Materials simulations are conventionally programmed in some old-fashioned programming languages, which is a reminiscence of computational science before machine learning become popular, rendering it challenging to communicate between simulation and machine learning (ML). This communication barrier prevents the efficient construction of an ML model from simulation data, and by construction, the ML model generally lacks extrapolability to predict the material structure with target properties unseen in the training set. A team led by Prof. Han Liu from SOlids inFormaTics AI-Laboratory (SOFT-AI-Lab)@Sichuan University (SCU, CHINA), together with collaborators at University of California Los Angeles (UCLA, USA) and Google Brain Team (Google LLC, USA), introduced a computational inverse design framework that addresses these challenges, by programming both simulation and machine learning in a differentiable programming language platform, allowing simulation to gain backward differentiability the same as machine learning. Their seamless integration enables ML model to be trained directly by the differentiable knowledge of simulator—rather than a predefined training set, which is key to enhance the ML model extrapolability. Taking the example of sorption simulation in porous matrices, it is interesting that a deep generative model is trained to enable accurate prediction of porous structures by inputting arbitrary sorption isotherm. Moreover, the computational inverse design pipeline is found to be accelerated by tensor processing unit (TPU) computing, demonstrating the TPU hardware accelerator specialized for machine learning is flexible enough for intensive scientific simulations. Overall, by fusing simulation and machine learning in differentiable programs and hardware accelerators, this approach holds promise to accelerate inverse materials design.