From Animal Instincts to Machine Emotions: A Physics-Informed Multispecies Framework for Emotionally Intelligent AI

The ability to feel emotions is needed for a species’ survival and social engagement, although emotions have not been deeply integrated into AI technology with biological understanding. PI-EI (Physics-Informed Emotional Intelligence) is a framework presented in this paper that combines zoological instinct modelling, emotional theories in psychology, and physics-based energy models to develop AI with emotional understanding. The proposed system processes animal behaviour with a CNN-Bilstm Instinct Encoder, takes emotional changes into account with Lagrangian mechanics and energy terms, and ends with a Transformer-based Emotion Decoder using Ekman’s taxonomy. It was developed and tested on a cross-domain dataset that incorporated 1200 zoological clips, 800 EEG files, and 17340 emotion labels given by experts. Results from the experiments make it clear that PI-EI performs better, with $93.6 \%$ accuracy, $94.5 \%$ F1-score, and $96.1 \%$ AUC, topping the conventional CNN-LSTM and ResNet-Transformer baselines by $5.4 \%$ and $3.0 \%$ in accuracy, respectively. Results from ablation studies show that F1-score decreases by 3.5% without physics-based regularisation, meaning it supports stable and biologically sound emotion modelling. It performs well on avian species it has not previously seen, reaching a $90.4 \%$ accuracy level. ROC-AUC scores per class were 0.91 for Happy, 0.94 for Sad, and 0.96 for Fear, suggesting that the model can identify emotions well. Grad-CAM found that the model focuses on essential animal behaviours like caregiving and aggression. PI-EI’s use of biological science, physical boundaries, and interpretability creates the basis for improved emotion recognition in affective AI systems.