Hybrid Thermodynamic and Machine Learning Optimization of Plate Heat Exchanger Performance for Brewery Cooling Systems in Uganda

Background Efficient thermal management in brewing is critical for product quality, energy efficiency, and sustainability. At Nile Breweries Limited, Mbarara, Uganda, the Plate Heat Exchanger (PHE) cooling system has faced persistent challenges in achieving the target wort outlet temperature of 283 K (10 °C) or lower. Understanding and optimizing the thermodynamic and operational performance of this system is essential, particularly in resource-constrained Sub-Saharan production environments. Methods An integrated optimization framework was developed by combining thermodynamic modeling, empirical diagnostics, and machine learning. Over 50 production batches were monitored using a high-resolution (1 Hz) dataset to assess short-term performance, while a longer dataset of 480 batches was analyzed for fouling trends. Statistical regression, LMTD analysis, and correlation tests were applied to evaluate thermal relationships, and a feedforward Artificial Neural Network (ANN) was trained on seven operational parameters in MATLAB/Simulink to enable predictive control. Results Only 14.60% of production batches achieved the cooling target (95% CI: 11.2%–18.7%). Regression analysis showed a significant positive association between wort mass flow rate and outlet temperature (slope = 1.24 K/(kg/s), p < 0.001, R 2 = 0.42), and a negative correlation with heat exchanger effectiveness (slope = –0.0037 per kg/s, p = 0.012, adjusted R 2 = 0.28). LMTD decreased with increasing wort flow (slope = –0.42 K/(kg/s), p = 0.002, R 2 = 0.36), revealing a throughput–efficiency trade-off. Water flow rate improvements diminished beyond 12.1 kg/s, showing no significant effect on effectiveness (p = 0.18) or LMTD (p = 0.15). Fouling analysis indicated a gradual decline in mass-corrected effectiveness across 480 batches (slope = –0.00001 per batch, p < 0.001, R 2 = 0.22), while wort pH exerted a minor negative effect on effectiveness (r = –0.1645). The ANN demonstrated strong predictive accuracy (R 2 = 0.8762, MSE = 0.0578, MAE = 0.0938), with residuals confirming minimal bias. Optimization identified optimal operating ranges of 16–17 kg/s for water flow and <16.5 kg/s for wort flow. Conclusions The study highlights a significant performance shortfall in the brewery’s cooling system, driven primarily by throughput–efficiency trade-offs, diminishing thermal returns, and progressive fouling. By integrating classical thermodynamic analysis with machine learning, the research demonstrates a robust and scalable framework for enhancing energy efficiency and adaptive process control in brewing and other energy-intensive industries. This work advances the application of intelligent thermal management strategies in Sub-Saharan industrial contexts, with future research recommended in real-time fouling detection, economic optimization of flow control, and hybrid predictive modeling for dynamic production environments.