Neural Network solutions for Anaerobic Digestion

Description

Anaerobic digestion is a sequence of processes via which microorganisms biodegrade organic matter into stable compounds and methane-rich biogas in the absence of oxygen. Anaerobic digestion is well-established in wastewater treatment, waste management and agriculture as a predictable source of renewable energy and to produce low-carbon-footprint fertilizers.

Despite anaerobic digestion’s obvious opportunities and benefits, performance deterioration and failure caused by unpredictable variations in loading rates and environmental conditions are still common and contribute to increasing start-up and running costs. At the same time, co-digestion of sludges coming from different sources has been widely shown to improve process stability and increase methane yield by 25-400% compared to digesting single-sourced sludges.

A number of laboratory tests [1] are available to understand and mitigate anaerobic digestion’s operational risks and maximise co-digestion’s potential. However, such investigations are costly and lengthy, and require specific expertise.

Biokinetic modelling vastly enhances the predictive power of laboratory tests – and in some instances, can replace them. Anaerobic Digestion Model No. 1 (ADM1) is the standard biokinetic model because of its generality [2]. However, this generality comes with the price of having 67 state variables, 45 initial conditions, 28 influent variables and 100 model parameters, most of which are not experimentally measurable. As such, the procedure of calibrating ADM1 to determine the model parameters remains complex and application-related – in fact, an art – still requiring a variety of laboratory tests, and expertise beyond the capabilities of most operators.

Thanks to their potential to extract higher-level abstract features from raw input data and use them to make predictions, neural networks would be capable – in theory – of simplifying the complexities involved in calibrating and using ADM1. However, current neural-network-based approaches suffer from two major limitations: (i) a lack of generality, as single models only reproduce very specific laboratory experiments or single digesters; and (ii) a paucity of training data.

Dapelo and Catenacci [3] recently proposed a new approach, consisting of: (a) generating a large artificial dataset through systematic application of ADM1 over carefully controlled random variations of ADM1 state variables and parameters; and (b) train a neural network over the artificial dataset, to predict the ADM1 parameters, given a generic set of state variables. This overcomes the above-mentioned paucity of data, provides a general neural-network-based method, and greatly simplifies the task of calibrating ADM1.

Candidates will extend Dapelo and Catenacci’s [3] method from steady-state to transient. The work will include re-generating the synthetic dataset and implementing an extended neural network. The novel model will allow predicting the inhibition constants (which are not possible to predict in the steady state [3]), and improved accuracy in general. ADM1 standard extension will be considered for a larger set of inhibiting factors.

Students will have access to the University of Liverpool’s High-Performance-Computing facility for neural network training. Throughout the PhD, they will have the opportunity to build an expertise in neural network development and acquire skills in C++, Python and TensorFlow.

Applicant Eligibility

Candidates will have, or be due to obtain, a master’s degree or equivalent from a reputable University in an appropriate field. Exceptional candidates with a First Class bachelor’s degree in an appropriate field will also be considered.

Application Process

Candidates wishing to apply should complete the University of Liverpool application form [How to apply for a PhD - University of Liverpool] applying for a PhD in Civil Engineering and uploading: Degree Certificates & Transcripts, an up-to-date CV, a covering letter/personal statement and two academic references.

We want all of our staff and Students to feel that Liverpool is an inclusive and welcoming environment that actively celebrates and encourages diversity. We are committed to working with students to make all reasonable project adaptations including supporting those with caring responsibilities, disabilities or other personal circumstances. For example, If you have a disability you may be entitled to a Disabled Students Allowance on top of your studentship to help cover the costs of any additional support that a person studying for a doctorate might need as a result. 

We believe everyone deserves an excellent education and encourage students from all backgrounds and personal circumstances to apply.

Enquiries

Candidates wishing to discuss the research project should contact the primary supervisor Davide Dapelo, those wishing to discuss the application process should discuss this with the School PGR Office [soepgr@liverpool.ac.uk].

Availability

Open to UK applicants

Funding information

Funded studentship

The EPSRC funded Studentship will cover full tuition fees of £4,786 per year and pay a maintenance grant for 4 years, starting at the UKRI minimum of £19,237 pa. for 2024-2025. The Studentship also comes with access to additional funding in the form of a research training support grant which is available to fund conference attendance, fieldwork, internships etc.

Supervisors

References

  1. van Loosdrecht, Mark C. M., Nielsen, Per H., Lopez-Vazquez, Carlos M. & Brdjanovic, Damir. Experimental methods in wastewater treatment. (IWA Publishing, 2016).
  2. Batstone, D. J. et al. The IWA Anaerobic Digestion Model No 1 (ADM1). Water Sci. Technol. 45, 65–73 (2002).
  3. Dapelo, D., Cantenacci, A. A Neural Network for ADM1 Calibration. Water Sci. Technol. (under review).