Energy‐Aware Multi‐Objective Workflow Scheduling Using Proximal Policy Optimization and Kolmogorov–Arnold Networks in Heterogeneous HPC Environments

ABSTRACT Designing adaptive and energy‐aware scheduling algorithms for scientific workflows has become increasingly important as high‐performance computing (HPC) systems evolve toward large‐scale heterogeneous architectures. This paper proposes PPO–KAN, a hybrid workflow scheduling framework that integrates proximal policy optimization (PPO) with Kolmogorov–Arnold Networks (KAN) to jointly optimize makespan and energy consumption. PPO enables adaptive policy learning in dynamic resource environments, while the KAN‐based policy representation enhances expressiveness by modeling complex nonlinear task–resource relationships. The proposed framework supports three reward formulations—makespan‐oriented, energy‐oriented, and weighted multi‐objective—allowing explicit control over optimization trade‐offs. Extensive experiments conducted on four benchmark scientific workflows—Epigenomics (904 tasks), LIGO (922 tasks), Montage (902 tasks), and SIPHT (1004 tasks)—using simulated heterogeneous HPC clusters demonstrate that PPO–KAN achieves up to a 41.7% reduction in makespan and a 28.3% reduction in energy consumption compared with NSGA‐II and PPO with feedforward neural networks. Furthermore, convergence analysis shows faster and more stable learning, with policies stabilizing within 50–100 episodes. Ablation studies further confirm that architectural enhancements improve energy efficiency and policy robustness. Overall, the results indicate that PPO–KAN provides an effective and scalable solution for multi‐objective workflow scheduling in heterogeneous HPC environments.