As more than one-third of global greenhouse gas emissions are related to the operation of buildings, reducing building energy demand is a key area in the architecture, engineering, and construction (AEC) industry. One promising means is to conduct early-stage building energy optimization. Early-stage energy optimization offers substantial potential, as influential architectural design variables (ADVs) can be adjusted at low cost to achieve significant efficiency gains. However, existing optimization workflows depend heavily on energy simulations, which are time-consuming and computationally expensive. To address this, this thesis investigates the application of machine learning (ML) for accelerating early-stage building energy optimization, focusing on three key areas: developing ML prediction models, extending their generalizability through transfer learning (TL), and embedding them into practical optimization workflows. A key contribution of this research lies in systematically identifying influential ADVs through both literature review and stakeholder surveys. Findings highlight building plan, window-to-wall ratio (WWR), and wall material as consistently important across sources, while practitioners additionally emphasize orientation, shading devices, storey number, storey height, roof type, and roof material. The thesis incorporates ADVs from both evidence-based and practice-based perspectives to ensure the development of robust and practically relevant ML models. Comparative ML experiments further provide recommendations for algorithm selection: Support Vector Machine (SVM) for small datasets, Multiple Linear Regression (MLR) for limited and low-diverse datasets, Artificial Neural Network (ANN) for larger and diverse datasets, and Random Forest (RF) when accuracy outweighs computational efficiency. Guidelines are also proposed for synthetic dataset generation, stressing the need for adequate size and diversity to achieve reliable predictions. To evaluate generalizability, an ANN model trained on Gothenburg data is transferred to five cities with different climates through transfer learning (TL). TL substantially improves prediction accuracy in heating-dominant contexts (Stockholm, Seattle, Chicago), reducing the need for up to 1,600 training samples and saving over 180 hours of computation. Its effectiveness declines in cooling-dominant climates (Madrid, Miami) but remains beneficial when data availability is limited. While its effectiveness is highest in heating-dominant contexts with data scarcity, the results confirm TL’s potential to reduce training requirements and computational time. Finally, the ML model is integrated into a Grasshopper-based optimization workflow and exemplified with a case study. Results show that while ML-based optimization yields slightly higher energy demand than simulation-based methods, it drastically reduces computation time and provides comparable design outcomes. Overall, this thesis advances methodological knowledge on selecting ADVs, algorithms, and datasets for ML-based building energy prediction, while also confirming the feasibility of cross-climate adaptation and workflow integration. The findings offer valuable guidance for researchers, software developers, and practitioners seeking to accelerate sustainable building design.