adm1-numba

adm1-numba is an Anaerobic Digestion Model No. 1 (ADM1) reactor simulation package using Numba and numbalsoda.

ADM1 is a dynamic model for anaerobic digestion introduced by Batstone et al. in the original IWA Scientific and Technical Report No. 13: Anaerobic Digestion Model No. 1 (ADM1). It describes biochemical conversion, acid-base chemistry, inhibition, and gas-liquid transfer in anaerobic digesters. adm1-numba implementation closely follows the work Aspects on ADM1 Implementation within the BSM2 Framework by Rosen, C. and Jeppsson, U.

Existing Python ADM1 implementations such as pyADM1 are convenient but can be slow and may not use the latest implementation choices. ADM1F provides a fast implementation, but it requires a C++/PETSc-based installation, which can add deployment effort. adm1-numba keeps the interface simple for Python users while accelerating the model with Numba and numbalsoda. On default steady-state run results, it shows minimal differences compared with ADM1F while remaining easy to install and integrate into Python workflows.

This package is designed for repeated model calls in control, optimization, and reinforcement learning (RL) applications. See our publication using this implementation.

Installation

Install with:

pip install adm1-numba

Quickstart

from adm1_numba import DEFAULT_F0, DEFAULT_U0, run_stage

sol, qch4, qco2, ph = run_stage(
    DEFAULT_U0.copy(),
    deltaT=10.,
    tint=1e-2,
    f0=DEFAULT_F0.copy(),
)

print(sol.shape)
print(qco2[-1])

Use .copy() when passing default arrays if you plan to mutate input values.

Direct ODE Integration

run_stage is a wrapper around numbalsoda.lsoda. You can also integrate the ADM1 ODE model directly with the exported function pointer:

import numpy as np
from numbalsoda import lsoda

from adm1_numba import DEFAULT_F0, DEFAULT_U0, calculate_observables, funcptr

t_eval = np.arange(0.0, 1.0, 1e-3)
sol, success = lsoda(
    funcptr,
    DEFAULT_U0.copy(),
    t_eval,
    DEFAULT_F0.copy(),
    rtol=1e-7,
    atol=1e-9,
)

if not success:
    raise RuntimeError("ADM1 integration failed")

qch4, qco2, ph = calculate_observables(sol)

print(sol.shape)
print(qch4[-1], qco2[-1], ph[-1])

Multi-stage Control Example

The example below runs multiple stages. A dummy agent changes the input flow rate Q with random exploration noise, and each stage starts from the final state of the previous stage.

import numpy as np

from adm1_numba import DEFAULT_F0, DEFAULT_U0, run_stage

class DummyAgent:
    def __init__(self, baseline_q: float = 170, noise_scale: float = 10.0):
        self.baseline_q = baseline_q
        self.noise_scale = noise_scale

    def act(self, stage: int, rng: np.random.Generator) -> float:
        periodic_signal = 15.0 * np.sin(stage / 5.0)
        exploration_noise = rng.normal(0.0, self.noise_scale)
        return max(1.0, self.baseline_q + periodic_signal + exploration_noise)

Q_index = 23
rng = np.random.default_rng(1)
agent = DummyAgent(baseline_q=DEFAULT_F0[Q_index])

u = DEFAULT_U0.copy()
f = DEFAULT_F0.copy()
history = []

for stage in range(30):
    f[Q_index] = agent.act(stage, rng)

    sol, qch4, qco2, ph = run_stage(
        u,
        deltaT=1.0,
        f0=f,
    )

    u = sol[-1].copy()
    history.append(
        {
            "stage": stage,
            "Q": f[Q_index],
            "Qch4": qch4[-1],
            "Qco2": qco2[-1],
            "pH": ph[-1],
        }
    )

print(history[-1])

Citation

To cite this repository:

@article{gao2024reinforcement,
	title = {Reinforcement learning-based control for waste biorefining processes under uncertainty},
	author = {Gao, Ji and Wahlen, Abigael and Ju, Caleb and Chen, Yongsheng and Lan, Guanghui and Tong, Zhaohui},
	journal = {Communications Engineering},
	year = {2024},
	volume = {3},
	number = {1},
	pages = {38},
	doi = {10.1038/s44172-024-00183-7},
}