Abstract
Drilling fluid properties change continuously during operations due to temperature effects, solid buildup, influxes, degradation, and interventions. Conventional periodic testing provides delayed data, hindering timely corrections and increasing risks of NPT, well control issues, and formation damage. This study introduces a real-time forecasting methodology for key mud properties – density, Marsh funnel viscosity, plastic viscosity, yield point, and gel strength – with prediction horizons of 15-120 minutes. The approach combines high-frequency surface sensor data (density, temperature, flow, rheology, drilling parameters) with physics-based temperature/pressure corrections and hybrid machine learning models (XGBoost + attention-augmented LSTM). The method was developed and validated using full-scale mud loop experiments and field data from three directional wells (onshore shale gas and offshore HPHT). On unseen test sets, models achieved R2 >= 0.95 for horizons up to 60 min, RMSE <= 0.012 g·cm-3 for density and <= 3.5 s·L-1 for viscosity at 60 min, and directional accuracy >= 90%. Probabilistic forecasts enabled reliable trend alerts. Field trials showed that early predictions allowed preemptive treatments, reduced ECD fluctuations, mitigated barite sag, and improved hole cleaning. The framework enables proactive mud management, with the potential to reduce mud-related NPT by an estimated 15-30% under similar operating conditions, based on observed response times in the test cases. Limitations include reliance on sensor quality and need for recalibration after major changes. Future work may integrate downhole data and automated treatment suggestions.