As part of subproject P03 of SFB TACO, I will focus on the application of the recently developed on-the-fly machine-learning force field (MLFF) method to complex realistic problems such as solid-state phase transitions and molecular adsorption on surfaces. Specifically, I will be devoted to the development and application of the delta machine-learning approach. The goal of this approach is to develop a MLFF that approaches the accuracy of high-level quantum chemical calculations (e.g., the random phase approximation). This goal will be achieved by machine learning the difference between high-level and low-level methods, ideally using a small number of representative structures. In addition, I will contribute to other topics within P03, such as extending the current MLFF to more challenging ultivalent oxides and predicting tensorial properties. This project will involve many collaborations within SFB TACO. For instance, in cooperation with Franchini (P07), we will help to develop an accurate MLFF for polaron transport in materials, and with Parkinson (P04) and Rupprechter (P08), our development of a concise ML framework for the prediction of tensorial properties will greatly assist the interpretation of their measured infrared spectra.