Flow channelization limits chemically enhanced lithium recovery from fractured geothermal reservoirs: a discrete fracture network reactive-transport study

Code, model inputs and simulation output supporting a study of lithium in-situ recovery from a fractured enhanced geothermal system, using three-dimensional discrete fracture network reactive-transport simulation. Scope. Fracture networks are generated and meshed with pydfnWorks; flow and multicomponent reactive transport are solved with PFLOTRAN on a purpose-built Cooper Basin lithium-silicate thermodynamic database produced with pyGCC. The model is parameterised on the Habanero reservoir of the Cooper Basin, South Australia. The domain is 500 m across at 4.2 km depth, 240 °C and 60 MPa, with a doublet at 250 m spacing and a thirteen-species reactive system in which spodumene is the lithium source. The result the package supports is that fracture-network topology, rather than brine chemistry, is the primary control on recovery. Flow is confined to a small set of well-connected fractures, so the producer samples dilute fluid while bypassed rock retains near-saturated brine. 96% of the dissolved lithium resides in cells that supply 1% of production. Cumulative production scales as the injection rate to the power +0.41 and produced concentration as the rate to the power −0.59, the two exponents satisfying the identity mass conservation requires. Chemical enhancement is limited to under 10%, because injected reagents follow the same pathways as the flow and are consumed before reaching the bypassed rock. Contents. 137 simulation run directories in 16 sets: a 5-rate injection sweep at 20 and at 50 years a 10-member network ensemble at 5 rates injection-pH and carbonated-water sweeps a 3 by 4 grid in the spodumene rate constant and reactive surface area, with a 4-point extension below it a fracture-intensity sweep a 750 m domain comparison, with its drainage-geometry control 3 discretizations of one fixed network, for the mesh-convergence test a one-dimensional column as a reference case time-step and velocity-output verification runs Each run directory holds its PFLOTRAN mass balance and input deck, from which every published value can be recomputed without rerunning the solver. Also included are the geochemical database and the scripts that build it, the network generation and meshing drivers, the analysis and figure scripts, and the job submission scripts. Three verification tools check the archive against the published values, enumerate the runs against what the paper describes, and compare results computed on two machines. Reproducibility. Every network is generated from a fixed seed, so the fracture geometry regenerates exactly and can be compared by checksum. The results were recomputed on a second cluster with an unrelated toolchain — a Spack PFLOTRAN build with OpenMPI and PETSc 3.24.5, against the Cray build and cray-mpich used for the published runs. Produced concentrations agreed to 5 significant figures. Threshold-based swept-fraction statistics differed by 2 to 6%. That is expected: they depend on whether individual cells fall above or below a flux cutoff, and a change in the rank decomposition alters the order of summation. The corresponding Gini coefficient, an integral over the whole distribution, agreed to 0.1%. Mesh convergence was established by refining one fixed network: a 6.8-fold increase in element count moves the produced concentration by 0.46%. Because the target resolution also governs which fractures the generator accepts, the test holds that resolution fixed and varies the far-field cell size, which is a meshing parameter alone. The reproduction guide sets out why this matters and how to verify that the three meshes describe the same network. Limits. The kinetic parameters are literature-default values and are not calibrated; the model is not history-matched to any Cooper Basin production, tracer or assay data, none of which exists for this undeveloped resource. The reported concentrations are therefore a conservative estimate under the assumed kinetics rather than a resource assessment. The simulations are isothermal and hold permeability constant; the reproduction guide and the paper's appendices quantify both simplifications. Full field output in HDF5 is retained for a subset of runs only, the mass balances being sufficient for every published number. Requirements. The analysis and figure scripts need only Python and are listed in requirements.txt. Rerunning the simulations additionally requires PFLOTRAN v6 built against PETSc, and dfnWorks 2.7 with LaGriT. The reproduction guide records the configuration details that are easy to get wrong, among them the MPI launch setting under Slurm and the dfnWorks executable paths. Files in this record, and how to reassemble them. The deposit is split in two so that the smaller part can be taken on its own. cooper_li_isr_non_h5.zip, 5.7 GB, holds everything except the PFLOTRAN field output: the code, the geochemical database, the meshes, the input decks, the solver logs and the mass-balance files from which every published value is computed. This file is sufficient on its own for the whole of the reproduction guide except the two figures that need per-cell fields. cooper_li_isr_h5.7z, 43.6 GB, holds the field output in HDF5 for 94 of the 104 completed runs. REPRODUCE_STEPS.md, the reproduction guide, also included inside the zip. Both archives store the same directory prefix, so extracting them into one place reassembles the original tree. unzip cooper_li_isr_non_h5.zip 7z x cooper_li_isr_h5.7z To confirm an archive is intact before extracting it, 7z t cooper_li_isr_h5.7z. On the field output. Ten runs have no HDF5 in this deposit. They are the three mesh-convergence runs, the two pH runs at the finer resolution, four earlier runs at that resolution, and one of the two realizations at the highest fracture intensity. The four earlier runs predate the correction described in Section 3.4. Every one of the ten has its mass balance, input deck and solver log in the zip, and no published value depends on their spatial fields. Getting started. REPRODUCE_STEPS.md gives the commands in the order they are run, from reading the archived output to recomputing every simulation, with the expected result and a check at each step. Steps 0 to 2 verify the published values and regenerate every figure in about 30 minutes and need no simulation software. The remaining steps rerun the simulations and take about 40 hours at 16 MPI ranks. Contact. Yongqiang Chen, WA School of Mines, Curtin University ([e-mail]), and Klaus Regenauer-Lieb, WA School of Mines and ARC Centre of Excellence for Carbon Science and Innovation, Curtin University ([e-mail]). Released under the Apache 2.0 licence.