In recent years, with the rapid development of urban high-rise and high-density buildings in China, urban ventilation, heat island effect, air pollution and thermal comfort at pedestrian level is becoming an important topic in the study of high-density and high-rise urban design. Since the environmental assessment in the design stage can have a direct and significant impact on the final quality of the city, it is particularly important to rethink the relationship between architectural form and urban micro climate, especially the internal logic between urbanmorphology and natural ventilation, and to explore the design method driven by wind environment optimization in the early design stage. Therefore, with the continuous development of environmental performance simulation and experimental tools, the past “trial-and-error” passive design method is gradually replaced by logical generation design, and the role of early conceptual design decision-making for urban wind environment impact is becoming increasingly prominent. However, at present, a large number of performance simulation tools are subject to the tradeoff between simulation accuracy and time consumption, which makes it difficult to bring timely feedback in the early stage of design. In fact, the iterative generation method based on digital simulation tools only chooses the best among the limited random solutions, not from the environmental performance to the inverse solution of the scheme form. This “post-evaluation” paradigm cannot really satisfy the architect’s comprehensive requirements for the environmental performance design. Firstly, this paper introduces the experimental platform using physical wind tunnel as simulation tool, including data acquisition and performance visualization design method exploration. It has the advantages of easy operation, controllable cost, and stable flow field and so on, guided by the initial shape generation design of the scheme. Its perception and real-time acquisition of environmental data make the wind environment simulation change from the postdesign verification to the front-end design exploration.After that, this study summarizes several commonly appliedurban morphologydesign strategies for wind environment performance optimization of high-rise buildings in high-density cities, including rotation, twisting, concaveconvex, hollowing, lifting-up, and develops a set of dynamic model devices for high-rise buildingsfor physical wind tunnel test. In the experiment of urban morphology generation in physical wind tunnel, firstly, the method of optimizing the layout of ventilated buildings is adopted, and then the angle between building orientation and wind direction is determined by wind tunnel experiment under the condition that the building layout is initially determined by sunshine. Finally, the specific form is determined to achieve the optimal control of the overall wind environment of the block. The dynamic model controls the parameters of different servosdirectly by the program, and uses the servos and gears to drive each building model to move. It constantly changes the orientation and shape of the building, including the size of the hollow facade, the external shape of distortion or indentation, and the height of the overhead. The dynamic model has the possibility of generating a large number of different morphology data instantaneously, which meets the needs of machine learning for massive sample data, and opens up the relationship between building geometry generation and urban wind environment data.On this basis, an optimization algorithm of building geometry genera?摩敯獮椠条湮? design based on neural network algorithm is proposed in this paper. The generation system based on neural network includes four parts: morphology optimization strategy, mechanical dynamic system, simulation evaluation system and intelligent prediction system. The parameter logic can be described as the data circulation process of “geometric parameter group - mechanical parameter group - environmental parameter group”. In this study, the twisting form in high-rise buildings are selected for wind tunnel experiments under different plane and elevation strategies, and the mechanical parameters of each morphology scheme and the corresponding environmental data obtained from experiments are taken as sample data to construct a neural network regression model to predict the optimal morphology of twisted buildings under the control of various wind environment independently. Thus, this paper presents a design-oriented, performance-based, high-rise building group morphology generation, prediction and optimization design method based on natural wind environment in high-density urban districts. The application of physical wind tunnel is trying to solving the problems of complex operation and feedback timeconsumption of using CFD (computational fluid dynamics) software. The introduction of dynamic model makes it possible to obtain a large number of entity building models in a short time. The combination of neural network algorithm maximizes the application of wind tunnel test data to realize the direct conversion of design-oriented from environmental performance to architectural geometry. It has found that, in the early stage of urban design, the optimal control of urban form according to the wind environment can be realized, which may avoid the economic and resource losses caused by the repeating modified design based on the local wind environment criteria and pedestrians’ requirement in the later adjustment stage of design, which overturns the “post-evaluation” mode of environmental performance of contemporary architectural