A Benchmark Suite for Verifying Neural Anomaly Detectors in Distillation Processes

Inspired by the success of machine learning in other domains, the application of AI in the field of chemical process engineering has increased in recent years. While neural networks often show paramount performance, they are error-prone in general which is particularly problematic when employed in safety-critical applications, such as chemical plants. This has given rise to the development of verification techniques for neural networks, which aim to autonomously verify (i.e., to mathematically prove) that a neural network fulfills a set of correctness properties, thus that it is safe and reliable. However, there is no general definition for safety and reliability of neural networks. In this paper, we start bridging this gap by introducing a benchmark suite for verifying neural networks used to detect anomalies in distillation processes. With this benchmark suite we aim at confronting existing verification methods with complex, ‘real-life’ properties and thereby foster new advances in the field of neural network verification.