Hybrid Modelling in Terramechanics

Hybrid modelling is a concept that has been introduced in various fields of research but has yet to be explored in the field of terramechanics. The concept of a hybrid model is to find a middle ground between representing some system or process using traditional parametric models that are derived from first principles and data-based approaches that use statistical analysis or machine learning (ML) algorithms based on large amounts of experimental data gathered on the system that is being studied. The parametric models have the advantage of being based on known physical principles and provide a clear understanding of the system. Unfortunately, the systems being studied are often too complex to be accurately described by such models, and the necessary assumptions that are made in their development can also limit the scope of their applicability. Data-based methods can be used to represent systems that are difficult to model in other ways by simply using ML to map inputs to outputs. Given enough training data, this approach can be representative, but it provides no physical understanding of the system. A hybrid model combines both approaches to make up for each of their shortcomings. In such a model, a well-established parametric model is used to represent a system and a ML algorithm is trained to augment the model in an effort to represent the effects of phenomena that it does not capture. Training data is still required to train the ML component of the model, but the ML component learns only the necessary additions to the parametric model rather than the entire system. By using a hybrid model, the physical insights and understanding that the parametric model provides are maintained while ML is used to augment it in cases where its limitations prevent it from accurately describing the system being studied.