Justin’s Homepage

We are in the midst of the third wave of Artificial Intelligence (AI) research. More than ever before, computer science deals with investigating and employing human-made stochastic systems so complex, that the methods required to deal with them come close to natural science. Even though we constructed these systems ourselves and they are fully observable, we have neither mechanistic control nor a higher-level understanding. For that, we have to experiment on them with empirical methods. To describe this situation, the term “artificial science” comes to mind. It already surfaced in the past, for example in the writings of Herbert Simon in 1969, but it gains relevance once again. Also, on how our reality, in part, approaches Isaac Asimov’s Robot stories.

A visual inductive modelling approach and also a game to teach the fundamentals of the (CRN) inverse problem. Paper and blog post coming soon. In the mean time, you can already try it yourself!

We provide an accessible introduction and overview of the machine-learning and system-identification problem of inferring chemical reaction networks (CRNs) from time series data. Specifically, given data on the temporal evolution of abundances or concentrations of entities, the task is to infer a set of reactions that could have produced this data, including their kinetic laws and corresponding parameters. Solving this problem is expected to significantly accelerate the prediction and understanding of population dynamics by automating the time-intensive modeling process. In a systematic survey of peer-reviewed and preprint articles published until December 2025, we identified 71 publications detailing 68 distinct methods motivated by chemical and biological applications. Based on this large sample of the literature, we provide a current perspective on the methodological developments, propose a three-dimensional taxonomy, and highlight promising future directions. Motivated by the absence of a recent state-of-the-art synthesis and scattered progress across the application domains of CRNs, we view this work as an important step toward further methodological progress.