Workshop
Brief
Machine learning methods are now regularly applied across industry and the physical sciences to automate and accelerate the design of expensive-to-evaluate and structurally complex systems. Typical tasks include design space exploration and optimisation, what-if analysis, uncertainty quantification, and more. These methods are typically designed to be fully automatic, in that they require no interactions with the domain expert running the design process. However, our combined experience in industry and academia continuously reminds us about the need to interact with domain experts when we approach solving practical machine learning-based decision-making problems. This workshop aims at bringing together scientists developing state-of-the-art methodologies, and practitioners that implement these tools to solve real-life problems. We intend to have a combination of talks and posters from a varied number of fields (such as physics, astronomy, engineering, Earth sciences, material sciences, chemistry, biology, medicine, etc.) that all intersect at the point of making optimal decisions using machine learning models, where explicit domain knowledge can play a major role in the downstream modelling tasks.
1. ML Meets Explicit Knowledge
Different branches in the natural science domains have different paradigms for formulating, expressing and incorporating explicit knowledge in order to model the phenomena of interest. For instance, in domains such as physics and Earth sciences, this can relate to studying differential equations to modelling the behaviour of dynamic systems; in biological domains, explicit domain knowledge can be incorporated through e.g., knowledge graphs. The ML approaches applied in the natural sciences often build on these paradigms, exploiting domain knowledge either explicitly in the design of the ML method or implicitly through various techniques such as loss functions, regularisations, logic programming, constraint programming, semantic nets, etc.
2. Expert-informed Machine Learning for Engineering
In engineering design, data-efficient machine learning methods have become more ubiquitous, as an increasing number of applications have arisen that are sufficiently complex that cannot be efficiently solved by purely data-driven approaches. Therefore, it is becoming increasingly important to incorporate domain expertise into these previously purely data-driven methods. Such hybrid approaches, where data-driven optimization approaches are designed to work in tandem with domain experts, can unlock an additional level of optimization efficiency.
Explicit domain/expert knowledge can often be complementary to the one learnt from data. Exploring and exploiting the strengths of these complementary forms of knowledge can lead to a better understanding of the problem at hand, the design of interpretable and trustworthy ML approaches, and ultimately to better downstream performance. However, there are a number of challenges in studying the synergy of these two forms of knowledge:
Understanding the explicitly formalised domain knowledge often requires specialised scientific background. The complexity of the domain knowledge itself can hinder the comprehension of the underlying ML problem.
Researchers in different scientific disciplines take different approaches to formalise explicit knowledge. Lack of overview of possible approaches, their differences and similarities hinders the development of generalisable ML solutions that are applicable beyond one application domain.
There are a limited number of easily accessible research resources, such as standard benchmark datasets. This hinders method development and robust evaluation.
The goal of the workshop is to bring attention to these challenges in different science and engineering domains, and offer a platform for knowledge exchange, discussing efforts, applications and solutions across the different ML disciplines. The workshop will present the potential of recent ML approaches to tackle decision-making problems through exploiting explicit knowledge and will highlight the overarching goals and challenges of designing such ML methods, formalising domain knowledge and analysing scientific data.