Structure-guided operator fusion for GPU-accelerated thermodynamic property evaluation in process simulation

Flash calculations, parameter regression, and nonlinear solvers in process simulation repeatedly evaluate thermodynamic properties for large batches of states. On a GPU, performance depends strongly on how consecutive equation blocks exchange intermediate results. Fine-grained kernel partitioning increases global-memory traffic, while deeper fusion increases the computation and state carried by each thread. The appropriate fusion boundary therefore depends on the thermodynamic model, GPU architecture, and problem size. We present a structure-guided method for selecting operator-fusion boundaries. The method represents Peng–Robinson (PR) and Soave–Redlich–Kwong (SRK) cubic equations of state and non-random two-liquid (NRTL) and Wilson activity-coefficient calculations as equation-level dataflow graphs. The size, lifetime, and reuse of intermediate quantities then reveal candidate boundaries. This analysis replaces open-ended trial and error with fusion-boundary recommendations derived from the thermodynamic equation graph; a matched-output comparison confirms the deployed choice on the target device and workload. We evaluate the method on four NVIDIA GPU architectures and across multiple component counts. For the full fugacity-coefficient path with 20 components, full-path fusion improved performance on all four GPUs. At larger component counts, the benefit of full-path fusion narrowed and, on some devices, reversed, showing that the appropriate fusion depth depends on hardware and problem size. On the RTX 4090 system, six controlled device-side batch workloads ran 6.2–8.5 × faster than a 24-thread CPU. This work provides a practical route for implementing GPU acceleration in thermodynamic property evaluation and chemical-process simulation.