Abstract
<title>Abstract</title> <p>Activation functions play a critical role in introducing non-linearity to Artificial Neural Networks (ANNs). However, conventional functions like ReLU and Sigmoid often suffer from limitations such as the ”dying ReLU” problem, gradient vanishing, or lack of adaptability to complex data distributions. Inspired by quantum electronics and optical computing, this paper introduces MSPPP-Act, a novel, physics-informed adaptive activation function modeled after the photophysical properties of the highly stable 1-(4-methylsulfonyl phenyl)-3-(4-N,N-dimethyl amino phenyl)-2-propen-1-one (MSPPP) chalcone laser dye. The mathematical formulation of MSPPP-Act abstracts the dye’s distinctive Amplified Spontaneous Emission (ASE) threshold and optical gain profile into a smooth, continuously differentiable function. To address the inherent chemical dependencies of the dye, the solvent effects and concentration dynamics are mapped onto trainable parameters, allowing the activation function to dynamically adapt its shape during backpropagation. Furthermore, leveraging the exceptional photostability of MSPPP, A stabilization mechanism was integrated to prevent gradient saturation and ensure a steady gradient flow under high-pressure training stages. Also, the performance of MSPPP-Act was evaluated across various standard metrics, including image classification and regression tasks.</p>