Knowledge-guided machine learning deciphers spatiotemporal dynamics in chemical reactor systems
En palabras de los autores
Characterizing the spatiotemporal dynamics of complex chemical reactors, especially localized engineering-critical points or regions, remains challenging in process engineering. Although data-driven methods offer computational advantages, the incorporation of domain-specific physical knowledge can further enhance their predictive performance. Here, we present a physics-informed machine learning framework that systematically embeds coking mechanisms and heat transfer principles into neural network architecture and training constraints. Using a coke-accumulating cracking reactor as a prototypical case, the proposed approach reproduces the spatiotemporal distributions of key operational parameters obtained from conventional simulations with high fidelity and achieves a 105-fold computational acceleration. Crucially, the incorporation of mechanistic knowledge enhances prediction fidelity for critical safety indicators, particularly the maximum external wall temperature (error < 0.5%), enabling precise identification of optimal decoking intervals. This paradigm bridges data efficiency with domain-informed modeling, providing a useful reference for developing real-time monitoring and optimization models for complex industrial systems. A physics-informed neural network embeds coking and heat transfer knowledge via an output integration layer and region-weighted loss to predict reactor spatiotemporal fields; it matches conventional simulation results, improves peak wall temperature accuracy, and runs about five orders of magnitude faster than conventional simulations.
Apareció: miércoles, 18 de noviembre. Nature Communications. Revista con revisión por pares.