Artificial neural networks with conformable transfer function for improving the performance in thermal and...

Area of Science:

• * Applied Mathematics
• * Artificial Intelligence
• * Chemical Engineering

Background:

• * Traditional neural networks often require extensive data and complex configurations.
• * Non-integer order transfer functions offer adaptive capabilities for complex modeling.
• * Existing transfer functions may limit neural network efficiency in specific engineering tasks.

Purpose of the Study:

• * To introduce a novel conformable transfer function (CTF) for neural network applications.
• * To evaluate the CTF's adaptability and performance in diverse engineering scenarios.
• * To demonstrate the potential for reduced model complexity and data dependency using the CTF.

Main Methods:

• * Development of a novel transfer function integrating hyperbolic tangent and Khalil conformable exponential functions.
• * Integration of the CTF into neural network architectures for three distinct experimental cases.
• * Validation of model performance using statistical metrics (determination coefficient, adjusted R-squared, slope-intercept) and overfitting checks (MSE).

Main Results:

• * High correlation between models and experimental data: 99% (annular Nusselt number), 97% (volumetric mass transfer coefficient), and 95% (solar thermal efficiency).
• * Demonstrated absence of overfitting through Mean Squared Error (MSE) analysis on training and full datasets.
• * Successful reduction in the number of hidden layer neurons required for effective learning.

Conclusions:

• * The proposed conformable transfer function (ANN-CTF) enables efficient neural network training with reduced data.
• * The CTF exhibits significant adaptability and accuracy across varied engineering applications.
• * This research paves the way for more data-efficient and streamlined AI models in scientific and engineering domains.