Title: Input Convex Lipschitz RNN: A Fast and Robust Approach for Engineering Tasks

Abstract: Computational efficiency and non-adversarial robustness are critical factors in process modeling and optimization for real-world engineering applications. Yet, conventional neural networks often fall short in addressing both simultaneously, or even separately. Drawing insights from natural physical systems and existing literature, it is known theoretically that an input convex architecture will enhance computational efficiency, while a Lipschitz-constrained architecture will bolster non-adversarial robustness. However, integrating both properties into one model is a nontrivial task, as enforcing one property may compromise the other one. Therefore, in this work, we develop a novel network architecture, termed Input Convex Lipschitz Recurrent Neural Networks, that inherits the strengths of both convexity and Lipschitz continuity. This model is explicitly designed for fast and robust optimization-based tasks, which outperforms existing recurrent units in terms of computational efficiency and non-adversarial robustness. Additionally, we have successfully implemented this model in various practical engineering applications, such as optimization of chemical processes and real-world solar irradiance prediction for Solar PV system planning at LHT Holdings in Singapore. Source code is available at this https URL