GENIE V2.1: learned representations of industrial equipment, paired with validated physics

Industrial energy is a physical system. Every kilowatt that enters a plant through its main meter leaves as heat, motion, pressure or product, and each machine between the meter and the work obeys conservation, its own performance curves and the limits its operators set for safety. An answer about that plant is useful only if it respects this structure. It has to know which machine is running, how much of the load that machine carries, what the load costs, and whether the reading in front of you is physically possible.

General-purpose models do not carry this structure. A language model reasons over text, while a plant speaks in watts and degrees. A generic time-series model learns to continue a curve with no knowledge of the plant that produced it.

GENIE is Ampere’s AI for industrial energy, and it rests on a partnership between learning and physics that we call physics-grounded learning. Learned encoders read raw meter and sensor streams and learn the equipment behind them from unlabeled data, so that naming a machine takes only a simple layer trained on labeled examples. Physics, in GENIE’s design, carries what must hold, from the limits of each machine to the structure of the meters. It carries the most weight where data are thinnest: rare faults, few labels, and operating states a plant has never visited.

GENIE V2.1 is the first architecture release of that design. Its signal encoders cover four modalities and are pretrained by self-supervised learning on unlabeled commercial and industrial energy and equipment signals. They work alongside an equipment knowledge graph of more than 127,000 nodes, a specification corpus of more than 125,000 certified equipment performance records, a tariff library of more than 11,000 industrial rate schedules, and weather and carbon feeds. A physics validation layer screens operating states. Its rules enter enforcement as each passes validation on real instrumentation, and V2.1 enforces the first of them. Everything is served through an inference API.

The pretraining does measurable work. In offline evaluation of equipment identification from raw signals, measured with no hint of the answer in the inputs, the pretrained encoder reaches about twice the macro-F1 of an untrained encoder of the same design.

Learned representations and validated physics are the two foundations for where GENIE is going: an action-conditioned world model for industrial energy, one that predicts how a plant responds to what is done to it. A plant operator does not need another dashboard. They need to know what happens if they move a setpoint, shift a schedule or take a machine offline, and they need the answer to obey the physics of their own plant. V2.1 lays the foundations. The releases and papers that follow will report our progress toward that model.

Industry is where energy does physical work. In 2023 the industrial sector used 35% of all end-use energy consumed in the United States (U.S. Energy Information Administration). That energy runs through compressors, chillers, boilers, pumps, fans, furnaces and motors, each a physical machine with a rated input, an efficiency curve and its own way of failing. An AI that helps run this equipment has to carry five kinds of structure, and no quantity of generic text or generic curves will supply them.

Conservation

Energy and mass balance. The site meter records the sum of what every load beneath it draws, plus losses. An answer that assigns more load than the meter saw, or invents a machine to absorb a gap, is wrong before anyone acts on it.

Equipment physics

A compressor cannot ramp faster than its motor allows. A chiller’s capacity is bounded by its compressor and the temperatures it works between, and a cold room cannot warm faster than its thermal mass permits. These limits are known, written down and certified long before anyone collects data.

Metering topology

Meters nest. A circuit meter sits under a feeder and the feeder under the site meter, while most of a plant’s load is never metered at all. Inferring the loads behind an aggregate meter has been a field of its own since the first non-intrusive load monitors (Hart, 1992), and the problem only gets harder in a plant whose equipment was installed over decades.

Degradation

Equipment drifts before it fails, and failures are rare. Machinery prognostics is a long chain, from data acquisition through health indicators to remaining-life prediction (Lei et al., 2018), and the events that matter most are the ones a plant has the fewest examples of.

Safety

Every action in a plant happens inside an envelope of product temperatures, pressures and demand limits. Learning to act within such limits has its own research literature (García and Fernández, 2015). In a plant there is no acting outside them.

General-purpose models carry none of this. Large language models are extraordinarily capable with text, yet careful studies find that they reason poorly about time series (Merrill et al., 2024), and that removing the language model from language-model forecasters does not hurt their accuracy (Tan et al., 2024). Generic time-series foundation models (Das et al., 2024; Ansari et al., 2025) model a series and the covariates they are given. They do not see the plant behind it: which meters nest under which, which machine produced a signal, and what limits that machine must respect.

That leaves a question. Can a model learn industrial equipment from raw signals, without labels, and still carry the physics an engineer would require of any answer about their plant?

Our answer is a partnership between learning and physics. We call it physics-grounded learning, and it can be summarized as: learn the signal, keep the physics.

It starts with the signal. A plant reads its meters every second and logs a sensor stream on every loop, and almost none of that record carries a label outside the closed hardware system. Labeling it is slow and expensive, and the labels belong to a specific site. Self-supervised learning puts the unlabeled volume to work: a model learns the structure of the signals themselves without being told what they mean (Balestriero et al., 2023). We want GENIE’s encoders to learn what tells one machine from another, such as the shape of a compressor’s start-up in a power trace or a fan’s load tracking the weather, before any label is attached. Naming a machine should then take only a simple layer trained on labeled examples.

Some of what a plant does is not up for learning. Conservation holds whether or not the training set happens to show it. A machine’s rated limits are on record, and the meter hierarchy is a fact of the wiring. Machine learning has a growing literature on carrying knowledge of this kind (Karniadakis et al., 2021; Willard et al., 2023), and a model built to respect conservation has shown better energy conservation and long-term prediction on the systems that obey it (Greydanus et al., 2019). In GENIE’s design, known physics is primary and the system relies on it for structure.

One school of thought holds that physical priors limit generalization, and that a model should learn its physics from data alone.

In industrial energy we take the opposite view.

Conservation, metering topology and equipment limits are what an answer has to satisfy before anyone can act on it. They carry the most weight where data are thinnest. Take a fault seen twice in ten years, a machine with no labeled history, or an operating state the plant has never visited. Learning gives GENIE breadth across equipment and sites. Physics keeps its answers accountable in the cases the data cover worst.

A rule can be right in principle and still wrong as written for a particular meter or sensor. This is often evident in industrial equipment whose performance curves rest purely on predetermined control logic. That is why, in GENIE’s design, a rule enters enforcement only after it has been validated on the real instrumentation it will screen, against an acceptance bar fixed before the test.

GENIE V2.1 is the first release built on this design.

V2.1 has a signal encoder for each of four supported modalities: building and site energy, electrical power signatures, vibration, and HVAC control. Each turns a window of raw readings into a representation of the equipment behind it. The encoders are pretrained by self-supervised learning on unlabeled commercial and industrial energy and equipment signals, and task-specific heads, trained on labeled examples, read their representations. The model served by V2.1 has about 115 million parameters.

The rest of the system supplies what no meter reading contains. An equipment knowledge graph of more than 127,000 nodes connects each machine to its thermodynamic and energy type. A specification corpus of more than 125,000 certified equipment performance records supplies the rated performance of each equipment type, the baseline for comparing what a machine does with what it was built to do. A tariff library of more than 11,000 industrial rate schedules turns energy into cost, and weather and carbon feeds supply the conditions that drive load and the emissions it carries.

The physics validation layer screens operating states against rules drawn from the physics of the equipment. Rules enter enforcement one at a time, each only after it passes validation on real instrumentation against an acceptance bar fixed before the test.

Ramp rate is the first rule to clear that bar. V2.1 enforces it on the real-power streams it was validated on, in normalized units, and reports it as not applicable elsewhere. Each further rule enters the same way, once it clears the same bar.

Answers leave through the inference API, which serves identification, physics validation and representations to applications. Load attribution combines identification with the plant’s onboarding records. Ampere’s Marichi, the product built on GENIE, is the layer that acts on GENIE’s answers.

V2.1 offers four capabilities.

Equipment identification is the base of everything else, because the questions a plant asks are about its machines. Given a window of power readings from one meter, it says which piece of equipment the window comes from, with a confidence for each window.

We measured what pretraining adds. In offline evaluation on a held-out, time-ordered test split that training never sees, measured with no hint of the answer in the inputs, the pretrained encoder reaches about twice the macro-F1 of an untrained encoder of the same design. Macro-F1 averages the identification score across equipment classes, so a rare machine counts as much as a common one.

The pretrained encoder scores higher in every one of five training seeds. Architecture, identification head and data are identical across the comparison; the only difference is whether the encoder was pretrained.

Load attribution answers the question an energy manager actually asks: where did the energy go? It starts from the site meter’s readings and the plant’s onboarding records. GENIE combines identification with the equipment nameplate ratings and operating hours collected at onboarding, and reconciles the result against the site meter, so every kilowatt-hour on the meter is accounted for, in units the energy manager can bill.

In V2.1 the physics screen has one enforced rule: ramp rate, on the real-power streams it was validated on. Given an operating state as streams from meters and sensors, the screen reports for each enforced rule whether the state violates it, how often and by how much. On any other input the rule is reported as not applicable, never as a pass and never as a flag. Within that scope it tells an operator when a power reading moves faster than the rule allows.

The fourth capability returns a fixed-length representation of the equipment’s behavior from a window of readings in any of the four supported modalities. Applications can compare these representations across machines and sites, search them, or build on them.

Take a food-processing plant overnight. A site meter records the plant’s load every second. Two circuit meters cover the refrigeration compressors’ circuits, and a monitoring stream reports the real power of the refrigeration plant room. Between 02:00 and 06:00 the site meter records 800 kWh. The plant’s onboarding records list two refrigeration compressors, A and B, their condenser fans and an evaporator defrost heater, each with its nameplate rating and scheduled operating hours.

Identification comes first. The encoders read each circuit meter’s stream window by window, and identification names the equipment on each circuit: compressor A on one, compressor B on the other. The windows in which each compressor is identified, and the near-zero windows on its circuit between them, show both machines stopping during the defrost and restarting together when it ends, at 03:45.

Attribution then puts energy on each machine. Identification places each machine on its meter, the onboarding records give its nameplate rating and operating hours, and the result is reconciled against the site meter. Each machine’s share is its nameplate rating over its operating hours, so the share stays flat while the machine runs and the rest of the site’s load absorbs the meter’s variation. Compressor A is attributed 315 kWh, compressor B 120 kWh, the condenser fans 70 kWh and the defrost heater 20 kWh. The remaining 275 kWh is the rest of the site’s load, and the five shares add up to the meter’s 800 kWh.

At 04:12 the plant-room stream jumps by about 80 kW within 1 second and falls back. This is the kind of stream the ramp-rate rule was validated on, real power in normalized units, and the physics layer screens it with that rule and flags the step as a violation: the reading moved faster than the rule allows. The answer carries the flag, so the plant’s team reviews the reading before anyone acts on it. When they do, the site meter shows no matching rise. It was a telemetry glitch; no machine was behind it.

The context feeds fill in the rest. The tariff library prices the night’s energy at the off-peak rate and shows how the plant’s demand charge is set. The carbon feed shows the grid cleaner overnight than at the afternoon peak, and the weather feed supplies the outdoor temperature behind the condenser load.

What the plant receives is an equipment-level account of its night: which machine each circuit carried and when it ran, the energy attributed to each machine with every kilowatt-hour on the meter accounted for, and one flagged reading to review.

A plant will act on what GENIE tells it. Three practices, built into how we develop GENIE, are what that trust rests on.

Physics rules are enforced only after they pass validation on real instrumentation. Every rule is tested against violations planted into real data and against the same data left clean, and it is held to an acceptance bar fixed before the test. A rule is enforced on the input it was validated on and reported as not applicable elsewhere.

Served behavior is checked against offline results. We send the evaluation’s own test items through the served code, in the units a client sends, and compare the result with the offline number.

Results are scored on held-out test splits that training never sees, and the splits follow time, so a model is tested on a period it has not seen, as it will be in a plant.

GENIE’s goal is an action-conditioned world model for industrial energy: a model that predicts how a plant responds to what is done to it. Move a chiller setpoint, shift a production run or take a compressor offline for maintenance, and the model should say what happens to the load, the cost, the emissions and the product before anyone touches the plant. World models of this kind learn how an environment evolves in response to actions, and they are an active area of research (Ha and Schmidhuber, 2018). Industrial energy is one of the places they matter most, because every action in a plant has a price, a carbon cost and a safety envelope.

V2.1 lays the two foundations that model needs. One is a learned representation of equipment, built from raw signals without labels; it is how the model knows what is running and how it behaves. The other is validated physics, and it matters even more once a model is asked about actions. An action takes a plant toward states it has rarely or never visited. The data are thinnest there, and conservation, equipment limits and the meter hierarchy have to carry the answer.

An AI that helps run a plant has to know which machines are running, where the energy goes, what it costs, and which readings cannot be true. General-purpose models were not built to carry that structure. GENIE is built around it, as a partnership: encoders that learn equipment from raw, unlabeled signals, and physics that carries what must hold, with rules that earn their place on real instrumentation.

GENIE V2.1 is the first release of that design, with signal encoders for four modalities, an equipment knowledge graph, a specification corpus, tariff, weather and carbon context, a physics validation layer and an inference API. Its pretraining already does measurable work, and its physics enters enforcement one validated rule at a time.