1. A smart feeder is four different weighing jobs

The brief always arrives the same way: “we need a load cell for our smart pet feeder.” That sentence covers four jobs with four different masses, four different error budgets and, in my experience, four different answers. Until I get the person on the call to separate them, the numbers we exchange are noise.

Job A is the one the product is sold on and the one with the smallest signal — the food in a cat’s bowl is about 1 % of the mass of the hopper you were probably thinking of instrumenting. Job B is a level gauge, not a meter. Job C is not a weighing problem at all, it is a verification problem, and it only works if A is trustworthy. Job D is a completely different product that happens to share a shopping list.

There are also three technology families competing for the same feature, and it is worth being clear about which one you are in before we talk about sensors:

If you are still deciding between a single-point cell, a set of planar beams and a four-cell summing tray, the general logic is in How to choose the right load cell for your project. This article assumes you have already chosen strain gauges and now have to place them.

2. A cup is not a mass, and it is your real competitor

Before we price anything, it is worth knowing what the load cell is being asked to beat. It is not another load cell. It is a plastic measuring cup, and the reason it is a serious competitor is that the Association for Pet Obesity Prevention’s 2025 pet weight and nutrition survey found something I have not been able to forget since I read it.

So the competing instrument is used by almost everybody, and it does not work. Why not? Because a cup measures volume, and the thing you are trying to control is mass. The conversion factor between them is bulk density, and bulk density in this category is not a number — it is a range wide enough to swallow the entire clinical problem.

The same scoop, the same instruction on the bag (“one cup twice daily”), and a 3.3× spread in the mass delivered. Even if you stay inside extruded kibble, which is what most of your customers feed, you are still looking at 84 g against 132 g — a 1.6× spread, with no way for the owner to know which end they are on, because pet food packs almost never print bulk density. The only reliable fix is the one the vets have been recommending for years: weigh it.

Here is why that margin matters more than it sounds like it should. Take a 4 kg indoor cat on 55 g a day — roughly 198 kcal at 3.6 kcal/g, which is an ordinary maintenance ration. Now express the error as time instead of grams, because that is how owners experience it:

Read the last column. A 10 % overfeed is not 10 % — it is 36.5 extra days of food per year, which is what it feels like to the animal and to the owner’s wallet. And look at the 4 kg cat row at 20 %: 4.02 kg of extra food a year, on a 4 kg animal. The measuring error is larger than the pet. I have used that sentence in three customer meetings and it has ended the argument each time.

3. The precision band is four kibble pieces wide

Now I can say what precision you need, and it is much less than anybody asks for. Work back from what the veterinary literature treats as a meaningful change:

• Weight-management plans typically target 1–2 % of body weight lost per week. Faster than that and you risk muscle loss.
• Obesity risk rises meaningfully at just 10 % above ideal body weight — well before an animal looks overweight.
• A small published study of dogs with osteoarthritis found lameness improved starting at only a 6.1 % reduction in body weight. The clinical response curve is steep at the bottom.
• A University of Liverpool / Waltham study across more than 50,000 dogs in 12 breeds found overweight dogs’ lifespans were up to 2.5 years shorter than ideal-weight dogs’. In a species that averages 10–13 years, that is the whole tail of the animal’s life.

Put those together and the control band — the error you actually need to stay inside — is roughly ±2 % to ±5 % of the daily ration. Not 0.1 %. Not 0.01 %. Two to five percent.

The entire daily error that changes a 15 kg dog’s life is four pieces of kibble. That is the size of the problem. Everything a feeder does beyond getting that arithmetic right is comfort features.

4. Repeatability is a percentage of capacity, not of reading

This is the section I wish every pet-tech engineer read before they send me a drawing. It is also the reason I have had to tell four customers that the sensor they had already chosen was the wrong one, and in every case the sensor was fine.

Every strain-gauge specification that sounds like accuracy — repeatability, non-linearity, hysteresis, creep, temperature effect — is written as a percentage of full scale. Full scale is the cell’s rated capacity, not the load you happen to put on it. So the error a cell contributes to your reading is:

That ratio is the whole design. Buy a cell whose capacity is 3× your gross and you have just divided your resolution by three, at the same price and with the same datasheet. And if you put four cells under a tray, you multiply the capacity by four while the load stays the same — the averaging only claws back the square root.

Here is what that looks like with parts out of my own catalogue. Every one of these cells is specified at 0.05 %F·S repeatability except the two planar beams at 0.02 %. Nothing in the left-hand column is a trick.

Look at the span of column five: 0.25 g to 5.00 g, a factor of twenty, on cells whose datasheets all say the same number. The only thing that differs is how much capacity you bought that you were not using. Four bars of the same accuracy class, four completely different products.

The four-cell row is worth doing by hand, because it is the counterintuitive one. A 20 kg tray set is twenty times the capacity of a single 1 kg cell. Summing four cells averages their random errors, and averaging four independent things buys you √4 = 2. So you divide by twenty and multiply by two: ten times worse than one small cell, before a single real-world effect is added. The arithmetic is not subtle, but the intuition runs the other way — “more sensors must be better” — and I have watched four engineering teams go that way.

One nuance before we leave it. Repeatability is not accuracy. Repeatability says the same load reads the same number twice; accuracy says that number is near the truth. A feeder needs the second one only on the day it is calibrated, and lives on the first one forever after. If your firmware computes “eaten = before − after”, a constant offset cancels completely and a wandering one does not. So the purchase spec should lead with repeatability, zero stability and creep, and put comprehensive error second — which is the opposite of how most quotes in this category are laid out. On a cell that reads 300 g for a living, 0.05 %F·S comprehensive error is 0.15 g, below everything else in the chain.

5. The pet is 17–40× the signal

Now put the animal in the picture, because the bowl scale has to share its platform with something that weighs more than the food by a very large factor and does not hold still.

An ordinary cat, resting a paw on the rim while it decides whether it is hungry, puts sixteen times the entire food mass onto the platform. A large dog puts forty. And unlike the classic tare problem, this load is transient and it arrives exactly during the window you are trying to measure. The bowl scale cannot distinguish “the pet ate 20 g” from “the pet is standing on me” by looking at mass alone. There is no filtering trick around that; the information is not in the mass.

What does work is not a better cell but a better question. Three conditions in firmware, in this order:

1. Stillness first. A reading only counts if the platform has been quiet to within a fraction of a gram for several seconds. A paw arriving is a step; a pet settling is a step; both are rejected by the same gate.
2. Then direction and size. Eating removes mass in steps. Kibble removal is discrete, and the smallest meaningful step is one piece, which is 0.2 g for cat kibble and 2 g for large-breed kibble. Anything smaller than the local piece size is a crumb, or noise.
3. Then persistence. The new level must hold. A paw leaning is a step and a reversal. Accept the step only if the level is stable after it, which is typically 20–60 seconds of patience that costs you nothing in the app and everything if you skip it.

I should be honest about the boundary here: that state machine is firmware, it is yours, and I do not write it or sell it. But it is the reason the sensor choice matters. A cell with poor zero stability makes condition 1 impossible — the platform never looks quiet, so nothing is ever accepted, and the product ships with a threshold loose enough to let the cat through. Weighing hardware that drifts is not a small quality problem. It silently sets the sensitivity of every decision above it. I wrote about the same mechanism from the other direction in 3D printer filament weighing, where a drifting load cell makes the spool-empty alert fire at the wrong time.

6. Pieces average out. The last piece does not.

Pet food is not a continuum. It is a count of discrete objects, and a smart feeder sits on top of that fact. A 240 g daily ration of medium kibble is about 200 pieces; a 55 g cat ration of 0.2 g kibble is 275 pieces. Your app reports a smooth number of grams. The physical system is counting.

Two different effects live inside that, and they behave in opposite ways. Individual kibble pieces are not identical — a premium extruded diet holds piece mass to roughly ±5 %, an economy diet closer to ±15 %. That is a random error, and random errors average away. But the granularity — the fact that mass can only arrive in 1.2 g steps — is not an error at all, and it does not average away. It stays exactly one piece wide no matter how long you average.

Read the last two columns together. Over a full bowl the food’s own piece-to-piece variation is worth less than a gram, and the granularity is half a percent — a 0.25 g cell is comfortably finer than both, which is why the cat bowl line in the previous section is not a compromise. But when three pieces are left, the step is 33 % of the reading. Your app can print “2 g remaining” and the truth is anywhere from 0 to 6 g, because the number cannot exist between pieces.

7. The food gains mass while nobody eats

Here is the failure mode that nobody budgets for, and it took me two projects to see it because it looks like the sensor drifting. Dry pet food is hygroscopic. Extruded kibble leaves the plant at roughly 8–10 % moisture and then spends the rest of its life trying to reach equilibrium with the room it is sitting in. In a bowl on a kitchen floor, over twelve hours, that can move the moisture content by a percentage point or two.

Moisture is expressed as a fraction of the as-fed mass, so the relationship between dry matter and total mass is m = m_dry / (1 − w). Going from 8 % to 10 % is therefore a mass ratio of (1 − 0.08) / (1 − 0.10) = +2.22 % on the same amount of food. Nothing has been eaten. The scale is right. The mass genuinely increased.

The direction is what makes this expensive. Moisture adds mass, so a feeder that infers consumption from a falling bowl will under-report the meal every single time. On the cat’s day — two meals of 27 g — that is 1.20 g of phantom food, which is 2.2 % of the daily ration. Look back at section 3: the control band is 2–5 %. The humidity alone consumes the bottom of the band, systematically, in the safe direction for the pet and the wrong direction for your data.

On the hopper the same effect is an order of magnitude larger in absolute terms: 80 g on 3.6 kg, which is two whole 40 g portions. Compare that to the ±3 g a four-cell planar set resolves on that hopper and you can see the shape of the argument. The hopper’s humidity drift is more than twenty times its own resolution, so it is unmistakably visible — and completely useless as a meter.

8. Volumetric dosing, and the bowl as the only witness

Most automatic feeders dispense by turning an auger or a rotary drum for a set number of revolutions. That is a volume device. To turn it into a mass you multiply by bulk density, and the multiplication has two problems that no amount of motor control fixes.

The first is the recipe. Extruded kibble runs 0.35–0.55 g/mL. Change the food, or change the formula, and the same ten revolutions deliver 1.6× the mass. Support gets a ticket saying “the feeder overfeeds” and the customer’s real problem is that they switched brand.

The second is the hopper level, and it is the one that catches people out. An auger does not fill completely. Under a full hopper the flights pack well; as the hopper empties and the material above the inlet stops pressing down, the fill factor falls. Take a flight volume of 12 mL and ten revolutions:

The same thirty revolutions, the same firmware, the same food, and a 16.2 g difference — 41 % of a 40 g portion — purely because the hopper was emptier. A feeder calibrated on day one with a fresh hopper drifts by roughly 40 % of its target as the hopper drains. That is eight times the entire clinical control band, and it happens slowly enough that no owner connects it to the refill.

There is a third problem, and it is the reason I get called. Bridging. Kibble arches over the auger inlet, the auger turns in the air, and the firmware faithfully reports a successful feed. Nobody is home. Nobody notices for a week.

I have described the same structural idea in three other applications, and the geometry keeps repeating: in smart vending machines, in 0.1 g espresso dosing, and in smart shelves with one cell per bay. Wherever the machine actually moves something physical, the weighing sensor stops being a sensor and becomes a check.

9. Evaporation is a slope. Drinking is a step.

Most of the products that ask me for a food-bowl cell also want a water side — a fountain, a reservoir, a wet-food bowl. The physics is different in a way that matters, and I have seen a launch delayed by two months because of one number that nobody estimated.

Water in an open bowl at room temperature evaporates. Indoor still-air evaporation rates sit somewhere between roughly 0.5 and 2.5 mm/day equivalent depending on humidity, air movement and whether the pump is aerating the surface. Multiply by the surface area and you get mass leaving the bowl with no pet involved. For a 22 cm fountain basin that is 0.038 m²:

A cat drinks 10–20 mL per visit. So in a warm dry room, evaporation removes the equivalent of three to nine drinking visits per day that never happened — and because the loss is real and the scale is honest, the app reports them as intake. Note the direction: on the water side, the physical nuisance inflates the number, where on the food side humidity deflates it. Both are wrong, in opposite directions, and a product that tracks both had better know which is which.

There is also a counterintuitive detail: a circulating fountain is worse than a still bowl, not better. The pump keeps the whole basin wetted, aerates the water, and usually increases the exposed surface. Customers ask for the fountain and get more evaporation with it.

The reason I can be specific about all of the above is that a pet feeder is one of the few product categories where the engineering constraints are unusually favourable and the commercial constraints are unusually hard. There is no legal metrology requirement anywhere in this article — nobody pays for pet food by weight, so no approved instrument is needed and no type approval applies. What that means in practice is that the only thing standing between you and a good product is physics, not paperwork. The next section covers the parts of the physics that live in the enclosure rather than the electronics.

10. The four things that break a feeder scale

Every one of these has come up in a real project, and none of them is a load cell specification problem. They are the reason I ask for photographs before I quote anything.

(1) The bowl comes off, twice a week, for the life of the product. Pet bowls get washed. That means the weighing platform has a removable member, which means the seat repeatability is part of your measurement chain. Work out the tolerance and you will see how tight it is: if you want ±1 g on a 60 g meal sitting on a 310 g gross, the bowl must return to the same position to within ±1 g, which is ±0.32 % of the gross. On three locating pins on a 120 mm bolt circle, 0.1 mm of slop at one pin is enough to move load between supports and drift the zero several grams. That number is a moulding tolerance, not a sensor specification, and no cell on the market fixes it. If your bowl is dropped onto a soft silicone mat instead of hard points, the mat’s compression set will add a slow drift on top of it.

(2) The bowl also goes in the dishwasher. A domestic cycle runs at 60–70 °C with steam and detergent, and the cells in this family are temperature-compensated over −10 to +40 °C. A cell inside the bowl will be at the edge of its compensated range every wash, then wet, then hot, and its zero will move. The design answer is simple and worth stating plainly: put the cells in the non-removable base and weigh the bowl through the seat. Then the dishwasher only ever sees plastic or stainless steel, and the electronics never leave the room.

(3) IP65 is not “washable”. Several cells in this family are IP65 (WST107, WST109, WST1101, WS1102, WST102), which is protection against dust and water jets from any direction — a good rating for a kitchen floor. It is not a dishwasher rating and it is not a submersion rating. Two practical details decide whether IP65 holds: cable entry direction and sealing of the cell in its recess. A gland or cable exit facing up collects water and turns into a bucket; I have written about that particular failure mode in how to tell whether a load cell is damaged. And where four cells are summed, the junction box is the other half of the water problem — use a sealed one with gasketed entries, and read the junction box wiring and adjustment guide before you decide where to put it.

(4) The animal tests the enclosure, not the sensor. Cables get chewed, a 40 kg dog leaning on a 600 g feeder tips it, a cat launches off the lid. None of that is a weighing problem, but all of it terminates in the weighing chain. Two habits cover most of it: run cables in a channel rather than in free air, and give the base enough footprint and mass that the pet has to try. The smart cat litter box guide covers the four-cell summing and matching logic a tray-based feeder inherits, in a far harsher mechanical environment — an animal deliberately climbing into the machine.

11. What to put in the spec sheet

Here is the sheet I actually want to receive. If you send me these nine lines I can quote the same day and the quote will be right the first time.

If you have not chosen a topology yet — single-point, planar set, or summing tray — start with the selection guide, and if you want the fundamentals first, The Complete Guide to Load Cells covers strain gauge physics and the difference between the families in more depth than I can here.

12. What I supply, and what I do not

This is the part of the article I would ask you to read even if you skip everything else, because getting it wrong wastes both our time. I am a weighing-component supplier. I am not a feeder manufacturer.

There is no legal metrology requirement in this application. Nobody pays for pet food by weight, so no approved instrument is needed, and I will not sell you one. If a competitor offers you “OIML-approved” cells for a pet feeder, ask what the approval is for. Component-level OIML R60 describes the cell; it does not approve your machine, and your machine does not need approving. What matters to your customer is that the portion is right and repeatable — a design and assembly question, not a certificate question.

13. Field notes: Hangzhou, and an app that said 5 g

A Hangzhou OEM called me in the spring of 2026 with a WiFi feeder they had already shipped about three thousand of. The complaint was specific and strange: the app said the cat had eaten 5 g, and the owner could see an empty bowl. The support queue was full of it. Their engineering conclusion was that the load cell was no good, and they had already picked a replacement from a competitor.

I did not quote anything. I asked for one thing: log the raw gross reading for seventy-two hours with no cat, no food and no lid. They thought it was a waste of a week. It took about four hours to write and it answered the question before lunch on the first day.

Two mechanisms, neither of them in the sensor. The bowl sat on a 2 mm silicone mat; under a 310 g load that mat takes a compression set over about three days, its effective height drops, and part of the load starts finding its way around the load cell into the housing. That is the smooth 7 g. And when the bowl came back from the wash, the keeper tabs were not fully home — the gross came back 4 g high, and the firmware helpfully zeroed on that as its new reference. The drift then stopped, because the firmware had just redefined the truth.

Here is the number that ended the discussion. A cat meal in that design is 27 g. Eleven grams of zero wander is 41 % of the signal, and it moves in both directions. Detect a 5 g bite against a reference that wanders by 11 g and the arithmetic produces exactly the support tickets they were reading: phantom meals, missing meals, and “5 g eaten, bowl empty.”

Phantom meal reports went from around six a week to fewer than one a month on the next firmware build. Same cells. Same tray. Same everything that cost money. I have had this conversation often enough now that I ask for the empty-platform log before I open a catalogue, and I recommend you make it the first acceptance test in your own process — forty-eight hours, no cat, and you will find out whether your weighing chain or your mechanics is the thing that moves.

14. Field notes: Shenzhen, four cells under an 800 g bowl

A Shenzhen team sent me a tray drawing in 2026 and a requirement I have now seen four times: “we want ±1 g portion accuracy, and we are printing it on the box.” The tray was a formed stainless plate on four half-bridge cells rated 5 kg each, summed into one bridge, with the bowl sitting on top. Gross load, with a large-dog portion and a steel bowl: about 800 g.

I did the capacity arithmetic from section 4 in front of them. Four cells at 5 kg is 20 kg of capacity. Each cell carries ±0.05 % × 5 kg = ±2.5 g. Summing four averages the random part, buying √4 = 2, so the set resolves ±5 g. They were asking a platform for ±1 g that could not deliver ±5 g on a bench, and their 800 g gross was only 4 % of the capacity they had bought.

Then the second problem, which was worse and which I had not expected to have to raise. The half-bridge micro family publishes a four-angle error of 1.0 %F·S — I pulled up the WST105 and WST106 pages and showed them. On a 20 kg set that is a 200 g figure. Whether you see all of it or a fraction of it depends on how far off centre the bowl sits, and nobody was controlling that. They were having a ±1 g argument on a platform whose published corner sensitivity is two orders of magnitude larger.

They took option 1 for the tray product and option 2 for a compact model, which is the answer I would have given. What I did not do was sell them the ±1 g story. If a customer tells me a number is going on the packaging, my interest in that number goes up sharply, because a claim on a box is a claim you have to demonstrate at end of line on every unit, and a corner-sensitive platform makes that a per-unit measurement problem rather than a design property.

Two useful side effects came out of that project. They moved the level sensor from the bowl to a single WST1101 planar beam under the hopper, which at ±0.02 %F·S is the most repeatable part in this article and is exactly the right tool for “is there enough food for tomorrow.” And they started specifying gross utilisation as a design rule internally: nothing under 50 %, except the one part that is deliberately a level gauge.

15. Field notes: Utrecht, and a fountain nobody refilled

The most recent one came in during the summer of 2026 from a small team in Utrecht building a combined feeder and circulating water fountain. Their beta users had noticed something odd: on warm days the app reported the cat drinking more — sometimes 90 mL when the owner knew the animal’s normal was closer to 40 mL. Their read was that the load cell was too coarse and they wanted a finer one.

I asked for the raw trace rather than the daily totals, and the shape of it told the whole story. Underneath the visit steps there was a smooth downward slope of 2 to 4 g per hour, depending on the room, present day and night, that never stopped. Their visit threshold was set at 2 g, so the algorithm was happily counting the slope as a continuous stream of tiny drinking events. Their fountain basin is 22 cm across — 0.038 m² of open water with a pump keeping it moving and aerated.

The mechanism is not subtle once you see it, and the fix is not a filter — it is a different question. Evaporation and drinking live at completely different timescales. In five seconds, evaporation removes 0.0055 g; a genuine visit moves 10 g or more. That is a gap of nearly two thousand times, and it means you do not have to be clever. You have to stop asking for a daily total and start asking for steps.

Reported intake fell to a believable 35–55 mL/day and, more importantly, stayed believable on warm days instead of jumping. And the cell did not change. Their WST109 2 kg resolves 1 g against a 5 g decision threshold — a five-to-one margin on a 1.25 kg gross. A finer cell would have made the evaporation more visible, not less, which is a nice illustration of why “buy more resolution” is so often the wrong reflex.

Of the three projects in this article, two were solved without touching the sensor and one needed a smaller cell — and in the one that needed a smaller cell, the customer had been asking for a bigger one. That ratio is not unusual for this category. If you take one thing away, take the capacity-to-gross relation from section 4 and the empty-platform log from section 13.

16. Questions I get asked every time

Recommended models

Six cells that cover every position discussed above: the bowl, the tray, the washable side, and two sizes of hopper. Capacities are the range members that matter for this application; every model has more sizes than the ones listed.