Machine learning-augmented high-throughput catalyst design for NH3 synthesis

Project summary

Haber–Bosch process is a key industrial advancement in terms of NH3 synthesis. The process involves reaction between N2 and H2 feed streams at high pressure and temperature over Fe-based catalyst. Major thermodynamic incentives of energy efficiency and production economics are driving the research towards the development of novel catalysts which are capable of conducting the reaction at moderate temperatures and pressures. Designing a novel catalyst, to enhance the rate of a chemical reaction, is a challenging task which requires significant experimental work. Conventionally, discovery and optimization of new catalysts is done intuitively with time consuming trail-and-error based approaches. The recent progressions in machine learning and high-throughput platforms have made it possible to rapidly screen broader spectrum of materials and identification of relevant features contributing to the targeted reactions facilitating catalyst design and discovery. Moreover, processing of big volumes of data libraries is now viable, with the improvements in computational algorithms, allowing formulation of predictive models leading to systematic process optimisations and understanding of reaction mechanisms.

 

The current PhD project overlays a broader perspective from catalyst screening towards process optimisation. The preliminary component of the project requires data generation by a combination of high-throughput screening experiments, characterizations, and data mining from literature. Subsequently, unsupervised machine learning techniques will be utilised to capture unique patterns in the data targeted towards high reactivity of lattice N2 in metal nitrides catalyst structure signifying activity in the reaction. Relevant materials will be sorted and tested for NH3 synthesis with reinforcement learning based optimiser directing the experimentation process to achieve input parameters leading to optimum process efficacy with minimum experimentation runs. Data libraries generated from this study will be extended by mapping with quantum chemistry simulations, and subsequently be used to train machine learning models with enhanced predictive capability aiding further research in this area.