Arash Pourghorban

This study presents an electroencephalogram (EEG)‐based framework for thermal comfort classification by examining epoch sensitivity, feature importance, and interpretability, and the integration of spectral and microstate descriptors. Ninety sessions were conducted in a climate‐controlled chamber under neutral, warm, and cool conditions. Channel‐wise relative‐power (RP) features across five frequency bands were extracted using multiple epoch lengths (2‐, 4‐, 8‐, 16‐, and 32‐s) and evaluated using three classifiers: linear Support Vector Machine (SVM), radial‐basis‐function (RBF) SVM, and Random Forest (RF). Among the tested configurations, SVM‐RBF using 16‐s epochs achieved the highest overall performance when the full‐feature set was used. When reducing feature dimensionality, RF‐based analysis identified a compact top‐15 feature subset that achieved 76.67% accuracy, comparable with full feature set. Feature importance and stability validation consistently indicated that several frontal‐temporal RP features, particularly gamma_T5, theta_T4, gamma_T4, and beta_FP1, were the most reliable contributors to thermal discrimination. Supporting this observation, interpretability analysis showed that theta_T4 was most informative for identifying neutral from warm states, whereas gamma_FP1 activity provided stronger contrast under cool conditions. Finally, combining the selected top‐15 RP features with EEG microstate temporal descriptors further improved RF accuracy to 88.57%, surpassing the SVM‐RBF baseline. Collectively, these findings demonstrate the feasibility of EEG‐based thermal comfort modelling under within‐subject settings and highlight the value of integrating spectral with microstate features, while further validation using larger populations is needed to improve cross‐subject generalizability.

Enclosed airspaces (EAs) are a common component of different energy efficient technologies in building envelopes which have intrinsic adaptive behavior under various climatic conditions. The need for comprehensive accurate numerical models without restrictions for different applications still exists. Thus, for the first time, machine learning and regression-based techniques (Ordinary least squares Linear Regression (OLR), Support Vector Regression (SVR), Kernel Ridge Regression (KRR), k-Nearest Neighbors (kNN), r-Nearest Neighbors (rNN), and Artificial Neural Network (ANN)) were applied to develop precise models derived from the most credible experimental data with stochastic logic of testing and training in several runs. It was found that application of KRR, SVR, and ANN leads to the most desirable outcomes with highest R₂ (0.97–0.99), and lowest errors, however OLR does not provide satisfactory results (R₂ <0.6). Moreover, major deviations are observed for calculations by OLR, kNN, and rNN in horizontal EAs (thicknesses = 40mm) with downward heat flow direction.