assessment essential for contaminated site management and remediation. Soil–water characteristic curve (SWCC) and adsorption isotherm (AI) are two key relations controlling soil heavy metal migration. Uncertainties in fitting parameters of SWCC and AI and model selection uncertainties are unavoidable due to a lack of test data. It is therefore necessary to quantify these uncertainties and factor them into risk assessment of heavy metal polluted sites. This study develops a physics-informed Bayesian model averaging (BMA)-based risk assessment framework of soil heavy metal migration, illustrated through a real contaminated site example. Results show that the proposed framework can not only quantify the uncertainties in the estimated SWCC and AI, but also predict the SWCC and AI with reasonable accuracy. No single SWCC–AI model combination dominates (probabilities do not exceed 0.7), and different combinations lead to about fivefold differences in maximum predicted migration distance, which could mislead decision-making. By integrating parameter and model selection uncertainties, the framework provides a robust and model-independent estimation of soil heavy metal pollution probability, thereby reducing the decision risk of relying on a single model combination.</p>