Source-linked AI summary
Conservative Hybrid Graph Networks for Process Systems with Learned Routing
Paolo Guida
TL;DR
Industrial process networks change effective topology and operating regime while their routing mechanisms may remain latent, challenging models that must both predict trajectories and preserve physical meaning. CHGN separates routing, regime, and removal predictions and inserts them into a fixed conservative transport law. It transfers across unseen graph sizes without retraining, while its fixed-graph accuracy is not always best and real-plant performance is limited when interventions are unobserved.
Problem
Industrial process models often observe state trajectories without observing the routing actions and operating regimes that generated them, while unconstrained graph models may not assign stable physical meaning or conserve material.
Method
CHGN uses separate learned routing, regime, and removal mechanisms within a fixed incidence-based transport formulation that preserves internal balance by construction.
Results
CHGN retains low state and mechanism errors on unseen 25- to 40-node graphs after training on 10- to 20-node graphs, unlike the matched graph baselines.
Takeaways & Limitations
The architecture supports zero-shot transfer, stable named mechanism roles, threshold-response evaluation, and conservative internal transport across the studied synthetic networks.
Takeaways & Limitations
Identifiability can remain ambiguous in cyclic, insufficiently excited, or poorly measured networks, and the real-plant fault categories include sessions with only one example.
Abstract
from arXiv · showhide
Industrial process networks do not maintain a single effective topology while operating: streams are throttled or bypassed, and units move between idle, transition, and active regimes. Models of such systems are typically trained on measured state trajectories while the operating mechanisms that generated them remain latent, and an unconstrained graph network can fit such a trajectory without assigning stable physical meaning to the recovered routing. We address both problems with the Conservative Hybrid Graph Network (CHGN), which learns routing, regime assignment, and removal rates as data-driven surrogates and inserts them into a fixed transport equation, so that the mass balance holds by construction for any predicted routing. CHGN trained on networks of 10-20 nodes transfers zero-shot to unseen graphs of 25-40 nodes without retraining, reaching an RMSE of 2.1e-3 against 6e-2 to 9e-2 for GNN baselines under the same protocol, with a gate MAE of 7.9e-3 and regime accuracy of 94.3% (1.2e-2 and 96.4% respectively on the fixed training topology). On a fluid-mixing pilot plant, CHGN improves on a persistence baseline for held-out physical faults but does not predict manual interventions, for which the governing valve actions are unobserved. The model therefore transfers across process topologies without retraining and exposes the latent mechanisms governing plant behaviour to inspection.
1 Introduction
Industrial process networks have changing effective connections and partially observed routing or operating regimes, creating challenges for graph models trained only on state trajectories. CHGN addresses identifiability and conservation by constraining routing within a fixed transport formulation.
- Changing streams, bypasses, and offline equipment make industrial process topology dynamic rather than fixed.
- Measured state trajectories may be available even when the routing actions and operating regimes that generated them are unobserved.
- A flexible graph model can reproduce state updates while assigning non-unique physical meaning to routing weights through gate-message rescaling.
- Identifiable latent mechanisms require additional structure, such as decoder restrictions, priors, or direct mechanism information.
- Unconstrained direct state increments can create or destroy conserved material during rollout because they lack an explicit balance restriction.
- CHGN reconstructs internal transport from the incidence matrix, known rate coefficients, source inventories, and predicted routing while separating mechanism-specific inputs.
2 The Conservative Hybrid Graph Network
CHGN combines shared graph representations with separate heads for routing, unit regimes, and removal rates, then inserts their predictions into a fixed process balance. The incidence-based construction makes internal transport conservative while supporting autoregressive rollout and mechanism supervision.
- Process graph and governing dynamics: CHGN represents process dynamics on a directed graph with units, internal streams, static edge features, and a permutation-invariant global context.
- Process graph and governing dynamics: Known rate coefficients and predicted routing weights determine each internal edge flow from the source inventory.
- Routing and regimes: Local regime heads represent idle, transitioning, and active unit states, while a sink head predicts bounded state-dependent removal rates.
- Algorithm architecture: The spatial encoder maps state history and unit information into topology-dependent latent representations for downstream mechanism heads.
- Routing and regimes: Routing weights are learned for switch-controlled edges using complementary branches that share transport-rate coefficients.
- Assembly: The model reconstructs the future state with an autoregressive Euler rollout, applying a non-negativity safeguard that was inactive in reported experiments.
- Assembly: The signed incidence matrix ensures each internal edge contributes once at its source and once at its destination, so routing redistributes material rather than creating it.
- Training: Joint training combines state prediction with binary gate supervision and categorical regime supervision, while the sink head receives no direct supervision.
3 Results
Across synthetic networks and a real pilot plant, CHGN transfers across unseen topologies, recovers named mechanisms, preserves structural consistency, and improves prediction of held-out physical faults. Its fixed-graph benchmark is less favorable, and manual interventions remain difficult when valve actions are unobserved.
- Transfer to unseen graphs: CHGN transfers from 10–20-node training graphs to unseen 25–40-node graphs without retraining, while matched shared-weight baselines remain substantially less accurate.The comparison uses the same training distribution and zero-shot evaluation protocol.
- Transfer to unseen graphs: 2 × 10^-3 state RMSE is maintained across unseen topologies, compared with approximately 6 × 10^-2–9 × 10^-2 for shared baselines.The separation persists across graph sizes and independently trained seeds.
- Mechanism ablations: Fixing routing increases state RMSE from 0.01267 to 0.05293, whereas fixing regimes reduces regime accuracy to 30.7% with little change in aggregate state error.Across matched seeds, the routing increase has a paired bootstrap interval excluding zero, while the regime-state change interval includes zero.
- Structural consistency: CHGN preserves internal transport consistency, with maximum residual approximately 1.2 × 10^-7 across independently trained seeds.The non-negativity safeguard is inactive, and remaining derivative error is attributed to non-transport approximation and integration differences.
- Real fluid-mixing system: On the fluid-mixing pilot plant, CHGN improves over persistence for leakage-plus-clogging and stirring-error sessions but is worse for manual mode because valve actions are unobserved.State MSE improves by factors of 16.0 and 6.3 for the two physical-fault categories, while manual mode is 2.0× worse in MSE.
4 Discussion and limitations
CHGN transfers to larger process graphs while recovering named mechanisms and preserving their roles during rollout, but its benefits depend on the combined architecture and remain bounded by identifiability and real-plant data limitations.
- CHGN transfers to larger process graphs without retraining, recovering named routing and regime variables while preserving internal transport balance by construction.These properties do not improve trajectory prediction on their own.
- Only CHGN retains low state and mechanism errors on unseen 25- to 40-node graphs, whereas conservation alone is insufficient.The comparison supports a combined role for the fixed transport law, named variables, information partitioning and graph-local parameter sharing.
- Routing is strongly load-bearing: fixing it increases state error, and routing interventions propagate through closed-loop trajectories.Regimes are recovered accurately but have a much smaller effect on aggregate state error in the principal benchmark.
- Different internal routing choices can still produce the same observable state evolution in cyclic, insufficiently excited or poorly measured networks.Intervention and threshold-sweep results support meaningful named gates only on the systems studied.
- The fluid-mixing experiment has satisfactory held-out physical-fault results, but two anomaly categories contain only one session each.The supported use case assumes relevant state and external forcing are observed while routing and operating mechanisms remain latent but physically defined.
5 Conclusions
CHGN constrains latent, dynamic routing and local regimes within a fixed transport law, preserving internal balance while supporting inspection, intervention and reuse across process topologies.
- CHGN predicts switch-controlled routing weights and local unit regimes while constraining how they enter the state dynamics.The formulation combines learned and fixed entries to remove direct gate-message rescaling.
- Internal transport redistributes material without creating or destroying it, and graph-local mappings can be reused across process topologies.A less-constrained sequence model remained more accurate at point prediction on the fixed-graph benchmark.
- On a real fluid-mixing system, CHGN retained forecasting ability for held-out physical faults not explicitly associated with manual operations.
A State reconstruction on the training topology
On the twenty-unit training topology, CHGN’s autoregressive inventory rollout closely follows the true trajectory, while a held-out instant exposes localized regime and gate errors.
- The predicted and true inventory trajectories are almost indistinguishable across a complete autoregressive rollout without teacher forcing.Visible residuals are concentrated in product and recycle tanks and grow with opposite-signed biases.
- At one held-out instant, the process graph reports inventory, regime predictions and switched-branch gate values with their errors.
- The dryer is predicted active rather than transitioning, while the S2 bypass gate has absolute error 0.11 versus 0.01–0.02 for the other switches.Transition is the least accurate regime class, and S2 is near the sigmoid’s most threshold-sensitive midpoint.
B Interventional audit and counterfactual response
The audits show that learned mechanisms are operationally meaningful: routing is load-bearing, recovered responses remain monotone under threshold sweeps, and closed-loop intervention effects are quantified on held-out trajectories.
- One-step substitution: 4.115× 10^-4 is the all-unit one-step RMSE with learned variables, compared with 4.239× 10^-4 after replacing both mechanisms with true values.Fixing routing increases error by a factor of 4.7 to 1.932 × 10^-3, while shuffling both mechanisms raises it to 1.987 × 10^-3.
- Recovered mechanism: Figure 7 visualizes each unit’s inventory and regime status and each switched branch’s true gate, predicted gate and error.The figure also uses edge colors to encode absolute gate-error ranges.
- One-step substitution: Learned gates can compensate for approximation error in other operator components, so mechanism importance is indicated by error increases after fixing or shuffling variables.
- Fixed-history threshold sweeps: A monotone fraction of 1.000 is obtained across all five switch and four regime threshold sweeps, with switch-gate MAE from 9.2 × 10^-3 to 2.1 × 10^-2.Regime agreement ranges from 96.4% to 99.7%, while S2 shows an approximately 0.04 low-threshold offset.
- Closed-loop intervention effects: Closed-loop audits roll the simulator and CHGN forward for 40 steps after changing one switch or regime threshold, without teacher forcing.The intervention effect is defined as the counterfactual trajectory minus the factual trajectory for the same initial history.
- Zero-shot transfer robustness: CHGN retains lower state-rollout RMSE and routing-gate MAE and higher regime accuracy than shared-weight baselines across unseen graph sizes and independently trained seeds.The figures present graph-size trends and seed means with 95% Student-t intervals.
- Closed-loop intervention effects: 6.4 × 10^-4 is the intervention-effect RMSE across 240 routing interventions, versus 8.0 × 10^-5 across 192 regime interventions.The comparable absolute counterfactual-state errors are 4.05 × 10^-3 for routing and 3.99 × 10^-3 for regimes.
C Additional zero-shot transfer results
Additional evaluations show that CHGN’s zero-shot advantage persists across held-out graph sizes and seeds, while graph-specific retraining is a distinct secondary comparison.
- The transfer separation persists across unseen graphs with 25–40 nodes and independently trained seeds.This indicates the result is not attributed to a small number of favorable topologies or one initialization.
- Table 4 compares zero-shot CHGN with comparators retrained separately on each held-out topology, so the two result blocks are not directly comparable.The protocol tests graph-specific fitting rather than shared-weight zero-shot transfer.
- Approximately 0.0305 state RMSE is reached by the best retrained models, compared with 2.118 × 10−3 for CHGN without parameter updates.These values come from a secondary graph-specific protocol and are not a replacement for the matched shared-weight comparison.
D Fixed-graph predictive benchmark
On the fixed graph, model ordering is stable across seeds, while the LSTM leads three benchmark panels.
- The model ordering remains stable across seeds, and one outlying run does not produce the LSTM advantage.
E Numerical physics audit
The numerical audit separates approximation error from structural consistency, finding exact update reconstruction and negligible internal-transport residuals while identifying integration and non-transport approximation error.
- 9.8 × 10−3 state-derivative RMSE remains across independently trained seeds, reflecting finite-model and discretisation errors.
- Zero RMSE is obtained when the reported state update is reconstructed from the predicted derivative, confirming numerical consistency with the implemented Euler step.
- Approximately 1.2 × 10−7 maximum internal transport is consistent with floating-point accumulation.
- The non-negativity correction has zero RMS across every entry and rollout step.
- Remaining derivative error reflects non-transport approximation and the mismatch between CHGN’s Euler update and the RK4-generated synthetic trajectories.Baselines lack a separately exposed transport term, so the same structural audit cannot be applied to them; this does not prove conservation violations on every baseline trajectory.