New Breakthrough in Hybrid Modeling: Debutanizer Simulation Combining Physical Principles and Neural Networks

This article introduces the open-source project debutanizer-hybrid-model, which combines classical chemical engineering principles with deep learning to implement a differentiable hybrid model for industrial debutanizers using JAX, demonstrating the application potential of AI for Science in process industries. Maintained by Danny-Taehyun-Kim, the project was released on GitHub (link: https://github.com/Danny-Taehyun-Kim/debutanizer-hybrid-model) on May 28, 2026. The core idea is to retain the physical model structure and use neural networks to compensate for model deficiencies, achieving efficient and robust simulation.

The debutanizer is a key equipment in petrochemical and natural gas processing, used to separate LPG mixtures into light components (propane, butane) and heavy components (pentane and above). Traditional modeling is based on the MESH equations (Material balance, Equilibrium relations, Summation equations, Enthalpy balance), requiring accurate thermodynamic models (e.g., Peng-Robinson equation of state) and a large number of physical property parameters. However, it faces issues such as complex computation, parameter uncertainty, and difficulty in handling uncertainties in actual processes.

The core of hybrid modeling is to retain the basic structure of the physical model and use neural networks to compensate for model deviations. The physical part uses the Peng-Robinson equation of state to calculate phase equilibrium; the neural network learns the residuals between physical model predictions and actual observations, improving interpretability and extrapolation capabilities. The project chooses the JAX framework, which supports automatic differentiation, enabling functions such as end-to-end training, sensitivity analysis, and optimal control.

The project implements a stage-by-stage MESH solver, considering mass transfer resistance to more accurately predict separation performance; JAX's JIT compilation accelerates the solving process. It also provides synthetic plant data, simulating different operating conditions (changes in feed composition, adjustment of reflux ratio, etc.) for model validation. Benefits include testing extreme operating conditions, evaluating prediction errors, and reducing experimental costs.

Advantages: 1. High data efficiency, requiring less data than pure black-box models; 2. Strong extrapolation capability due to retaining physical constraints; 3. Good interpretability: the physical part is understandable, and the neural network only compensates for errors. Challenges: 1. Reasonable division of model structure is needed; 2. Training stability is affected by different dynamic time scales; 3. Validation requires simultaneous evaluation of physical model and neural network performance, making test design complex.

The methodology of this project has wide applicability: 1. Digital Twin: The hybrid model has both interpretability and continuous learning capabilities; 2. Real-time Optimization: Differentiable characteristics support real-time adjustment of operating parameters; 3. Fault Detection: Residual patterns reveal equipment anomalies; 4. Process Design: Accelerate the development of new processes and reduce experimental costs.

The debutanizer-hybrid-model project shows that AI for Science does not replace traditional methods but seeks the optimal combination of physical modeling and data-driven approaches. For process industries, hybrid modeling is a key path to intelligent operation, respecting domain knowledge while embracing machine learning capabilities. The use of JAX provides a reference for scientific computing, and automatic differentiation is changing the paradigm of scientific modeling.