A Process Simulator Learns to Use Machine Learning

KBC, the Yokogawa-owned process engineering business, has released Petro-SIM 7.7, adding machine-learning-enabled hybrid modelling to a simulation platform built on first principles. The interesting choice is the word beside rather than instead: the physics stays, and the models are trained to cover what the physics does not.

That distinction is the whole argument. A first-principles model of a reactor is trustworthy because it is derived, auditable and behaves sensibly outside the range you have data for. Its weakness is everything the equations do not include — fouling, catalyst ageing, feedstock variability, the particular behaviour of a particular unit. A pure data model has the opposite profile: it captures those effects and then produces confident nonsense the moment it is asked about conditions it has never seen. Hybrid modelling keeps the derived structure and lets a learned component absorb the residual.

The release includes an embedded ML Utility Hub that generates synthetic data and lets engineers develop, train and deploy hybrid models inside the Petro-SIM environment without specialist data science skills. Synthetic data generation is the part worth noting, because the classic blocker for plant-level machine learning is not algorithms but labelled operating data: a well-run unit spends most of its life in a narrow band of conditions, which is exactly the dataset least useful for training. Generating cases from the physics model to cover the space the plant does not visit is a reasonable answer.

Other additions are more conventional but no less useful: an expanded optimizer library with new algorithms for complex refinery operations, an extended reactor suite including a new lube hydrocracker model, enhanced physical property methods, and new support for bio-oil processing, electrolysis, pyrolysis and biomass gasification along with decarbonisation correlations.

Petro-SIM 7.7 also provides the engineering foundation for KBC Acuity Process Twin Pro, which continuously monitors model health, identifies deviations from plant performance and keeps models current across their lifecycle. That addresses the other well-known failure of process digital twins, which is not that they are built wrong but that they silently stop matching the plant. A twin nobody has revalidated in two years is a simulation of a refinery that no longer exists.

Chief technology officer Simon Rogers framed the position directly, arguing that industrial AI is only valuable when engineers can trust it through transparency, engineering rigour and validated results — which is the correct standard, and one that hybrid models are better placed to meet than the alternatives.