Energy optimization with an AI digital twin — a badly tuned plant burns steam and electricity every single shift.

The problem

The plant has a thousand setpoints; only a few of them are genuinely optimal.

The operations / energy manager of a chemical plant lives with the same friction every week:

1. The DCS is tuned to the design, not to today. The original setpoints were calculated for a given recipe, a given capacity and a given energy cost. The plant has evolved — the mix changed, raw materials change with the supplier, electricity prices move every hour — and the setpoints are still the ones from commissioning.
2. The veteran's intuition is the best tool available. The operator with the most years knows that “if we drop 3 °C in reactor 2 when we switch to recipe B, the steam header gets a rest”, but that knowledge is not in any system. The day they retire, it leaves with them.
3. Trying to optimize on the live plant is frightening. Moving a setpoint without knowing the effect on quality or safety is a gamble. And simulating it in an external proprietary model is expensive and takes months to calibrate.
4. The energy bill arrives late. Specific consumption per unit of product shows up at month-end; the operator gets no real-time feedback on how much each shift is costing.

The classic system works — the plant operates and produces. But it operates far from the energy optimum most of the time, and nobody in the plant has a way to prove it. This is not a people problem — it is a problem of information (what is happening) and simulation (what could happen) living on separate islands.

How it fits the IRIS system

iLEAN does not add a fourth system — it seals the cracks between the DCS, the veteran's knowledge and the energy bill.

The plant already has a DCS, meters, quality analyses and an MES or an APC. Each one does its own job well. The problem lives in the joints — between real-time operational data and the question “what would happen if…?”, between plant knowledge and the energy bill, between the improvement proposal and the operations approval. iLEAN acts as the putty that fills those gaps, without asking you to change your DCS or your APC.

Edge captures what the DCS does not see. Brain keeps the digital twin calibrated against history. The agent proposes setpoints. The operator signs — the system never touches the DCS without a validated envelope.

The three iLEAN pieces applied to energy optimization with a digital twin:

• Brain (digital twin) — a model of the sub-process trained on the real DCS history and calibrated against live data. It answers questions like “if I raise the feed flow by 8% and drop pressure by 2%, how much steam do I save and how far does quality move?”. It is not a monolithic proprietary simulator: it is a living model that learns with every shift and also captures the veteran's knowledge when the agent cross-references it with real data (what the veteran knew becomes formalized).
• Edge + Tracer — they capture what the DCS does not integrate: the local panel of an old valve by vision, the electricity meter of the utilities, the analysis that is still written down on paper. Without complete capture, the digital twin has gaps; with complete capture, the twin can be trusted.
• Agent — every shift, it proposes optimal setpoints to the operator on their screen — not in an email at 10pm. It cites the source of the calculation (the history, the recipe, the conditions), the expected saving in kWh/t and the effect on quality. The operator accepts, modifies or ignores. The system measures the result against the baseline. What worked gets formalized in the twin for next time.

See the full IRIS architecture →

Before and after

Operating on inherited setpoints vs. the iLEAN digital twin

Aspect	Classic operation with a DCS + a monthly bill	With iLEAN Brain + Edge + Agent (digital twin)
Where the setpoints come from	Commissioning + the veteran's intuition	A twin calibrated on real history + the conditions of the shift
Capturing the veteran's knowledge	In their head, gone on the day they retire	Formalized in the twin when the agent cross-references it with data
Feedback to the operator	Monthly energy bill	Savings in kWh per accepted change, on their screen
Simulating changes	Try it on the live plant or use an expensive external tool	The twin answers in seconds, without touching the plant
Energy baseline	Approximate annual calculation	Live EnPI, per shift, adjusted for conditions
CSRD / customer reporting	Duplicated work, a separate template	The same twin data feeds the reporting

Impact estimate

Impact estimate for your plant — to be validated with your numbers.

The block below is an estimate to be validated with the specific data of your plant. We put it forward so the committee has an order of magnitude; we refine it during the diagnostic.

• Mid-sized chemical plant with a DCS in place, a steam header + electrical supply + one energy-intensive sub-process (reactor, distillation column, pumping train), and at least 12 months of operating history available in the DCS or the historian.
• Digital twin pilot on a priority sub-process (the one with the highest energy cost or the highest consumption variability). First value expected within a few weeks.
• Expected energy savings in the pilot sub-process between 5% and 15% in the first months, depending on the baseline and the initial variability.
• Indicative payback between 4 and 9 months, depending on the energy cost of the pilot sub-process.
• The hard lever adds up: the energy bill + a reportable footprint reduction + a basis for future decarbonization investments + retaining the veteran's knowledge.
• Context data point: the most digitalized sectors raised productivity by up to 40% compared with the least digitalized ones^[2]. Serious digitalization is not a cost — it is a lever.

And the operations manager's reasonable doubt

“What if the twin gets it wrong and proposes a setpoint that affects quality or safety?” — hallucination is a problem of free generation, not of anchored tasks. Proposing an optimal setpoint is an anchored task: the twin is calibrated against real history and proposes within the known operating range. In this kind of task, the best models brought error below 1.5% ^[1]. And even so, what is critical is never decided alone: iLEAN proposes to the operator and never writes to the DCS without a human signature. The three safety rings exist precisely for this: the OT network only accepts validated changes.

^[1] OpenAI paper “Why Language Models Hallucinate”, 2025 — on the reliability of AI in anchored tasks.

^[2] Fundación BBVA / Ivie — productivity of the most digitalized sectors vs. the least digitalized ones (2000-2021).