Artificial Intelligence and the unfashionable end of the energy system: Improving gas yields from biomethane plants

Fixing these little problems matters

Solar has scale. Wind has drama. Hydrogen has a mad optimism from its devotees bordering on theology. Biomethane, by contrast, is faintly dull. It smells a bit, requires planning permission near residential areas, and depends on an unstable biology that produces molecules rather than electrons. And yet, if the UK and Europe are serious about security of supply, resilience, and decarbonising the parts of the energy system that electrification won’t reach for decades, biomethane is not a sideshow. It has a substantial speaking role in the main act, Part II.

What is changing, quietly, incrementally, and without much fanfare, is how Artificial Intelligence (AI) is being used to improve gas yields from anaerobic digestion (AD) and biomethane plants. Not in the “AI will solve climate change” sense, but in the more interesting “AI might finally help the system run as it was supposed to” sense.

The unspoken thing of biomethane performance

It’s a bit of a secret in the industry that most plants do not operate at their theoretical optimum. Thousands of plants were built in the last decade during a rush to qualify for subsidies. At the time, getting molecules or electrons into the system was more important than optimising gas yields. There is an installed European capacity of 7,000 plus plants, often characterised by suboptimal feedstock menus, overly conservative operating regimes, poor data integration, and reactive rather than proactive maintenance. These factors combine to keep yields perhaps 10–25% below what could be achieved from the same installed capacity.

Why 15% is a big deal

Europe now has more than 1,600 operational biomethane plants injecting close to 7 billion cubic metres (bcm) per year into gas networks. The UK, by contrast, has just over 100 grid-connected plants, producing roughly 10–20 TWh annually – around 1–2 bcm of gas.

In a system where revenues are capped by subsidy limits, feedstock availability, and rising costs, gas yield is everything. A few percentage points of additional methane recovery add up to potentially more than a billion cubic metres on a European scale. This is where AI, properly applied, starts to matter.

What AI actually means in an AD context

AI in biomethane plants isn’t about radical change. It usually boils down to three rather prosaic things:

1. Collecting more data points and tracking patterns
2. Machine learning-based process and menu optimisation
3. Predictive, rather than reactive, decision-making

AD plants already generate vast volumes of data, including feedstock composition, volatile fatty acids, pH, temperature, loading rates, gas composition, parasitic energy use, etc. The problem has never been data scarcity; it has been a lack of usable insight. AI systems can identify patterns invisible to a spreadsheet and turn them into actionable guidance.

Feedstock management: where yields are won or lost

Gas yield is ultimately a feedstock problem. AD plants don’t rely on electronics or mechanics; they rely on Mother Nature, which is messy and unpredictable. No two tonnes of food waste are identical. Agricultural residues vary seasonally. Fats, oils, and greases are both literally and metaphorically slippery customers.

AI models can help overcome this complexity by creating digital twins that predict how specific feedstock blends are likely to behave in a given digester. Operators can then push the biology to the optimum mix while taking account of gate fees, carbon intensity and gas prices. This isn’t about chasing maximum output every day. It’s about knowing how far to push without upsetting the bacteria.  Often, it unlocks meaningful gains without a single new piece of hardware.

Process optimisation beyond rules of thumb

Many AD plants still operate on decades-old mental rules developed by experienced operators who smell vaguely of manure and have feedstock under their fingernails: keep temperatures steady, avoid shock loading, watch your Volatile Fatty Acids, that sort of thing.

Digital twins build on this experience by spotting early warning signs of digester stress. Subtle shifts in gas composition and micro-trends in alkalinity are predicted well before they become obvious. This enables earlier, gentler interventions. The results are happier bacteria, faster recovery, and more consistent methane production.

Predictive maintenance and parasitic losses

Yield is also about what steals energy before gas ever reaches the grid. Like my shower, compressors, pumps, and membrane units degrade slowly. AI-driven predictive maintenance can identify declining performance weeks before inefficiency shows up in monthly reports. Fixing a compressor early can deliver more gas than endlessly tweaking feedstock ratios. These are boring gains, but they are real.

Carbon intensity: the invisible constraint

AI is increasingly intersecting with carbon intensity (CI) scores, which drive certificate prices and profitability. Every megajoule of biomethane injected now carries a CI score: a quantified measure of the CO2 emissions created from feedstock collection through digestion, upgrading, and grid injection.

The policy shift from straight subsidies to CI-driven certificates has made your average plant operator’s life just a bit harder. Monoculture maize plants were much easier to run. Not easy, but easier. Today, multi-feed stock plants require a daily or even hourly multi-variable calculation of gate fees, CI scores, biological impacts and certificate prices. Choosing between expensive high-yielding maize or cheap low-yielding cow manure is hard maths to do while on the phone to an impatient farmer. AI models can track these inputs in real time to maintain the balance. They weigh the cost of feedstock against the needs of the bacteria and the impact on CI scores

Using AI to improve CI scores is subtle, nuanced and deeply unglamorous. It is precisely the kind of incremental improvement to the installed capacity of existing plants that we need to move towards the 2030 targets without rewriting the entire energy system.

Necessary, but not sufficient

Even if AI helped every plant in the UK and Europe improve yields by 10–15%, it would not solve the underlying scale problem. AI isn’t magic. It’s a quiet janitor sweeping up existing inefficiencies off the floor, one datapoint at a time.

Europe’s current biomethane output (~7 bcm per year) is in contrast to gas demand of roughly 400 bcm annually. If biomethane is to make a serious contribution, it requires higher efficiency, but more importantly, the construction of many more plants. We need builders as well as janitors. AI also plays a role here by improving reliability and revenue stability, which makes projects more bankable. Yield optimisation improves today’s plants; scaled deployment of AI helps make tomorrow’s plants investable.

The limits of AI and why they matter

Despite the hype, AI is no cure-all. It cannot compensate for poor plant design, chaotic feedstock contracts or maize projects optimised for subsidies rather than carbon intensity. Over-automation and black-box decision-making carry real risks, and it will be a while before most operator’s hand full control of their multi-million-pound biological baby to an algorithm.

There is also the issue of grid constraints, particularly for electricity plants. Increasing gas yields is pointless if you cannot squeeze the product into the grid. If the plant is grid-constrained and has a positive gate fee, there is actually no upside to increasing yields.

Notwithstanding these outliers, AI’s most interesting contribution to biomethane is making something that works work a little bit better. It offers incremental improvements, fewer bad days, less foaming, and a few more CI negative cubic metres injected from the same feedstocks.

Fixing the plumbing

This brings us back to that dodgy shower. The fix wasn’t revolutionary. There wasn’t a redesign and no new water source. It was just a matter of replacing a worn-out part and paying attention to a system everyone had taken for granted because it “sort of” worked.

Biomethane is much the same. The pipes exist. The gas grid exists. The feedstocks exist. The plants sort of work. AI helps the system stop rattling, stops plants wasting energy, and starts delivering the maximum theoretical gas yield day after day, cubic metre after cubic metre.

Sometimes, the most important work in the energy transition isn’t inventing something new. It’s tweaking the plumbing before everyone gives up on the shower.