Correct the drift before the batch.
In a continuous process a small drift compounds. AnantState tracks each unit against its expected behavior, so a parameter heading off recipe is visible while the batch is still good.
In plain termsSee a process leaving its recipe before the product does.
Drift, and the correction that closes it
The entities are the process units, batches, feeds and utilities. The expected path comes from the recipe and the physics, which is why the baseline is strong in this domain and the learned part is a small correction rather than the main act.
A process drifting before it goes off spec
- FeedstockLeads to: Process unit
- Process unit: State tracked continuouslyLeads to: Still on spec?
- Still on spec?Leads to: On-spec product (yes); Priced correction (drifting)
- On-spec product
- Priced correction: Cost to act beside cost to waitLeads to: Process unit (adjust)
What moves in a unit
| State | Baseline drivers | Unexpected drivers |
|---|---|---|
| Process drift | Recipe, set points, throughput | Feedstock variation, fouling, sensor drift |
| Yield | Conversion, residence time | Catalyst decay, contamination |
| Equipment condition | Runtime, duty, maintenance interval | Wear, leakage, blockage |
| Energy posture | Load profile, tariff, weather | Equipment inefficiency |
Where it goes wrong today
- A parameter is inside its alarm limits and still on its way out of recipe, so the alarm arrives after the loss is committed.
- Corrections are made on experience, with no price beside them for the cost of acting now against the cost of waiting.
- An off-spec batch is explained afterwards, with no way to replay the decision that was made at the time.
A worked example
Forty units, and a correction made on the quarter drifting furthest.
Round, illustrative figures for process units: they show the arithmetic, leave out the subscription, and are for you to change in the calculator.
| Input | Value | What it stands for |
|---|---|---|
| Process units in scope | 40 | Everything the team is responsible for |
| Value at stake per one | $80,000 | An off-spec batch and lost yield on one unit |
| Cost of the action per one | $3,000 | Correcting a drifting parameter before the batch |
| Share acted on | 25% | The units drifting furthest: 10 of 40 |
| Modeled, USD | Pessimistic | Expected | Optimistic |
|---|---|---|---|
| Share of value at stake avoided | 8% | 18% | 30% |
| Cost to act, in total | $30,000 | $30,000 | $30,000 |
| Avoided loss | $64,000 | $144,000 | $240,000 |
| Net position | $34,000 | $114,000 | $210,000 |
Break-even is 3.8%: each action costs $3,000 against $80,000 at stake, so it has to avoid that share of the loss to pay for itself.
Integration reality check
- Plant data is often network-segmented. The read-mostly posture on the integrations page is why this is deployable in an operational-technology environment.
- We target the planning and prioritization layer, not control loops or interlocks.
What exists today
Chemicals and process exists as catalog templates and shares the industrial and IoT domain compiled into the engine. It is not a live simulated plant.
Bring a unit that drifts
Give us a month of process events and the cost of a correction. We will show how early the drift was visible.