Equations that need solved,
in order to provide better food for more people.

In plain termsPlants don't eat on a fixed schedule; feeding them to their actual daily demand saves a lot of fertilizer and water.

The scienceFarms dose a fixed recipe and top up to target Electrical Conductivity (EC), oversupplying some ions and undersupplying others as daily demand shifts with stage and climate; published demand-matched dosing saves up to ~40% fertilizer at equal yield and up to 6x water efficiency.

In plain termsWe can't yet watch every nutrient in the water in real time, and definitely not on a small farm.

The scienceNo sensor today reads nitrate, K, Ca, Mg, P, Na and micronutrients continuously in-line at farm-grade accuracy for a small grower.

In plain termsThe affordable sensors that read individual nutrients drift and foul as they sit in the water; right now keeping them honest means manual cleaning and recalibration, because nothing corrects them automatically.

The scienceIon-selective electrodes, the only affordable per-ion option, drift, react to interfering ions, and foul over a crop cycle, with no turn-key compensation keeping them accurate without recalibration.

In plain termsThe meter says 'salt is fine' while the plant is starving for one nutrient and drowning in another.

The scienceEC and pH measure only total salt, not which ion, so holding a target EC drifts into imbalance as fast-uptake N/P/K deplete while slow Ca/Mg/S/Na/Cl accumulate.

In plain termsWe can 'see' some nutrients with cheap light sensors, but not the ones that matter most in real water.

The scienceCheap optical chips (Ultraviolet-Visible-Near-Infrared (UV-Vis-NIR)) can infer some ions without wet chemistry but fail on P/Mg/S in real, changing solution matrices.

In plain termsMinerals build up in the water: calcium forms scale that steals nutrients and clogs pipes, and salt quietly piles up; we only notice after it's already happened.

The scienceCalcium and sodium compounds build up in a closed loop: calcium-phosphate/carbonate/sulfate scale strips Ca and P out of solution and clogs flow, while sodium and chloride accumulate with every top-up, and growers only react after it forms or forces a blind discharge.

In plain termsRoots breathe from a thin film of water, and the oxygen runs out before it reaches the end of the channel.

The scienceIn Nutrient Film Technique (NFT) the thin film is the only oxygen source and Dissolved Oxygen (DO) falls from ~7 mg/L inlet to ~4 mg/L outlet; what matters is oxygen at the root surface inside the matted film, not the reservoir.

In plain termsKeeping the water cool eats most of our power, and the chiller runs on/off instead of smartly.

The scienceHolding 18-22 C means running a chiller that is often 60-80% of system power with dumb on/off control.

In plain termsPlants can poison the water for the next crop, even when the nutrient test looks perfect.

The scienceRecirculating solutions accumulate allelochemical exudates that suppress the next crop (~30% lettuce yield loss) even when nutrients test fine.

In plain termsAlgae steals nutrients and oxygen and comes back fast after cleaning; we want to catch it the moment it starts.

The scienceAlgae consumes nutrients and Dissolved Oxygen (DO) and seeds pathogen biofilm; light exclusion is the proven near-free fix but blooms return within a week of any cleaning.

In plain termsEvery way to clean the water has a side effect, and nothing tells us which to use and how much.

The scienceEvery disinfection option has a documented downside: UV is blocked by organics and breaks iron chelates, ozone oxidizes micronutrients, and peroxide risks phytotoxicity, yet no standard model picks and doses the treatment for a given water.

In plain termsRoot rot kills from below, out of sight, days before any leaf shows a symptom.

The sciencePythium and Fusarium kill from the hidden root zone days before foliar symptoms, so a top-down camera sees nothing until it's too late.

In plain termsCameras can spot a nutrient shortage only after the plant looks sick; we need to see it days earlier.

The scienceRGB models classify late, already-visible symptoms but can't flag N/P/K shortage days before symptoms appear, when intervention still saves yield.

In plain termsThe two nutrients behind tipburn leave almost no early trace in any camera, so we can't see the problem coming.

The scienceCalcium and magnesium, the drivers of tipburn and leaf disorders, have almost no early spectral signature in Red-Green-Blue (RGB) or hyperspectral (R2 0.12-0.34 vs 0.60-0.88 for Nitrogen, Phosphorus, Potassium, and Sulfur (NPKS)).

In plain termsThe cameras that see problems earliest are out of reach and need too much computing power for a farm.

The scienceHyperspectral gives the early-warning signal but is 10-100x more demanding than a Red-Green-Blue (RGB) camera and needs server compute; the 'cheap camera, spectral richness, on-device' triangle is unclosed.

In plain termsA model trained on one variety or one room gets confused on the next; every farm has to start over.

The scienceA stress model tuned to one cultivar and room drops to ~80% accuracy or below on another, so every deployment becomes a re-collection project.

In plain termsBacteria hide inside leaves before any visible spot; a better camera sees them days earlier.

The scienceBacterial pathogens colonize leaves internally before sub-millimeter lesions appear; multispectral catches soft-rot 4-8 days post-inoculation vs 20-24 days for Red-Green-Blue (RGB).

In plain termsViruses spread silently for weeks before any sign, and there's no easy way to catch them early.

The scienceViruses replicate for weeks before color change and spread mechanically in a shared Nutrient Film Technique (NFT); per-virus hyperspectral works (90%+) but generic cheap 'a virus is present' detection across types does not exist.

In plain termsAphids and thrips are tiny and hide under leaves; by the time we see them, they're already a problem.

The science0.5-3 mm pests hide on leaf undersides and in canopy shadows, so whole-tray cameras see little until populations are economic threats.

In plain termsHeat and drought stress show up in leaf temperature before the plant looks bad; we can catch it days early.

The scienceTipburn, heat and drought show up first as canopy temperature or localized tip necrosis; thermal + Red-Green-Blue (RGB) fusion already detects water stress up to 84 hours early.

In plain termsThe public plant photos don't look like real farm canopies, so models trained on them fail in the real world.

The scienceThe big public plant-disease datasets are single detached leaves on plain backgrounds, not whole-canopy greens under Light-Emitting Diodes (LEDs), so they don't transfer (65.7% on real-world benchmark vs 99% on clean).

In plain termsArtificial Intelligence (AI) can tell us a plant is stressed but not why; we need it to point at the actual cause.

The scienceModels say 'stressed' but not 'the Vapor Pressure Deficit (VPD) spike Tuesday caused it'; fusion boosts detection accuracy, but trustworthy cause-and-effect needs designed experiments and causal methods.

In plain termsRoots tell us almost everything about plant health, and they're the part we never get to see.

The scienceRoots are the canary for nearly every disease yet are the least-observed tissue: white, matted, low-contrast, hidden in channels.

In plain termsWe don't even know what 'healthy roots' looks like at the microbe level, so we can't steer toward it.

The scienceNo validated reference profile exists that defines a healthy hydroponic root-zone microbial community (expected taxa, diversity metrics, pathogen-suppression markers), so growers have no quantitative baseline to monitor or steer toward, and candidate beneficial microbes cannot be evaluated against a standard.

In plain termsThe 'good bacteria' products were made for soil and usually don't survive in recirculating water.

The scienceThere is no reliable inoculation protocol for Bacillus, Trichoderma, and Plant Growth-Promoting Rhizobacteria (PGPR) in recirculating NFT; all three evolved for soil, often fail to establish in recirculating water, and some marketed beneficials measurably depress growth.

In plain termsDifferent tests give wildly different answers, so nobody knows the exact level at which to act.

The scienceDetection methods disagree by orders of magnitude (plating ~0.05 Colony-Forming Units (CFU)/mL vs bait 20-26 CFU/L vs quantitative polymerase chain reaction (qPCR)), so no actionable 'act now' concentration exists.

In plain termsDNA tests can see everything in the water but take days, too slow when disease spreads in hours.

The scienceEnvironmental DNA sequencing reads the whole loop community but is too slow for action: amplicon runs take 2-13 days and metagenomics weeks, while a Pythium outbreak reaches every tier in days; qPCR is faster but only detects known targets.

In plain termsSlime grows inside the channels, hides pathogens, and we can't see it while the system is running.

The scienceBiofilms on channel walls and drains harbor pathogens and consume dissolved oxygen, yet no practical sensor measures biofilm load in situ inside a working channel without stopping flow or disassembling the tier.

In plain termsA blocked drain or dead pump can kill roots in hours, and the leaves won't show it until it's too late.

The scienceA clogged drain or failed pump stops the film, roots dry in hours, and damage is done before any shoot symptom.

In plain termsOne shared water loop means one bad entry infects everything, and we can't trace where it came from.

The scienceOne shared reservoir and manifold means a single pathogen introduction reaches every tier, yet no field tool reconstructs where it entered or how fast it reached each tier, so growers cannot isolate the entry point before the whole loop is infected.

In plain termsWe control the whole room like it's one plant, but each shelf has different needs.

The scienceClimate and lighting are controlled from a single room-level sensor average, but conditions differ measurably by tier and position along the rack; no published study at small-farm scale quantifies what per-tier control actually buys, so growers pay room-level prices for tier-level outcomes.

In plain termsWe know light changes flavor and nutrition, but the exact recipe is still guesswork.

The scienceThere is no turn-key recipe for which Ultraviolet (UV)-A/far-red/blue:red mix moves a target compound, and published effects contradict across dose, background and cultivar.

In plain termsRunning lights during off-peak hours can cut energy use by over half, but nobody sells the controller for it.

The scienceSavings of 13-65% from shifting or dimming lighting to cheap hours are proven in simulation, but no turn-key controller exists for a small rack.

In plain termsThe air-moisture number we steer by is only as honest as cheap sensors that drift.

The scienceVPD is only as good as its temperature and humidity inputs, and cheap sensors drift and compound error worst at the high-humidity end where condensation risk lives.

In plain termsAdding CO2 helps only if the room is sealed, and sealing can take more than the CO2 gives back.

The scienceCO2 only helps inside a sealed room, but sealing fights free-air cooling and raises the dehumid load, and getting this wrong can wipe out 12x more value than it saves in energy.

In plain termsThe top shelf bakes while the bottom stays damp, and fixing one makes the other worse.

The scienceStacked racks create a real vertical gradient: top tier hot/dry under Light-Emitting Diodes (LEDs), bottom cool/wet, and airflow that fixes temperature can wreck humidity removal.

In plain termsPlants pump out moisture that eats more than half our climate energy, and most dehumidifiers waste it.

The scienceIn a transpiring room the latent load is often 50%+ of Heating, Ventilation, and Air Conditioning (HVAC) energy, and reheat/condensing strategies routinely waste it.

In plain termsPlants change the humidity minute to minute, but our climate system runs on last week's assumptions.

The sciencePlants are the fastest-changing humidity source, but no farm measures transpiration live (lysimeters expensive, stem gauges invasive), so HVAC runs on static assumptions.

In plain termsArtificial Intelligence (AI) can run the whole farm better than fixed rules in simulation. Making it work on a real rack is the hard part.

The scienceMPC/RL beats fixed setpoints in simulation but nothing ships at small scale: no calibrated digital twin to plan against, sensors drift, and models don't transfer across facilities.

In plain termsIndoor farming uses a fixed amount of energy per pound that scale alone can't fix; it's physics.

The sciencekWh/kg is set by fixtures, crop and physics (vertical ~10-18 vs greenhouse ~5.4 kWh/kg) and does not improve with scale; this meta-problem killed 14 Controlled Environment Agriculture (CEA) firms.

In plain termsWe want to promise a customer 'X pounds on Y date, within this range' - and actually keep it.

The scienceThis is the real commercial need: not just predicting a harvest date, but a calibrated forecast, a schedule that uses it, and a rule for who absorbs the shortfall. The pieces exist separately (harvest-date error around 2.4 days, honest confidence intervals) but no system combines them.

In plain termsA camera can weigh a plant in the lab, but on our stacked racks the same camera is off by up to 10x.

The scienceWeighing plants with a camera works in the lab (it tracks true weight closely), but falls apart in a working farm: leaves overlap on stacked racks, cameras cannot hold a fixed position, and each variety throws the calculation off, pushing error from about 2 grams in the lab to 17-25 grams on the rack.

In plain termsThe classic growth model exists, but every new variety or room means cutting plants up to re-calibrate it.

The scienceThe best-known public plant-growth model (NiCoLet, available since 1998) does not transfer to a new variety or a new room without re-measuring plants by cutting them up, so every new crop or space starts a fresh calibration campaign.

In plain termsMixing machine learning with plant biology is promising, but it is not accurate enough on the details that matter yet.

The scienceThe most promising approach is combining machine learning with plant physics, so the model learns from both mechanism and data. It predicts overall growth well but is still weak on the details that matter, like nutrient levels, and has never been proven over a full multi-week crop.

In plain termsA model can be '95% accurate' and still miss the date you promised a customer; we need forecasts that honestly show their own uncertainty.

The scienceA forecast can look excellent overall and still be off by 10-34% on any single prediction, and naive forecasts claim more certainty than they have. Simple fixes exist that make the stated range honest (coverage jumps from 40% to 81%), but growers are not getting them.

In plain termsModels trained early in a crop go out of date within weeks as sensors drift and data gets missed.

The scienceA forecast trained on week-one data goes stale by week four: sensors drift, logs get skipped, and early 'predictors' were never truly driving the outcome, so the model quietly stops matching reality.

In plain termsAll public yield models are trained on lettuce alone; nothing handles a real mix of greens.

The scienceEvery public yield study is single-crop lettuce; no published model spans lettuce plus basil plus arugula, and accuracy drops on species the model has never seen.

In plain termsCompanies promise big savings from 'digital twins', but nobody independent has verified them.

The scienceVendors advertise 25% energy savings and 30% less crop loss with no independent audit, while peer-reviewed assessments call the field embryonic (only about 9% actually virtualize the plants).

In plain termsHarvest schedulers assume yields are certain, so they plan against a number that is always wrong.

The scienceThe mathematics of scheduling is solved - mixed-integer linear programming, constraint programming, and reinforcement learning all work - but every published scheduler assumes yields are exact, so it optimizes against a number that is always wrong.

In plain termsWe know the environment changes how food tastes, but nobody has mapped 'grow it this way -> it tastes like this'.

The scienceLight, Electrical Conductivity (EC), temperature and stress each shift volatiles, sugars and phytonutrients, but there is no public model predicting a cultivar's chemotype from its climate recipe.

In plain termsNew food-safety law wants lot-level records for every harvest, and the tools assume field farming, not continuous harvest.

The scienceFSMA Section 204 requires lot-level records at critical tracking events (compliance Jan 2026), but standard tooling assumes field harvest, not continuous NFT harvest.

In plain termsWe promise every batch tastes the same, but nothing actually measures and corrects the flavor each cycle.

The scienceControlled Environment Agriculture (CEA) is marketed on 'identical every batch' but nobody runs a closed loop that reads each batch's chemistry and corrects the next cycle.

In plain termsThe 'way more nutritious' claims are everywhere, but the science behind them is genuinely conflicting.

The scienceThe '4-40x more nutrients' microgreen claim powers Controlled Environment Agriculture (CEA) marketing, but soil-vs-hydro studies contradict each other and even U.S. Department of Agriculture (USDA)'s two summaries disagree (~5x vs 4-40x).

In plain termsHow long greens stay fresh is decided while they grow, but nothing predicts it from the grow data.

The scienceShelf life is largely decided before harvest (nitrogen, light integral, water status), but no model predicts a batch's days-to-degradation from its grow conditions.

In plain termsIf a food-safety pathogen gets into the shared water, there's currently no way to kill it without hurting the crop.

The scienceOnce a human pathogen is in the shared solution there is currently 'no effective intervention' during production, and a 2024 indoor Listeria recall killed the 'indoor = safe' assumption.

In plain termsPathogens can get inside the leaf where washing can't reach, so 'wash-free' needs proof, not promises.

The sciencePathogens internalize from contaminated solution into edible tissue, and even triple washing removes only ~99% of surface bacteria.

In plain termsMicrogreens are a premium crop, but hand labor is the bottleneck; automation is the only way to scale it.

The scienceMicrogreens turn ~26 cycles a year but labor is the dominant input (12-18 min per tray), so the premium only survives at hand-labor scale.

In plain termsWe can measure flavor in a lab, but not fast or cheap enough to grade every harvest.

The scienceBrix, dry matter and volatiles are the measurable flavor proxies, but today it's a lab (Gas Chromatography-Mass Spectrometry (GC-MS)/Liquid Chromatography-Mass Spectrometry (LC-MS)) or taste panels.

In plain termsNobody can tell you exactly what automation is worth on your farm; the models are often half wrong.

SummaryThere is no tool that measures pounds-per-labor-hour per activity and computes the payback of automating this step vs that, and Controlled Environment Agriculture (CEA) labor models are routinely 40-60% wrong.

In plain termsWe don't know if a seed lot will germinate until weeks later, after it's already taken up space.

SummaryEvery seed lot is a gamble on germination you only learn 5-14 days later, after wasting channel slots.

In plain termsThe recipe for every variety under every condition exists only inside private companies, not in the open.

SummaryThousands of cultivars exist but no open database maps cultivar plus spectrum plus Electrical Conductivity (EC)/pH/temp to time, yield and flavor; that knowledge is private.

In plain termsEvery seedling is still moved by hand into the channel; there's no machine built for a small rack.

SummaryMoving a germinated plug into the channel at spacing is the classic manual 'extra touch' for Nutrient Film Technique (NFT), and no cheap retrofit device exists for a rack.

In plain termsDirty channels can infect the whole loop, but cleaning is by hand and there's no proof it's actually clean.

SummaryBetween cycles channels carry root mats and biofilm, and one dirty channel risks the shared loop, but hand-washing is slow and unverifiable.

In plain termsHarvest robots work only in giant purpose-built facilities; nothing retrofits onto a small rack.

SummaryOnly whole-facility systems prove harvest automation (SPREAD 30k heads/day, 99% uptime; Iron Ox shut down), and no retrofit harvester for a small rack is sold.

In plain termsPlants get carried by hand up and down the rack several times; the robots exist, but not the retrofit.

SummaryEvery plant moves several times and on a rack most of that is lifting and carrying; Autonomous Mobile Robot (AMR) navigation is commodity and open-source systems exist, but the gap is retrofit integration with an in-house 4-tier rack.

In plain termsTo breed for indoor farms you must measure how each variety responds, which is still done by hand.

SummaryIndoor breeding is meaningless without measuring cultivar response to a light recipe, and that phenotyping is manual or research-only.

In plain termsPicking individual leaves without bruising them is the hardest thing for a robot to do.

SummarySelectively picking individual marketable leaves without damaging the crown is where robotics fails completely, combining fine manipulation, occlusion and a 1-2% bruise tolerance.

In plain termsRobots crush wet, delicate leaves; the grippers that don't aren't reliable enough for a farm yet.

SummaryHolding a wet, fragile leaf without crushing it is a materials-and-control problem upstream of harvest automation; research grippers reach 42-73 fps on cheap edge nodes but none handle leafy greens at commercial reliability.

In plain termsCloning the best plants would fix inconsistency, but it's mostly hand labor and expensive machines.

SummaryFor premium slow-from-seed herbs, clean clonal propagation would fix genetics inconsistency, but micropropagation is labor-dominated and automation is hard to scale.

In plain termsBuilding a centralized vertical farm is 5-10x the undertaking of a greenhouse, and the easy funding that once fueled the boom is gone.

SummaryA centralized vertical farm must earn back 5-10x a greenhouse's capital on a commodity crop, and the easy venture capital that bridged the gap is gone.

In plain termsIndoor lettuce can't win on the open market against field lettuce; it wins on flavor, freshness and locality.

SummaryA lit stacked farm can never match a field head or a greenhouse head on commodity greens, so every venture that sold on price lost.

In plain termsLettuce is near its biological limit indoors, and no staple crop has been made to work vertically yet.

SummaryCurrent indoor genetics are near their biological ceiling (~230 kg/m2/yr lettuce), so 'more light, more layers' stops paying, and beyond leafy greens, microgreens and strawberries no calorie crop has vertical economics.

In plain termsEvery farm records data differently, so no two farms can compare; there's no common language yet.

SummaryControlled Environment Agriculture (CEA) has no agreed schema for recording crop, sensor or yield data, so farms aren't comparable and everyone rebuilds analytics from scratch (the International Organization for Standardization (ISO) standard is still in draft).