Multitasking Architecture Based on Deep Learning and OCR for High-Pressure Phase Equilibrium Detection

Phase transition detection is crucial in chemical industry to define operating conditions of separation processes. The synthetic visual method in high-pressure PVT cells is widely used; however, visual inspection introduces subjectivity that compromises reproducibility. To mitigate this limitation, this work proposed a semi-automated method based on computational approach and deep learning. The experimental dataset consisted of 28 videos of CO₂ + o-terphenyl mixtures (20 Liquid–Vapor transitions and 8 Liquid–Solid transitions). The data set was splitted: 17 videos for training, 5 for validation, and 6 for testing the developed model (4 LV, 2 LS). The system operated using a parallel architecture: a convolutional neural network, ResNet50V2, classifying the phase state (L, LV, LS), and an OCR (Optical Character Recognition) module, optimized through adaptive preprocessing and temporal continuity filters. The transition instant was confirmed through a temporal persistence logic (hysteresis), requiring prediction stability across multiple consecutive frames to prevent false positives caused by transient optical noise. The model was evaluated within the range of 4.3–75.7 MPa and 305.15–413.15 K, correctly identifying the transitions in all six test videos, achieving an R² of 0.99. The correlation showed mean absolute deviations of 0.015 MPa for pressure and 0.005 K for temperature. The largest pressure deviation observed (0.5 MPa) was associated with rapid depressurization during isothermal experiments. The validation of the method demonstrated that the proposed tool substantially reduces measurement subjectivity, being able to improve reproducibility, and providing a scalable basis for future studies of more complex phase equilibria (L–LL, L–LLS) in hydrocarbon systems and food-related products.