Methodology
How the system works
As fresh produce ripens and ages, its gas exchange changes in measurable ways. Low-cost sensors in a storage chamber can detect those changes, and a model trained on them can forecast remaining shelf-life early enough for an operator to change what they do with a lot.
Why this approach
What existing methods miss
Current shelf-life assessment in cold-chain operations relies on visual inspection, fixed-date labelling, or destructive sampling. Each method has the same structural problem: it measures the produce at a point in time rather than tracking the biochemical process as it unfolds. Visual inspection is subjective and misses early-stage senescence. Fixed-date labels are calibrated to average conditions, not the actual storage environment a specific lot has experienced. Destructive sampling is accurate but slow, costly, and cannot be applied continuously. Gas-exchange monitoring addresses all three limitations. The signal is continuous, non-destructive, and directly tied to the biochemical process driving decay — not a proxy for it.
Visual inspection
Subjective; detects decay only after it is visible
Fixed-date labelling
Calibrated to average conditions; ignores actual storage history
Destructive sampling
Accurate but slow, costly, and cannot run continuously
Gas-exchange monitoring (this approach)
Continuous, non-destructive, and tied directly to the biochemical process
The method
Five steps from signal to decision
Measure
CO₂ and ethylene-related VOC signals are logged continuously in storage chambers at partner sites. Sensors run passively alongside normal cold-chain operations — no changes to storage conditions, handling procedures, or lot management are required during the measurement phase.
Label
Observed shelf-life and quality outcomes are recorded as ground truth by site partners. These observations — when a lot was moved, discounted, or disposed of, and the condition it was in — form the training and validation dataset for the model.
Model
A machine-learning model is trained on the paired sensor and ground-truth dataset. The model learns to identify the gas-exchange patterns associated with different stages of senescence and to forecast remaining shelf-life from the current gas signature of a lot in storage.
Validate
Forecasts are tested against held-out lots and partner-run trials. Validation criteria — the accuracy thresholds the model must meet to be considered operationally useful — are agreed with research partners before data collection begins.
Decide
Outputs are framed as actions: move a lot, adjust atmosphere, discount now. The system does not produce a raw number — it produces a recommendation grounded in the operator's own historical outcomes for that produce type in that storage environment.
Sensor technology
What is measured, how, and why
The sensor array targets two classes of gas signal: CO₂ concentration and ethylene-related volatile organic compounds (VOCs). Both are well-characterised in the post-harvest biology literature as indicators of senescence progression.
CO₂
Cellular respiration accelerates as produce senesces, producing a measurable rise in CO₂ output. The rate and pattern of that rise varies by produce type and storage condition, and carries predictive information about the stage of senescence.
Primary signal for non-climacteric produce (berries); secondary signal for climacteric produce
Ethylene-related VOCs
Climacteric produce — apples, avocados, bananas — undergoes a pronounced ethylene burst at the onset of ripening. Ethylene-related VOCs are detectable before the burst peaks, providing an early-warning signal ahead of the rapid senescence phase.
Primary signal for climacteric produce; onset detectable before visible ripening begins
Hardware deployment: Sensors are housed in a compact unit designed for installation inside an existing cold-storage chamber without modification to the chamber itself. The unit logs data continuously and transmits it to the modelling layer. Installation at a partner site is intended to take less than a day and to require no specialist equipment or permanent infrastructure changes.
The model
Architecture, training, and outputs
The modelling layer takes the time-series gas-exchange data from the sensor array and maps it to a shelf-life forecast. The model is trained per produce type — the gas-exchange dynamics of apples and avocados are sufficiently different that a single cross-produce model would sacrifice accuracy for generality.
Time-series CO₂ and VOC readings from the sensor array, timestamped and associated with a specific lot and storage chamber.
Paired sensor readings and ground-truth shelf-life observations from partner sites. Training data is collected during the validation phase; the model is not pre-trained on synthetic or laboratory data.
A remaining shelf-life estimate for the lot, expressed as a range rather than a point estimate, together with a recommended action (hold, move, discount, dispose).
The model is calibrated to the specific storage conditions and produce-handling practices of each partner site. A model trained on one site's data is not assumed to generalise to another site without recalibration.
Validation protocol
How forecast accuracy is assessed
Validation is the central output of the current programme phase. The validation protocol is designed to answer a specific question: does the model's shelf-life forecast, produced from gas-exchange data alone, agree with the ground-truth shelf-life outcome for the same lot to within an operationally useful margin?
Validation criteria — the accuracy thresholds the model must meet — are agreed with research partners before data collection begins, not set retrospectively.
Forecasts are tested against held-out lots: lots whose sensor data was not used in model training, to assess generalisation rather than in-sample fit.
Partner-run trials provide an independent assessment of forecast accuracy under real operational conditions, separate from the model development team's own evaluation.
Results are reported against the pre-agreed criteria. Where the model does not meet a criterion, the shortfall is documented and informs the next iteration of the model or sensor specification.
Limitations and open questions
What the programme is designed to answer
The programme is at the pre-validation stage. The following are the principal open questions the validation phase is designed to address. They are stated here because institutional partners and funders are better served by an honest account of what is not yet known than by a premature claim of proven performance.
Does the gas-exchange signal carry sufficient predictive information to forecast shelf-life with operationally useful accuracy across the four target produce types?
How much does forecast accuracy vary across storage environments, chamber sizes, and lot densities? The model is calibrated per site, but the degree of site-to-site variation in the underlying signal is not yet characterised.
What is the minimum sensor deployment duration required to produce a reliable forecast for each produce type? Early-stage forecasting (before the ethylene burst or CO₂ rise is well-established) may carry higher uncertainty.
How does the model perform on mixed-lot storage, where produce at different stages of senescence is held in the same chamber?
Interested in the methodology?
We welcome technical enquiries from post-harvest scientists, food technologists, and cold-chain operators. We are happy to share more detail on the sensor specification, model architecture, or validation protocol design.
research@scifruvegai.com