Skip to the content.

Ontology infused AI for Materials Discovery

Project ID: 2228cd1380 (You will need this ID for your application)

Research Theme: Advanced Materials

UCL Lead department: Chemistry

Department Website

Lead Supervisor: Adham Hashibon

Project Summary:

Traditional materials development approaches rely on time-consuming bottom-up or trial & error cycles. These start from few processing and composition options, characterise the resulting structure and properties in the hope they give rise to desired performance, which is seldom the case. This necessitates multiple time-consuming cycles. A much more efficient approach is to start from the required Performance and work out backwards the needed Properties, Structure, and Processing (the so called PSPP chain). This is what the data-driven, materials design and discovery AI-powered approach is (top-down or inverse problem). Starting from the desired performance parameters (i.e., from the application perspective) it harnesses the power of AI and ML, via both supervised and unsupervised ML techniques, including physics informed machine learning approaches to determine the desired properties corresponding to the required performance. Then it finds the materials structure giving rise to such properties, and finally provides the desired process and composition giving rise to such structure. Key for the success of this approach is establishing clear relations between the Properties, Processing, and the structure and performance. The PhD Students will investigate and develop various surrogate AI/ML models, including physics informed machine learning and optimisation methods, e.g., Gaussian Process Regression (GPR) and Bayesian optimisation, and novel avenues of multi-agent and other hybrid methods. A particularly exciting cutting-edge approach consists of developing advanced novel methods utilising acyclic graphs based on ontological mereotopological causality relations as means to directly encode physics relations into graph neural network in the form of materials and process descriptors. This method promises to inject fundamental physics readily from the onset into ML thus resulting in novel domain-aware ML models that go beyond the state-of-the-art physics informed models. The PhD student will apply the methods for various materials applications, in particular materials for energy harvesting (thermoelectrics), hydrogen production or storage.