Conventional Digital Twins (DTs) in smart manufacturing rely on complex and time-intensive simulation models, hindering real-time DT-based decision-making. However, the availability of big data in Manufacturing Execution Systems (MES) enables training different Machine Learning (ML) models for fast and accurate predictions and decision assessments. Accordingly, this paper proposes an ML-Based Simulation Metamodeling Method (MLBSM) to facilitate DT-based decision-making for dynamic production scheduling in complex Stochastic Flexible Job Shop (SFJS) environments. The proposed MLBSM integrates three modules: a novel data vectorizing method (SPBM), multi-output Adaptive Boosting Regressor (ABR) models, and a new statistical risk evaluation method. SPBM converts unstructured production log data into numerical vectors for ABR training by calculating numeric penalty scores for each job based on the position of operations in the schedule queue. Each trained ABR predicts mean job completion times for various dynamic scenarios based on shift schedules. The risk evaluation method estimates the standard deviation of job completion times and calculates the delay probability scores for each job, aiding DT in promptly evaluating production schedules. Working seamlessly together, MLBSM modules present a novel way of using production log data for ML training and ultimately bypassing several computationally intensive simulation replications. In this research, a simulation model generates the synthetic MES data, focusing on the machining process at a photolithography workstation in the semiconductor manufacturing. Experiments demonstrate the MLBSM’s accuracy and efficiency, predicting high-risk jobs with over 80% recall and being at least 70 times faster than conventional simulation runs. Sensitivity analyses also confirm the MLBSM’s consistency under different workstation conditions.