The phrase “AI for carbon management” may sound as if it can explain a complex field setting all at once. I suggest first asking what decision the model is intended to support. Chapter 8 of Technology and Solutions for Demonstrating Carbon Reduction distinguishes what methane prediction models do from the scope in which they can be used. A model that detects faults and one that estimates emissions answer different questions.
A distinction to make first as you read the book
When evaluating alerts, look not only at what was flagged but also at anomalies that were missed. You need to check intervals without alerts too in order to estimate what may have been missed. Also consider whether the response is manageable for the staff responsible. Decide what to do when inputs fall outside the model’s scope; it may be necessary to flag limits in the output or hand the case over to a person. Remember as well that AI decision support and external verification of emissions reductions have different purposes and responsibilities.
Function checks. A bump test on a safety gas detector introduces gas to check sensor response and alarm functions; it does not replace quantitative calibration. For a research methane analyzer, perform checks suited to its purpose, such as zero and span checks, leak checks, and recovery tests following the manufacturer’s procedures. Passing a function check does not, by itself, guarantee the accuracy of emissions measurement.
Quantifying reductions requires both measurement traceability and traceability of the calculations. Link measurements to a documented calibration chain and uncertainty, and trace activity data, emission factors, and models through their sources, versions, and conditions of application to the final result. Not every reduction figure comes from direct measurement alone. Along with checking whether calculations can be reproduced, reviewers should confirm that the data and methods used are appropriate for the claim.
A realistic path for progress is to build up concentration observations and baseline emissions-rate measurements together, and link them with ventilation, weather, animal numbers, and feed information. Evaluate prediction error on farms and periods not used to develop the model, and state the scope of application. Even with more simultaneous observation data, model reliability will not improve automatically if the data lack representativeness or share common bias.
When comparing sensors, examine measurement range and resolution, cross-sensitivity to other gases, response time, and drift together. Also review environmental conditions such as temperature, humidity, and dust, along with calibration requirements. If the purpose requires distinguishing small concentration differences, check whether the device’s display units and uncertainty are fit for that purpose for each product and operating mode. The name of a sensor principle alone cannot substitute for these conditions.
AI can be used to identify candidate anomalies, estimate short missing intervals under defined conditions, and explore relationships among multiple variables. It can also help organize data, perform quality checks, and draft reports. The purpose and cost of failure differ across these functions, so do not assess them all using the same performance metric. Disclose the inputs and assumptions used to produce a result, and distinguish estimates from directly observed values.
What to look for in the field
Imagine a sensor producing a reading that differs from its usual values. AI may point to items worth checking. But that alert does not confirm equipment failure or prove emissions-reduction results. The staff member’s findings and actions need to be recorded so you can assess how the model helped with the work.
AI can organize observations and estimate missing values. Its value does not lie in hiding that a value is estimated; it lies in making clear what data and assumptions produced a value that can be trusted, and within what scope.
Separate evaluation data first. If data from adjacent times at the same farm are randomly mixed, performance that has memorized recurring patterns at that site may be mistaken for performance at a new farm. If the purpose is to predict future operation, separate by time order; if the purpose is deployment at other farms, separate by farm. Distinguish validation data used for model selection from data used for the final performance check, and apply preprocessing and missing-data imputation using rules established from the training data.
Evaluate AI materials according to their intended use. To predict future operation, consider preserving time order; to deploy at other farms, consider separating data by farm. Also separate the data used to select the model from the data reserved for final performance checks. Compare against a simple baseline model, and examine not only average error but also bias by farm and season and conditions under which the model fails. This is necessary to explain the model’s real scope of use.
Questions to ask in an investment review
In an investment review, ask what has improved compared with existing methods rather than focusing on the label “AI.” Also decide whether the goal is to issue many alerts or to miss fewer important anomalies. Considering the staff available to respond in the field and the cost of false alarms will make the usable conditions clearer.
If lineage is insufficient, reviewers may find it difficult to check how a number was produced, and acceptance of the data may be limited depending on the intended use. Set retention periods for source data and revision histories by checking the requirements of the applicable scheme, contract, and privacy rules. Do not arbitrarily overwrite records that must be retained, but do not make indefinite retention of all data the default either.
Finally, consider privacy and data rights. Farm data may be linked to information about animals, production performance, and finances. If such data are reused for finance, insurance, or marketing beyond the purpose of emissions verification, farms may lose trust. As mentioned in Chapter 1, losing trust also means losing data quality. Purpose limitation, separate consent, differentiated access rights, and a process for raising objections matter as much as technical security. Chapter 12’s contract discussion and Recommendation 12 in the policy proposals develop this design further.
A question to leave with before closing the book
Explaining AI’s role narrowly does not reduce its potential. It helps clarify what to delegate to a model and what a person needs to check. Try recording the model’s output alongside the actual decision and the result of the action taken.
About the author
Bongjun Lee graduated from the Graduate School of Management of Technology (MOT) at Sogang University. He spent seven years at Infitron Co., Ltd. and 20 years at Honeywell, working for a total of 27 years in gas measurement in service engineering, technical sales, and channel management. He is now the founder of AI Safety Korea Co., Ltd., where he is researching livestock methane monitoring and measurement, reporting, and verification (MRV) systems using sensors, AI, and data analytics.