Implementing artificial neural networks is commonly achieved via high-level programming languages such as Python and easy-touse\ndeep learning libraries such as Keras. These software libraries come preloaded with a variety of network architectures, provide\nautodifferentiation, and support GPUs for fast and efficient computation. As a result, a deep learning practitioner will favor training a\nneural network model in Python, where these tools are readily available. However, many large-scale scientific computation projects\nare written in Fortran, making it difficult to integrate with modern deep learning methods. To alleviate this problem, we introduce a\nsoftware library, the Fortran-Keras Bridge (FKB). This two-way bridge connects environments where deep learning resources are\nplentiful with those where they are scarce. The paper describes several unique features offered by FKB, such as customizable layers,\nloss functions, and network ensembles. The paper concludes with a case study that applies FKB to address open questions about the\nrobustness of an experimental approach to global climate simulation, in which subgrid physics are outsourced to deep neural network\nemulators. In this context, FKB enables a hyperparameter search of one hundred plus candidate models of subgrid cloud and\nradiation physics, initially implemented in Keras, to be transferred and used in Fortran. Such a process allows the modelâ??s emergent\nbehavior to be assessed, i.e., when fit imperfections are coupled to explicit planetary-scale fluid dynamics. The results reveal a\npreviously unrecognized strong relationship between offline validation error and online performance, in which the choice of the\noptimizer proves unexpectedly critical. This in turn reveals many new neural network architectures that produce considerable\nimprovements in climate model stability including some with reduced error, for an especially challenging training dataset.
Loading....