Solvents are widely used in chemical processes. The use of efficient model-based solvent selection techniques is an option worth considering for rapid identification of candidates with better economic, environment and human health properties. In this paper, an optimization-based MLAC-CAMD framework is established for solvent design, where a novel machine learning-based atom contribution method is developed to predict molecular surface charge density profiles (sigma-profiles). In this method, weighted atom-centered symmetry functions are associated with atomic sigma-profiles using a high-dimensional neural network model, successfully leading to a higher prediction accuracy in molecular sigma-profiles and better isomer identifications compared with group contribution methods. The new method is integrated with the computer-aided molecular design technique by formulating and solving a mixed-integer nonlinear programming model, where model complexities are managed with a decomposition-based strategy. Finally, two case studies involving crystallization and reaction are presented to highlight the wide applicability and effectiveness of the MLAC-CAMD framework.