<p>Cloud computing has recently emerged as a prominent approach to handling large-scale scientific workflows that require high computational capacity, scalability, and efficient resource utilization. Workflow scheduling, involving assigning dependent activities to virtualized resources in the cloud, ensures efficient model performance. Nevertheless, finding a reasonable trade-off between energy cost and delay remains a key challenge in managing large scientific workflows. Typically, metaheuristics like the standard Artificial Fish Swarm Algorithm (AFSA) face severe limitations, including early convergence, fixed parameter settings, and limited flexibility in high-dimensional problems. To this end, this work suggests a Modified Artificial Fish Swarm Algorithm (MAFSA) integrated into a scientific model to improve performance and efficiency. The technique proposed adopts three significant modifications: (i) an adaptive step control model for adjusting dynamic artificial fish movements to navigate between effective and efficient search spaces; (ii) a vision-assisted search model for improving global convergence and overcoming local convergence issues; and (iii) adjustment by introducing an attenuation factor that modifies precision for convergence in subsequent iterations. Extensive simulations were conducted using the CloudSim platform with heterogeneous virtual machines and DAG-based workflows ranging from 100 to 500 tasks. Experimental results demonstrate that MAFSA reduces energy consumption by up to 21–28% and latency overhead by 17–24% compared with the strongest baseline methods. In addition, workload imbalance and performance degradation are reduced by approximately 18% and 22%, respectively, while QoS is improved by nearly 15%. Statistical validation using t-tests and ANOVA confirms that these improvements are significant.</p>