ADM1 itself is well specified and freely implemented; what stops people using it is that a digester feed has to become a long vector of state variables before anything runs. Turning a written description into that vector, then offering a charge-balance check on the result, is where the time goes. Worth setting up deliberately: three reactor slots means you can run configurations side by side rather than one at a time, and the inhibition analysis is what tells you why a run underperformed instead of just that it did.
A wastewater process modelling server built on the Anaerobic Digestion Model No. 1, the international standard for anaerobic digester simulation. The hard part of ADM1 is normally translating a real feedstock into dozens of state variables; here you describe the waste in plain language and the server derives them, then runs the simulation and interprets the result.
- Feedstock definition from a written description — either state variables alone, or state variables together with kinetic parameters tuned to that feedstock — `describe_feedstock`, `describe_kinetics`
- Simulation setup as separate, checkable steps: influent flow rate with simulation time and time step, then per-reactor temperature, hydraulic retention time and integration method, for up to three reactor configurations — `set_flow_parameters`, `set_reactor_parameters`
- The simulation itself, run against whatever parameters are currently set — `run_simulation_tool`
- Stream analysis for the influent, any of the three effluents, or any of the three biogas streams — `get_stream_properties`
- Process diagnosis rather than raw output: inhibition factors with optimization recommendations, biomass yields and efficiency, and a nutrient-balance check on the carbon, nitrogen and phosphorus ratios — `get_inhibition_analysis`, `get_biomass_yields`, `check_nutrient_balance`
- A thermodynamic consistency check on the feedstock you just defined, before you spend a simulation on it — `validate_feedstock_charge_balance`
- Parameter access and a reset, so a run can be adjusted in place or started clean — `get_parameter`, `set_parameter`, `reset_simulation`
- Report generation with charts and methodology sections, built from the actual simulation data — `generate_report`
- A system prompt shipped in the README that walks a model through the correct order of operations, including which of the two feedstock tools to call and never both
A Google API key for the language-model step that turns a feedstock description into ADM1 parameters, set as `GOOGLE_API_KEY` in a `.env` file. Python 3.8 or higher. Clone the repository, create a virtual environment and install from requirements — QSDsan, which carries the ADM1 implementation, comes with the dependencies. The client entry points at `server.py` using the virtual environment's own python, and the README sets `MCP_TIMEOUT` to 600000 because simulations are not fast.