The United States has over 85,000 metric tons of commercial spent nuclear fuel, growing at a rate of 2,000 tons per year. As per the Nuclear Waste Policy Act, the federal government is responsible for disposing of this waste in a permanent geologic repository. However, such a facility has yet to be licensed. As a result, the government has paid commercial nuclear power plant operators over $9 billion dollars for costs incurred by storing spent fuel at reactor sites. These costs will continue to increase until a permanent solution is found. Crucial to evaluating potential repository concepts is the post-closure performance assessment, which models the release and transport of radionuclides through the subsurface to ensure that members of the public will not be exposed to significant radiation. State-of-the-art performance assessment codes such as PFLOTRAN couple modular “process models” to simulate the various physical processes (e.g. flow, transport, reaction) that control migration of radionuclides from the spent fuel to the biosphere. Due to the large timescales inherent in modeling long-lived fission products, these simulations are computationally expensive. Additionally, limited knowledge of subsurface parameters necessitates repeated simulation to account for model uncertainties. To address this issue, Illinois Rocstar LLC is developing NucBench, a data-driven system for producing reduced order models (ROMs) to predict radionuclide transport in geologic repositories for nuclear waste disposal applications. The system utilizes i) proper orthogonal decomposition (POD) to perform physics-based feature extraction from the results of high-fidelity simulations in PFLOTRAN and ii) long short-term memory (LSTM) neural networks to predict and parameterize evolution of the extracted features over time. POD achieves dimensional reduction by computing an orthonormal basis that efficiently spans the solution space with as few basis vectors as possible. The full order results are projected onto the POD basis, and this dimensionally-reduced representation is used to train an LSTM network to learn the system dynamics and dependence on model input parameters. The final result is a parametrized ROM that allows many-query applications such as uncertainty quantification and sensitivity analysis to be performed with significantly reduced computational cost. This method is applied to model the disposal of commercial spent nuclear fuel in a repository located within a shale rock layer. This enhanced modeling capability will support the performance assessment of nuclear waste repositories as new and better geologic disposal concepts are explored and ultimately implemented.