Green hydrogen is costly (at best, $4 a kg); the two key drivers of costs — electrolysers and electricity — are the focus of research, but economics can’t quarrel with physics. There is only so much you can do to hammer down costs.

Or so it was thought.

A new pathway is opening to tame recalcitrant costs — artificial intelligence (AI) and its subset machine learning (ML). AI/ML can assist in each step of the production — analyse real-time operational parameters such as temperature, pressure and input electricity; predict faults; and help discover catalysts that are alternatives to costly rare materials such as platinum and iridium (a gram costs $5,000).

International Energy Agency’s ‘Green Hydrogen Review 2023’ estimated that over 50 green hydrogen projects were using AI, involving some $500 million investment. In many pilot projects, AI usage has achieved efficiency gains of 5-15 per cent; this could go higher with more operational data, notes a June 2024 paper by Ashish Saxena, Master of Technology with Amazon, Seattle. He notes that Siemens Energy incorporated ‘reinforcement learning’ (RL), a type of ML, to enable online tuning of over 70,000 parameters for their electrolysis process to achieve 8-10 per cent enhancement. “The models used equipment control variables and performance measures to come up with the best control strategies,” says Saxena, noting that falling cost of software, sensors, and IoT connectivity is making AI affordable.

Experts such as Saxena underscore the value of AI in predictive maintenance and process optimisation. When run continuously, AI models feed on real-time data from the equipment’s performance to anticipate failures.

Catalyst finder

A bigger role AI could play is in identifying affordable catalysts for hydrogen evolution reactions (HER) in electrolysers. Catalysts, which speed up a chemical reaction without being consumed in the process, help in hydrogen and oxygen evolution in electrolysers.

The performance of a catalyst is influenced by the interplay of the structure, composition, electronic configuration, adsorption site geometry, and so on of the alloying materials. Analysing the parameters of hundreds of materials to determine the right combination is a daunting task, but could be made easier by AI.

But this begs the question why catalysts for hydrogen evolution reactions (HER) thrown up by AI are not commonplace. Indeed, AI has suggested some interesting catalysts, as we will see later, but the problem is the cost and complexity of experimental data generation needed to feed the AI models.

Prof KE Vipin and Prof Prahallad Padhan of IIT-Madras are developing novel ML methodologies for predicting and designing high-performance, cost-effective intermetallic catalysts for HER and oxygen evolution reaction (OER).

They used a dataset extracted from the Catalysis-hub database, representing a significant compilation of alloy catalysis data for HER and OER. It involved 16,226 distinct data points, with 8,856 entries focused on HER catalysis and 7,370 entries dedicated to OER catalysis, they say in a paper, which is yet to be peer-reviewed. Vipin and Padhan created a “systematic, data-driven approach to catalyst design that can overcome the limitation of traditional trial-and-error methods”. Incidentally, the model is good for designing any catalyst, not just for hydrogen evolution.

Using regression algorithms such as Random Forest, XGBoost, and Support Vector Regression, the model can “accurately predict the Gibbs free energy of hydrogen adsorption on bimetallic alloy surfaces — a key descriptor for catalytic activity,” Vipin told Quantum. Gibbs free energy is a thermodynamic quantity that tells us whether a chemical process can happen spontaneously, without needing extra energy.

“While we haven’t experimentally validated new catalysts yet, we are using the trained model to screen and predict promising new bimetallic candidates,” Vipin said. These predictions are further evaluated through density functional theory simulations to confirm their potential theoretically before going for experimental testing.

AI-driven catalyst design has already led to several candidates emerging, including a ‘nickel-incorporated carbon quantum dots’ (based on a Bayesian genetic algorithm); a ruthenium–manganese–calcium–praseodymium mixed oxide catalyst; and a copper-aluminium catalyst.

First key step

It must be noted that designing a catalyst is only the first (but indispensable) step — it must cross the ‘valley of death’ between research labs and industry. It must work fine for long years and not collapse after a burst of performance. It should integrate seamlessly with existing systems.

AI is helping win this major part of the battle. Once you have systems that help increase the efficiency of electrolysers (indeed the entire value chain of green hydrogen, including renewable energy assets), and an affordable catalyst, then green hydrogen can come within the reach of commerce.

“The path to a sustainable hydrogen economy is complex and multifaceted, requiring innovative solutions that transcend disciplinary boundaries,” write Vipin and Padhan.

“Through advanced machine learning approaches, we stand at the cusp of a potential breakthrough that could revolutionise green hydrogen production, bringing us significantly closer to a decarbonised, sustainable global energy system,” they add.