Artificial Intelligence (AI) has been called everything from a buzzword to a game-changer. In the biogas and renewable natural gas (RNG) industry, it’s becoming more clear that AI is neither hype nor magic—it’s a tool. And like any tool, its value depends on how well it’s understood, integrated, and used.

As investors, owners, plant managers, and operators look to modernize operations, well-trained and proven AI is emerging as a strategic asset. But its success hinges on one critical factor: data.

The Data Dilemma: Why AI Needs Better Fuel

AI thrives on clean, structured, and abundant data. Tesla accumulates almost 3 billion driving miles per year of real-world data and full self-driving still struggles in poor weather and unpredictable situations. Remember the Tesla recall, when they kept crashing into stopped emergency vehicles? In contrast, most biogas plants struggle to collect a few thousand data points, which is 0.000 000 001 times as much as Tesla collects each year[L3] .

Biogas and RNG facilities—struggle with few online sensors, easy to bias near-line measurements, slow and infrequent offline tests:

• On-line coverage often uses a small number of sensors, sometimes only the temperature is monitored in the anaerobic digester, and maybe in a pre-fermenter or pasteurization process. Few monitor the pH of these processes online.

Because the main product is RNG exiting the biogas upgrader, the RNG is usually continuously monitored as part of the upgrader control system. RNG quality and quantity measurement is also required to meet grid injection requirements and for payment purposes. However, almost no one has high quality continuous reporting of raw biogas composition—which would be very helpful for tuning and optimizing the biological processes inside the digester.

The operating site’s control computer records the data from these sensors. Keeping the online sensors cleaned, calibrated, and operating well presents a significant operational challenge in the harsh AD environment.

• Near-line testing includes samples that are collected and tested on-site. These samples may be collected and tested by different plant operators who work different shifts, thereby likely introducing variability into the sampling and testing program. When feedstock or equipment related upsets and breakdowns occur, this on-site lab testing may be skipped entirely for other, more urgent, issues.
• Offline samples are those that are shipped to testing labs, usually some distance away, and the test results may not be available for several weeks. Sample collection, storage, and aging are very important factors to consider when deciding how much weight to place on any set of lab results. The feedstock and digester contents are biologically very active, as a result, the samples are changing every minute they are in the sample bottles, which is important to consider when deciding what actions to take based on those values.

The quality of data sets collected from anaerobic digesters often contain too few points and contain very human idiosyncrasies that can confuse AI systems that are being training on sparse data.

• The frequency of manual sample collection and testing varies greatly by facility. Some farm-based anaerobic digesters may only sample their digesters once a month, or even less when they are in planting or harvest season. Larger market food waste and municipal source separated organics (SSO) processing sites may sample once per day for near-line testing, and monthly or quarterly for more detailed off-site lab analysis. Very few operating sites bother with routine sample quality programs like duplicate samples or split-testing amongst different labs for their routine process-control samples.

• The manual entry spreadsheet or simple data base is the most common data tracking tool after all the testing is done. Often, smaller site operators just keep a running hand-written list on a sheet of paper in the lab. Data recording and storage for these laboratory test results introduces another opportunity for error and inconsistent data collection.
• Inconsistent data formats are common across digester feeding, online control system, and off-line lab data storage databases. We have seen issues as simple as a European technology vendor listing the date as DD/MM/YY while our clients in the US use MM/DD/YY. This makes it hard to know if a result from 01/08/25 is from January, winter operations, or August and represents operations during warm summer weather.

BTW, in my mind the answer is obvious–really, BOTH formats are wrong. They are ambiguous so at high risk of error and therefore poor communication tools. Because at Azura we work with clients all across North America and overseas, I try to enforce that we be unambiguous and use letters for the month and 4 digits for the year, don’t care if the client prefers 03 JAN 2025 or Jan 03 2025. Some even like to date sort large data sets by year so they prefer 2025 JAN 03.

• Without operational context, all the data collected can be nearly useless. A fulsome understanding of the operational context of the samples should address questions like: Was a mixer broken, switched off, or running 25% of the time when a particular sample was collected? When was a change made to run a mixer continuously by switching to HAND mode?  How would stirring up long-settled solids alter sample quality? Was an opportunistic feedstock recently added that may have introduced some inhibitory cleaning chemicals? Or was everything operating per the original site design, at nameplate capacity, and at steady-state conditions for the past 4 weeks? (ha ha, has that ever happened?)

These data limitations make it difficult for AI models to learn reliably or make accurate predictions. Without this foundational data infrastructure, AI risks becoming an expensive experiment rather than a transformative solution.

Where AI Is Already Delivering Value

Despite the many challenges with data quantity and quality, AI is proving its worth in several high-impact areas in the water and wastewater infrastructure spaces and those of us in the organic waste to energy area can learn from these adjacent fields.

Predictive Maintenance

AI can analyze data from high quality sensors such as vibration or temperature sensors on motors and other specific equipment components like seals and bearings. Because there are billions of electric motors and various types of rotating equipment in the world, there is a tremendous amount of data readily available to train machine-learning systems.

Currently, an experienced plant operator may detect a looming problem with a pump or valve during their daily rounds. The operator uses sight, smells, and sounds to notice if anything seems different or unusual that day—the same way you might notice something isn’t right with your washing machine or your car because it starts making a funny sound when you brake. An AI system that monitors pump vibration every second of the day would also be able to spot the change in the vibration pattern that forecasts equipment failures before they happen. How much better is the AI at detecting an upcoming failure than a human operator and is that knowledge worth the cost of the sensors and the bots to monitor them?

Early detection of operating issues with equipment reduces downtime and extends asset life, but deploying leading edge tools over hiring experienced operators is ultimately a business decision.

Anomaly Detection

Modern AI tools are the advanced extensions of pattern recognition engines from the 1990s and use advanced mathematical models that grew out of the neural-network innovation around that timeframe. AI tools are excellent pattern recognition devices after they have been adequately trained and tested.

In addition to motor vibrations, a trained AI can flag unusual patterns in gas composition or flow rates, possibly enabling early intervention before disruptions occur. Detecting problems developing long before they become apparent to human operators allows plant managers to schedule maintenance and repairs and avoid unscheduled downtime that can be very costly!

Where AI Isn’t the Right Tool—Yet

AI isn’t a fit for every scenario. It struggles in:

• Low-data environments: Plants lacking sufficient instrumentation or historical records.
• Highly variable feedstock: Without metadata like which clients are dumping feedstock that day? What’s in the truck? How is the feedstock source performing and what is their waste consistency? Without that data, AI can’t correlate inputs with outputs.
• Regulatory decision-making: AI can assist but shouldn’t replace human oversight. Currently AI cannot swear an oath or provide expert or even basic testimony—humans need to own this loop.
• Economic modeling: Financial decisions still require human judgment and market analysis.

In these cases, rule-based systems or operator heuristics remain more reliable.

Process Optimization

Because of the generally poor-quality and very limited data available to train AI tools on process optimization of waste-to-energy biological process, they typically lag far-behind human experts and are not expected to catch up until after the online testing tools improve dramatically. For this to happen, the cost profile of the digester needs to also improve.

For example, on-line bench-top NMR spectroscopy is a tool used, with an expert operator, to optimize bioprocess systems used to produce expensive pharmaceuticals. These costly tools do not make sense for anaerobic digestion sites. Instead, a human operator can instantly recognize when the digestate smells ‘off’ long before the off-site testing indicates increased concentrations of propionic and butyric acid are building up in the digester.

The experienced operator is also continuously collecting information using all their senses, this is data is not yet coded into AI databases. For example, the operator, when going about their daily duties, is likely aware of the incoming feedstock truck traffic—is it more or less than usual? Does the feedstock look or smell different than it did last week or last month? Did the truck driver report something odd was happening at the food factory where the feedstock was originating—they didn’t know what was happening there but knew there was a sudden call for 20 extra trucks to haul away waste?

Unexpected Situations

If a certain event or upset only occurs once every 5 years, like the failure of a flare pilot flame, how do we get a few thousand repetitions to train an AI to detect this event, let alone respond?

In these situations, an operator who is well trained on the risks can make informed decisions in a crisis. They can also know when they are out of their depth and need to call help – something a control program struggles with.

Safety Matters!

Have AI cars stopped crashing yet?

Humans can perform real-time risk assessment and emergency response. In many risky situations, human operators can instantly sense when something’s off. It could be a strange vibration, an unusual sound, a minor earth shaking tremble from a tank failure, an odd smell, or a subtle shift in pressure. AI relies on point-sensors and thresholds measuring very specific parameters, but humans continuously take in data though all our senses and compare that to our years of lived experience at a site, in a room, or working with specific equipment.

In a dynamic safety event, like a leak, flare malfunction, or a tank overpressure, human operators are able to improvise under conditions of incomplete information and high-stakes decisions.

Physical safety protocols like essential Lockout/Tagout (LOTO) and other physical safety protocols have to be done in the field and cannot be done by instrumentation or by an automated control system. Safety procedures require hands-on verification, accountability, and trust. Only trained personnel can ensure equipment is truly safe to work on. AI can suggest steps, comb through manuals, remind staff of procedures, and present e-forms for work permits and sign-offs, but it can’t physically secure a site or confirm zero energy states.

Human eyes and ears are expert at detecting hazards in unstructured natural environments. In a workplace this translates into a cluttered plant floor, a corroded pipe, a dripping seal, or a distracted coworker. These are hazards that human eyes and instincts catch immediately. AI vision systems struggle with unpredictable, messy environments where context matters.

Wrapping it up

Artificial Intelligence and machine learning is not a silver bullet for biogas and RNG operations—it’s a tool, and like any tool, its effectiveness depends entirely on the quality of the data it’s fed and the expertise guiding its use.

The reality is, most facilities are still operating in a data desert, with sparse, inconsistent, and often unreliable inputs. Without a robust data infrastructure and deep operational context, AI risks becoming just another shiny object—expensive, impressive-looking, but ultimately disconnected from the messy realities of plant life.

That’s why human expertise remains irreplaceable. From interpreting ambiguous lab results to making judgment calls during equipment failures and safety events, seasoned operators and engineers bring the nuance, adaptability, and scientific rigor that AI simply can’t replicate.

At Azura, we’ve seen firsthand how real results come not from chasing hype, but from combining high-quality data with deep domain knowledge and sound engineering. AI has a role to play—but only when it’s working alongside the people who truly understand the biology, chemistry, and chaos of anaerobic digestion and the messy business world of organic waste management.

[p.s. This would have been a lot faster to write if I just let the bots churn out some quick-turn generic crap. Hope you found this useful. Systems thinking, bioprocess systems and modelling, infrastructure cybersecurity, regulatory compliance, safety, ethics, scalability, data ownership and security, transparency and trust, root cause and breakdown analysis, and expert witness sleuthing—these are all areas I think about during my morning workouts. Thanks for reading and let me know topics you’d be interested in reading about. Thank you. -Dave]

---

Get Insights from the Bioprocess Experts Right to Your inbox