Producing green gas: an algorithm to optimise methanization

France has set itself the target of incorporating 10% renewable gas into its energy networks by 2030, compared with the current 2%. There is no secret to achieving this: commissioning new methanizers and optimising their operation will be the key to success. Reliable predictive models that can improve the design and operation of these systems would be invaluable allies in this endeavour. “We have models that describe how methanizers work,” explains Roman Moscoviz, Head of Research & Innovation at Suez. “But they are based on data collected and processed manually. We were lacking the statistical and algorithmic tools to achieve more robust and automated optimisation.” 

For the company, which operates around fifty methanization plants processing sewage sludge and organic waste around the world, the stakes are high: “Shutting down a plant for three months can cost €500,000,” says Roman Moscoviz. “Similarly, improving the accuracy of methanizer dimensions can result in financial gain.” So, in order to design the ideal optimisation tools, in autumn 2021 the company launched a CIFRE (Industrial Training through Research) thesis in collaboration with Benjamin Guedj, Senior Research Scientist in the Modal project team at the Inria centre at the University of Lille and Scientific Director of the Inria London programme. Antoine Picard, who at the time held a Master’s degree in statistics, agreed to take up the challenge. “I knew nothing about methanization, but I wanted to move on from pure statistics to application,” he recalls. “So I had to immerse myself in the workings of methanizers and their models. The advantage of a CIFRE thesis is that it enabled me to benefit from the observation data supplied by Suez to train and test my models, and from the expertise of professionals to evaluate the data and the results obtained.” 

Step one: managing uncertainty using PAC-Bayesian theory

The thesis aimed to define algorithms for calibrating a digital twin of a methanization plant in order to provide relevant predictions on the evolution of methanization depending on the operating mode. This laid the groundwork for optimising the process. 

First, the PhD student had to solve a major challenge related to the complexity of methanization:

“However, the models describing the stages of this digestion must be calibrated to accurately represent the behaviour of each particular methanizer, because living organisms adapt: the characteristics of a given community of microorganisms can vary greatly between two methanizers”, explains Antoine Picard. The challenge therefore lies in identifying these characteristics, while taking into account the uncertainty associated with this identification, since it will be based on a finite amount of data that is not always reliable. Field analyses and sensors installed in methanizers inevitably have a margin of error. 

Under the guidance of Benjamin Guedj, his PhD supervisor at Inria and a specialist in the field, Antoine Picard opted for a specific machine learning approach, the PAC-Bayesian theory, to overcome this first obstacle. “This approach allows us to make predictions that are highly likely to be accurate, while incorporating uncertainty modelling”, he explains. “The idea is to control prediction error on new data based on empirical error, measured on training data, and prior knowledge”. 

Step two: doing just as well, but faster, by approximating the risk

But then the young researcher ran into a second major problem: these algorithms are very resource-intensive! They learn in small steps, comparing each simulation with the observation data. The result? It takes around 80 hours to calibrate a methanization model!

Here again, Antoine Picard found a solution to this problem. “We decided that we didn’t need to assess the risk precisely at each iteration of the algorithm, but that an approximation of the risk would be sufficient,” he explains. “Our method involved making a large number of risk estimates - 600 to 800 - during the first iterations to obtain an approximation of the risk, then learning with bigger steps. Ultimately, this solution allowed us to calibrate a model 10 times faster!”

The predictions made in this way were then verified experimentally, and the results exceeded expectations. “We obtain margins of error of 3 to 10% on predictions that can be made over several months, or even several years, which in our field is excellent,” says Roman Moscoviz. 

This success is reflected in a number of ways: four scientific papers have been published or are in the process of being published, and a fifth was presented at the NeurIPS conference, a key event in AI research, in December 2024. As for the algorithm, it has been adapted to the production monitoring tool at two methanization sites operated by Suez, and is set be rolled out in four others in 2025, and then to the company’s entire portfolio. “We are prioritising sites where we see potential for improvement, and we can already see a marked difference!” says Roman Moscoviz.

Antoine Picard defended his thesis in April 2025. “This research challenged me with industrial issues and operational difficulties, while giving me the satisfaction of finding a solution,” he says. “Plus, it was partly carried out at University College London (UCL) through the Inria London programme, which gave me international exposure and the chance to meet lots of interesting researchers.” This research is fully in line with the theme of the future Genesis team led by Benjamin Guedj - a joint team project involving the Lille University Inria centre and UCL - and will continue within Suez.

Antoine Picard should be joining Roman Moscoviz’s team for the rest of the project, with a new challenge to take up: “The goal now is to implement an algorithm that takes into account data from dozens of methanization sites in order to aggregate the various established models into a meta-model. This will optimise processes in all known methanization plants, as well as those for which we have little or no data.” The aim is to improve the efficiency of processing, increase biogas production, reduce costs and make a real contribution to the targets for incorporating green gas into energy networks.