In a recent article published in the journal Energies, researchers emphasize the significance of viscosity as a crucial indicator of fluid characteristics. The study explores the potential of artificial intelligence (AI) to advance sensor technologies capable of operating effectively in diverse conditions. Additionally, the article addresses the challenges associated with energy management in hypersonic airbreathing technology, with a particular focus on the need for precise measurement and monitoring of physical properties in extreme environments.
Experimental test configuration. Image Credit: https://www.mdpi.com/1996-1073/17/15/3635
Background
Accurate viscosity measurement is vital in industries such as energy management, chemical engineering, and materials science due to its influence on fluid behavior under varying conditions. Viscosity affects heat transfer, mass transfer, and fluid dynamics, making it essential for the design and operation of equipment like pumps, reactors, and heat exchangers.
While traditional methods, such as capillary viscometers and rotational rheometers, perform well in laboratory settings, they often struggle in extreme environments characterized by high temperatures and pressures. In applications like hypersonic airbreathing technologies, conventional sensors frequently fall short in providing accurate and timely measurements. This limitation can lead to inefficiencies and safety risks, highlighting the need for advanced measurement technologies that can operate effectively under such demanding conditions.
The Current Study
The study presents a New Generation Predicted Viscosity (NGPV) sensor that utilizes a machine learning framework, specifically a supervised artificial neural network (ANN), to predict fluid viscosity under extreme conditions.
Methodology Overview
Data Compilation
The methodology began with the compilation of a comprehensive dataset containing experimental viscosity measurements across various temperatures, pressures, and flow rates. Sourced from three distinct experimental setups, this dataset included over 600 test conditions. Each entry featured operational parameters such as Temperature (T), Pressure (P), and Mass Flow Rate (Q), with viscosity values obtained through rigorous experimental procedures to ensure high accuracy and reliability.
ANN Architecture
The NGPV sensor employed a deep learning model based on an ANN with the following architecture:
- Input Layer: Features corresponding to T, P, and Q.
- Hidden Layers: Comprising up to 8,000 neurons in total, enabling the model to capture complex non-linear relationships between the input parameters and viscosity.
- Activation Function: Rectified Linear Units (ReLU) are used in the hidden layers to improve training efficiency and mitigate vanishing gradient issues.
- Output Layer: A single neuron predicts the viscosity value, expressed in Pascal-seconds (Pa·s).
Performance Evaluation
The performance of the ANN was assessed using several metrics:
- Mean Absolute Error (MAE)
- Coefficient of Determination (R²)
- Root Mean Squared Error (RMSE)
This approach allowed the NGPV sensor to deliver accurate viscosity predictions under extreme conditions, overcoming the limitations of traditional measurement methods.
Results and Discussion
The ANN model was rigorously tested, and performance metrics were obtained during the evaluation phase. The model achieved an MAE of less than 4.8 × 10⁻⁷ Pa·s, indicating a high level of accuracy in viscosity predictions. The R² value was calculated to be above 0.99, suggesting that over 99 % of the variance in the viscosity data could be explained by the model.
This high R² value underscores the model's robustness and reliability. The RMSE was found to be significantly low, further confirming the model's predictive capabilities. The results indicate that the NGPV sensor can maintain accuracy even when subjected to flow rates that are ten times higher than those used during calibration.
The model's predictions were also compared against reference viscosity measurements obtained from traditional online rheometers. The NGPV sensor demonstrated a consistent ability to provide viscosity values that were within the expected accuracy threshold of 6 × 10⁻⁶ Pa·s. This performance is particularly noteworthy given the complexities associated with measuring viscosity in high-temperature and high-pressure environments.
Successful implementation of the NGPV sensor has several significant implications for industrial applications, particularly in energy management. The ability to accurately predict viscosity in real time allows for enhanced control over fluid dynamics in processes such as fuel injection, combustion, and heat transfer. This capability is crucial in hypersonic airbreathing technologies, where precise fluid behavior is essential for optimizing performance and ensuring safety.
Conclusion
In conclusion, the study highlights the advantages of using AI-driven models over conventional sensors. For instance, the NGPV sensor can adapt to varying operational conditions and provide continuous monitoring without the need for recalibration. This adaptability is particularly beneficial in dynamic environments where fluid properties may change rapidly due to fluctuations in temperature and pressure.
While the results are promising, the study acknowledges certain limitations. The dataset used for training the model, although extensive, may not encompass all possible fluid behaviors, particularly in highly complex multi-phase systems. Future research should focus on expanding the dataset to include a broader range of fluid types and operational conditions, which would enhance the model's generalizability.
Journal Reference
Gascoin N., Valade P. (2024). Virtualized Viscosity Sensor for Onboard Energy Management. Energies 17(15):3635. DOI: 10.3390/en17153635, https://www.mdpi.com/1996-1073/17/15/3635