From the surge-margin example to an uncertainty-aware anti-surge controller

[DanBoringName]

The question

Given its flow meter's real state-dependent uncertainty R(Q), does an active inference agent hold a centrifugal compressor's surge margin better than a fixed-margin anti-surge controller on the same plant? The example covers differential-pressure meters (orifice, venturi, nozzle) and Coriolis meters. "Better" is defined under the first result.

The continuation is open as a paper. Lead authorship is yours of course. I'm happy to co-author the meter derivation and the cpomdp side if that helps and my contribution is sufficient. If not I only ask you cite cpomdp.

Why it matters

The surge-margin example is cpomdp's first applied one. Its R(x) follows from how a flow meter works. Nobody chose it.

The meter's element type sets the sign of R′(Q). The sign decides the result. The agent's offset from its preferred margin has the sign of −d log R / dm, so it says which way the agent moves relative to surge.

• A DP meter reads flow through a square root, Q = k·√ΔP. A flat transmitter error then gives R(Q) ∝ 1/Q², and the epistemic term pulls the agent away from surge.
• Coriolis and ultrasonic meters hold roughly a constant percentage of rate. That gives R ∝ Q², and the pull runs toward surge.
• A meter with constant absolute error has R′ = 0. The epistemic term is then the same for every action, and the example's agent reduces to a controller.

Anyone putting an active inference controller on a compressor needs the answer. If the agent beats a fixed margin once a hard surge limit is in place, the meter's uncertainty model becomes a controller input worth measuring. If it loses, the project shows where the epistemic term costs margin. A hard limit then has to overrule it.

I built the example through the Coriolis case and won't be developing it further myself directly (yet). I'm more than happy to assist where I can.

What cpomdp already provides

Every link is pinned to v0.5.0.

• examples/surge_margin.py. The DP and Coriolis meters as R(m) functions, a fixed-R twin, a closed-form law for where the agent stands, and a closed loop through the library's own Agent. uv run --extra examples python examples/surge_margin.py --check runs its eleven falsifiers. Each is CORROBORATED, read off a grid. Its numbers are showcases and none is citable.
• Anti-surge control, the Applications page. It leads with the results and lists the limits that matter most for anti-surge.
• The derivation. It derives R(Q) and the law for where the agent stands. Its last section lists everything the demo leaves out.
• cpomdp.CallableSensor. Carries a state-dependent R(x) as a function of the state. A new meter model plugs in here.
• cpomdp.Agent with an ObservationGoal. The closed loop the example runs.
• cpomdp._lab.harness. Keeps the world and the agent apart, so the true meter can differ from the one the agent models. Its World is linear-Gaussian, and actions come from outside with no selector.
• cpomdp._lab.constructors. Builds models from a declared spec under named perturbations to their matrices. A CallableSensor passes through untouched, so a wrong meter model is built as a second sensor.
• cpomdp._lab.degraded. Filters that are worse than their model on purpose. WrongFixedRBackend filters on a fixed R. The model it is scored under keeps its R(x).

cpomdp._lab is private research apparatus. It carries no compatibility promise. A piece users need moves to a public module once it settles.

What's missing

A realistic uncertainty model per meter type. The example's meters are power laws. The DP meter has R(Q) = e_ref²·Q_ref⁴/Q², the Coriolis meter has R = (e_ref·Q)², and both are 0.2% accurate at design flow. The derivation calls that shared constant an assumption to re-anchor against a real datasheet. Real meters add calibration error, installation effects, drift and, in a Coriolis meter, a zero-stability term.

• The DP model is a first-order propagation through the square root. It rests on ISO 5167, which excludes pulsating flow. So it describes the meter only on the approach to surge.
• The Coriolis model has no zero-stability term, the fixed absolute error that dominates at low flow. There R stops falling and the pull toward low flow weakens.
• Ultrasonic meters sit in the derivation's sign table and have no model in the example.

A hard surge constraint. cpomdp v0.5.0 has no state constraint. Its grid selectors take action_bounds, the box the action grid spans. The example widens its valve grid past the surge line on purpose, so the grid can't act as an interlock nobody declared. Surge protection lives only in a soft economic preference, and the example shows what that costs. The ambiguity term is not a safety factor. A worse DP meter moves the agent closer to surge (rows 3 and 9). The Coriolis meter settles it past the surge line (row 11). The study needs a hard limit or a chance constraint on the margin, in its own loop or as a selector feature.

A baseline. The example compares the agent with a fixed-R twin of itself and nothing else. The Applications page lists this among its limits. A conventional anti-surge controller holds the operating point a fixed margin from the surge line by opening a recycle valve. Nothing in cpomdp implements one. The study needs it on the same plant with the same meter.

The standoff position. Where the agent stands is set by the economic weight Λ, a policy choice. The law makes the epistemic part of the offset scale as 1/Λ. Sweeping Λ over 10:1 moves the standoff further than the whole 8:1 meter ladder does. The prior's width scales the first step as much as Λ does. Only the direction of the offset is a result (row 5). Its size is a Λ artefact. Resolving that is part of the project. One route sweeps Λ and reports the agent's whole trade-off curve. Another prices Λ from what recycle costs.

Plant fidelity. The example has no compressor model. Its closed loop uses the agent's own model as the plant. The state is a single surge margin that persists from step to step and moves additively with the valve. The meter reads true and the run is noise-free. The surge line is a fixed number, and the operating point describes no machine. There are no Greitzer dynamics, no compressor characteristic, no speed line and no gas thermodynamics. Planning is one EFE step deep.

A credible study needs a plant the agent does not share. That means a compression-system model in the line of Greitzer (1976), or a dynamic process simulation, with recycle-valve dynamics and a surge line read off a performance map. The harness's World is linear-Gaussian, so a nonlinear plant runs in its own loop. The derivation names a further step. It calls treating the surge point as an unknown latent the interesting version of this problem.

What a first result looks like

One meter type, Coriolis. The example covers it, and it is the case where the epistemic term pulls toward surge.

Give it a measured or literature-sourced uncertainty model, zero-stability term included. The SEVA work below reports an uncertainty with every measurement and is one place to start. Put a hard surge constraint on the margin. Run the agent and a fixed-margin anti-surge controller on the same plant, with the same meter and the same disturbances. Sweep Λ for the agent and the margin setting for the baseline, and report both margin-versus-efficiency curves. Efficiency here is recycle avoided.

"Better" means the agent's curve dominates. At the same rate of surge-limit violations it holds a smaller mean margin. Or at the same mean margin it violates less.

A negative result counts. An agent that needs the hard limit to stay out of surge, and does no better than a fixed margin once it has one, is worth writing up.

Kind of work

Application / industry example, Validation against a benchmark or oracle (RxInfer, published data)

Rough size

Paper-sized

Status

Open, looking for someone to take it

Prior art and references

• Henry, M. P. and Clarke, D. W. (1993). The self-validating sensor: rationale, definitions and examples. Control Engineering Practice 1(4), 585–610. doi:10.1016/0967-0661(93)91382-7. The SEVA sensor reports an on-line uncertainty with each measurement. Its demonstrator is a Coriolis mass flow meter.
• Henry, M. P. et al. (2000). A self-validating digital Coriolis mass-flow meter: an overview. Control Engineering Practice 8(5), 487–506. doi:10.1016/S0967-0661(99)00177-X. A digital Coriolis transmitter that estimates its own measurement quality on-line.
• Koudahl, M. T., Kouw, W. M. and de Vries, B. (2021). On epistemics in expected free energy for linear Gaussian state space models. Entropy 23(12), 1565. doi:10.3390/e23121565. In fixed-noise linear Gaussian state space models with additive control, the epistemic term is constant and EFE minimisation reduces to KL control.
• State-dependent observation noise reintroduces epistemic value in linear-Gaussian active inference (2026). arXiv:2607.20306. Establishes that a state-dependent R(x) makes the epistemic term non-constant. The example's meters are instances.
• Discussion #76, Certifiable Active Inference. The certification programme this example sits beside. Its first thread asks where state-dependent sensing turns up in real models, and this is one answer.
• Sathish, S., Khader, M., Sayma, A. and Shah, S. (2024). Centrifugal compressor design and surge simulation for active inference based control. ASME Turbo Expo 2024, GT2024-124905. doi:10.1115/GT2024-124905. Proposes active inference for anti-surge control, on a natural gas compressor simulated in Aspen HYSYS. The example is my reading of it and makes no claim about what the paper did.
• Botros, K. K. and Henderson, J. F. (1994). Developments in centrifugal compressor surge control: a technology assessment. Journal of Turbomachinery 116(2), 240–249. doi:10.1115/1.2928358. A review of conventional surge control, for the baseline.
• Greitzer, E. M. (1976). Surge and rotating stall in axial flow compressors, Part I: theoretical compression system model. Journal of Engineering for Power 98(2), 190–198. doi:10.1115/1.3446138. The compressor-plenum model of surge dynamics. The example has none of it.

Skills that would help

Flow metering or process instrumentation, compressor or anti-surge control, Bayesian filtering, JAX

Get in touch

Responding here is the best way!
Active Inference Institute Discord cpomdp channel
hello@inferogenesis.com

Checks

• I searched existing Research Questions and this isn't a duplicate as far as I could find.
• Results from this project would be shared openly (preprint, repo or this thread).