NICOLET turns up in almost every academic reference on lettuce yield prediction, and increasingly inside the forecasting modules operators are being sold. It is a good model. It is also routinely deployed in ways that guarantee it will be wrong, and the failure is rarely in the math — it is in the parameters, the inputs, and the conversion step at the end.
NICOLET B3 is a nonlinear dynamic lettuce model that predicts dry and fresh matter from light, temperature, and CO2. In practice it drifts for three reasons: parameters calibrated in someone else’s greenhouse are never re-fit locally, LED spectrum is not represented in the model, and dry-to-fresh weight conversion is assumed rather than measured against harvest data.
What NICOLET B3 Actually Models
NICOLET belongs to a family of mechanistic carbon-balance models. The Seginet-lineage formulation divides the plant into virtual compartments — “structure,” “vacuole,” and “excess-carbon,” with soluble organic and mineral compounds inside the vacuole complementing each other to hold a constant total osmotic pressure. That last point matters more than it sounds: it is the mechanism that lets the model say something about nitrate content, not just mass.
The closely related Van Henten formulation that most software implements tracks two state variables, non-structural and structural dry weight, driven by incident photosynthetically active radiation, temperature, and CO2. Three inputs, two states, a handful of ODEs. That parsimony is why it runs on a controller instead of a GPU cluster.
It works when it is calibrated. A 2022 evaluation in the Spanish Journal of Agricultural Research reported R² of 0.9939 for fresh matter and 0.9858 for dry matter after calibration against greenhouse data, found that temperature had the most obvious impact on model performance compared with PAR and CO2, and noted the model could perform well with only two inputs.
Read that carefully, because it contains the warning. Those fits came after parameter calibration against local experimental data. Sensitivity work on NICOLET B3 published in Agrociencia identified three parameters with major effect on model behavior: the maintenance respiration coefficient, the radiation extinction coefficient through the canopy, and the uninhibited growth rate coefficient of a closed canopy. Those three are exactly the parameters that change when you swap cultivar, plant density, or rack geometry.
Why Does a Calibrated NICOLET Model Drift in a Vertical Farm?
Because the published parameter sets were fit under sunlight, in greenhouses, at field-like plant spacing — and a vertical farm violates all three conditions at once.
Spectrum is invisible to the model. NICOLET sees PAR as a scalar. A Michigan State and Virginia Tech group working with Erik Runkle’s lab showed that with an unvarying light-use efficiency coefficient, the dynamic growth model did not accurately predict dry mass across light treatments outside typical greenhouse conditions, and that fitting a spectrum-dependent coefficient brought prediction error down to roughly 6–8% for simple first- and second-order forms and about 2% average error for a fourth-order form. Same model, same crop; the difference was letting the efficiency term move with spectrum.
Canopy geometry is not greenhouse geometry. The radiation extinction coefficient encodes how light attenuates through the canopy. At vertical-farm densities with a fixed fixture height and no diffuse sky component, extinction behaves differently. Leaving the literature value in place systematically biases the back half of the crop cycle, when canopy closure dominates.
Microclimate is not room climate. The model is driven by canopy conditions. Most operators feed it a room sensor two meters from the plants. On an eight-tier rack, the top tier and the bottom tier are not running the same temperature or VPD, and temperature is the input the model is most sensitive to.
The model has no opinion about tipburn. NICOLET predicts biomass, not marketability. The Cornell CEA lettuce handbook pairs its 17 mol/m²/d DLI target with downward airflow over the plants, and notes that without that airflow lettuce has to be grown at reduced light — for example 12 instead of 17 mol/m²/d — which slows growth. A model can forecast a beautiful 160 g head that your QC line downgrades.
Which Model Fits the Sensors You Actually Have?
Model selection should follow instrumentation, not prestige. Running a nitrate-aware formulation on a farm with no tissue testing produces precision you cannot verify.
| Approach | Minimum inputs | Predicts well | Breaks when |
|---|---|---|---|
| Van Henten two-state | PAR, temperature, CO2 | Dry mass trajectory under stable climate | Spectrum or density changes without re-fitting |
| NICOLET B3 | PAR, temperature, CO2 (nitrogen status optional) | Dry and fresh matter; nitrate behavior via vacuole balance | Literature parameters are used uncalibrated |
| VPD-aware dynamic models | Above plus RH/VPD at canopy | Crops where humidity limits transpiration and expansion | Only room-level humidity is logged |
| Regression / ML yield models | Large labeled harvest history | Interpolating inside the conditions you have already run | Extrapolating to a new cultivar, recipe, or room |
Mechanistic models extrapolate; statistical models interpolate. If you are commissioning a new room, you need the former. This is the same tradeoff covered in Digital Twin Controlled Environment Agriculture: What It Actually Takes.
The Fresh-Weight Conversion Is Where the Forecast Usually Dies
NICOLET’s core currency is carbon. Your buyer’s currency is fresh weight in a clamshell. The conversion between them is a dry matter content assumption, and it is the least scrutinized number in most deployments.
Use Cornell’s benchmark as the anchor: maintaining 17 mol/m²/d produced a marketable head of 150 g fresh weight with nearly 7 g of dry weight at 35 days after seeding — roughly 4.7% dry matter. Now run the arithmetic on a model that nails dry mass at 7 g but carries the wrong conversion. At 5.5% dry matter the forecast reads 127 g; at 4.0% it reads 175 g. The biology was predicted correctly and the pack-out forecast is off by a third across that range. Dry matter content is not a constant: it moves with DLI, temperature, cultivar, and harvest time of day.
The osmotic side of NICOLET has a commercial edge too. Because vacuolar nitrate and carbohydrate substitute for one another, low-light production tends toward higher nitrate. If you ship into the EU, Commission Regulation (EU) No 1258/2011 caps nitrate in lettuce grown under cover at 5,000 mg NO3/kg for product harvested 1 October to 31 March and 4,000 mg/kg from 1 April to 30 September. A model that tracks the vacuole is telling you something about compliance risk, not just yield.
The fix is unglamorous: weigh and dry a sample at every harvest, log it against the batch’s actual climate record, and re-fit. A model without a calibration loop is a spreadsheet with ODEs in it.
How AGEYE Approaches This
CultivAid Predict (predict.ageyetech.com) is AGEYE’s agricultural digital twin and prediction engine, and it carries six crop ODE models rather than one: NICOLET B3 lettuce, Van Henten lettuce, Xu-Du VPD-aware lettuce, TOMGRO tomato, SIMPLE universal, and logistic herb. Model choice is handled by sensor-aware routing that classifies incoming streams as full sensors, basic sensors, or climate optimization, then selects the best model per crop — so a farm with canopy-level VPD data is not forced onto the same formulation as a farm logging temperature only.
The calibration endpoint records harvest actuals and tracks MAPE against prior predictions, which is what converts a static literature model into one that reflects your room. Predictions carry parametric confidence bounds and anomaly detection, and a scenario endpoint lets you test what-if condition changes before committing a recipe. The platform supports 16 crop varieties across lettuces, herbs, brassicas, tomatoes, strawberry, and microgreens, plus a multi-agent farm simulation with 6 agents and 18 action types for operational planning.
What This Means
Lettuce modeling is converging on hybrid designs: mechanistic cores for extrapolation, with empirical correction layers for spectrum, humidity, and specific leaf area that the original 1990s formulations never had to handle. The published work pointing that direction already exists. What most farms lack is not a better equation — it is the harvest-actuals discipline that lets any equation be corrected. Operators who instrument canopy-level conditions and log dry matter at every harvest will get useful forecasts from a thirty-year-old model. Operators who do not will get confident numbers from whatever they buy next, and the numbers will still be wrong. Forecast accuracy is a direct input to the labor and pack-out planning discussed in How to Lower Cost Per Pound in Indoor Farming: The Four Levers That Matter.
If you want to check whether your actual light environment matches what your growth model assumes, start with AGEYE’s free DLI Calculator.
